An odd fix for GCC 7

Disable warning as error
Add some include<functional>
2026-11-06 02:52:35 -05:00 · 2019-10-31 15:09:55 -07:00 · 2019-10-31 15:09:49 -07:00 · 2019-10-31 15:09:36 -07:00 · 2019-10-31 15:09:36 -07:00 · 2017-07-17 15:50:03 -07:00
179 changed files with 11329 additions and 2171 deletions
@@ -10,9 +10,8 @@ the CLA and we can cross-check with your GitHub username.

 Complete your CLA here: <https://code.facebook.com/cla>

-If you don't have a Facebook account, we can send you a PDF that you can
-sign offline. Send us an e-mail or create a new github issue to
-request the CLA in PDF format.
+If you prefer to sign a paper copy, we can send you a PDF.  Send us an 
+e-mail or create a new github issue to request the CLA in PDF format.

 ## License

@@ -1,5 +1,34 @@
 # Rocksdb Change Log

+## 3.3.0 (7/10/2014)
+### New Features
+* Added JSON API prototype.
+* HashLinklist reduces performance outlier caused by skewed bucket by switching data in the bucket from linked list to skip list. Add parameter threshold_use_skiplist in NewHashLinkListRepFactory().
+* RocksDB is now able to reclaim storage space more effectively during the compaction process.  This is done by compensating the size of each deletion entry by the 2X average value size, which makes compaction to be triggerred by deletion entries more easily.
+* Add TimeOut API to write.  Now WriteOptions have a variable called timeout_hint_us.  With timeout_hint_us set to non-zero, any write associated with this timeout_hint_us may be aborted when it runs longer than the specified timeout_hint_us, and it is guaranteed that any write completes earlier than the specified time-out will not be aborted due to the time-out condition.
+* Add a rate_limiter option, which controls total throughput of flush and compaction. The throughput is specified in bytes/sec. Flush always has precedence over compaction when available bandwidth is constrained.
+
+### Public API changes
+* Removed NewTotalOrderPlainTableFactory because it is not used and implemented semantically incorrect.
+
+## 3.2.0 (06/20/2014)
+
+### Public API changes
+* We removed seek compaction as a concept from RocksDB because:
+1) It makes more sense for spinning disk workloads, while RocksDB is primarily designed for flash and memory,
+2) It added some complexity to the important code-paths,
+3) None of our internal customers were really using it.
+Because of that, Options::disable_seek_compaction is now obsolete. It is still a parameter in Options, so it does not break the build, but it does not have any effect. We plan to completely remove it at some point, so we ask users to please remove this option from your code base.
+* Add two paramters to NewHashLinkListRepFactory() for logging on too many entries in a hash bucket when flushing.
+* Added new option BlockBasedTableOptions::hash_index_allow_collision. When enabled, prefix hash index for block-based table will not store prefix and allow hash collision, reducing memory consumption.
+
+### New Features
+* PlainTable now supports a new key encoding: for keys of the same prefix, the prefix is only written once. It can be enabled through encoding_type paramter of NewPlainTableFactory()
+* Add AdaptiveTableFactory, which is used to convert from a DB of PlainTable to BlockBasedTabe, or vise versa. It can be created using NewAdaptiveTableFactory()
+
+### Performance Improvements
+* Tailing Iterator re-implemeted with ForwardIterator + Cascading Search Hint , see ~20% throughput improvement.
+
 ## 3.1.0 (05/21/2014)

 ### Public API changes
@@ -74,7 +74,7 @@ depend on gflags. You will need to have gflags installed to run `make all`.
            * run `brew tap homebrew/dupes; brew install gcc47 --use-llvm` to install gcc 4.7 (or higher).
    * Install zlib, bzip2 and snappy libraries for compression.
    * Install gflags. We have included a script
-    `build_tools/mac-install-gflags.sh`, which should automatically install it.
+    `build_tools/mac-install-gflags.sh`, which should automatically install it (execute this file instead of runing using "source" command).
    If you installed gflags by other means (for example, `brew install gflags`),
    please set `LIBRARY_PATH` and `CPATH` accordingly.
    * Please note that some of the optimizations/features are disabled in OSX.
@@ -49,7 +49,6 @@ else
 	PLATFORM_CCFLAGS += $(JEMALLOC_INCLUDE) -DHAVE_JEMALLOC
 endif

-WARNING_FLAGS = -Wall -Werror
 CFLAGS += $(WARNING_FLAGS) -I. -I./include $(PLATFORM_CCFLAGS) $(OPT)
 CXXFLAGS += $(WARNING_FLAGS) -I. -I./include $(PLATFORM_CXXFLAGS) $(OPT) -Woverloaded-virtual

@@ -104,6 +103,8 @@ TESTS = \
 	stringappend_test \
 	ttl_test \
 	backupable_db_test \
+	document_db_test \
+	json_document_test \
 	version_edit_test \
 	version_set_test \
 	file_indexer_test \
@@ -111,7 +112,8 @@ TESTS = \
 	deletefile_test \
 	table_test \
 	thread_local_test \
-        geodb_test
+        geodb_test \
+        rate_limiter_test

 TOOLS = \
        sst_dump \
@@ -146,7 +148,7 @@ SHARED = $(SHARED1)
 else
 # Update db.h if you change these.
 SHARED_MAJOR = 3
-SHARED_MINOR = 0
+SHARED_MINOR = 3
 SHARED1 = ${LIBNAME}.$(PLATFORM_SHARED_EXT)
 SHARED2 = $(SHARED1).$(SHARED_MAJOR)
 SHARED3 = $(SHARED1).$(SHARED_MAJOR).$(SHARED_MINOR)
@@ -343,6 +345,12 @@ prefix_test: db/prefix_test.o $(LIBOBJECTS) $(TESTHARNESS)
 backupable_db_test: utilities/backupable/backupable_db_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) utilities/backupable/backupable_db_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@  $(LDFLAGS) $(COVERAGEFLAGS)

+document_db_test: utilities/document/document_db_test.o $(LIBOBJECTS) $(TESTHARNESS)
+	$(CXX) utilities/document/document_db_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@  $(LDFLAGS) $(COVERAGEFLAGS)
+
+json_document_test: utilities/document/json_document_test.o $(LIBOBJECTS) $(TESTHARNESS)
+	$(CXX) utilities/document/json_document_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@  $(LDFLAGS) $(COVERAGEFLAGS)
+
 ttl_test: utilities/ttl/ttl_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) utilities/ttl/ttl_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@  $(LDFLAGS) $(COVERAGEFLAGS)

@@ -352,6 +360,9 @@ dbformat_test: db/dbformat_test.o $(LIBOBJECTS) $(TESTHARNESS)
 env_test: util/env_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) util/env_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)

+rate_limiter_test: util/rate_limiter_test.o $(LIBOBJECTS) $(TESTHARNESS)
+	$(CXX) util/rate_limiter_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)
+
 filename_test: db/filename_test.o $(LIBOBJECTS) $(TESTHARNESS)
 	$(CXX) db/filename_test.o $(LIBOBJECTS) $(TESTHARNESS) $(EXEC_LDFLAGS) -o $@ $(LDFLAGS) $(COVERAGEFLAGS)

@@ -435,7 +446,7 @@ ROCKSDBJNILIB = ./java/librocksdbjni.jnilib
 JAVA_INCLUDE = -I/System/Library/Frameworks/JavaVM.framework/Headers/
 endif

-rocksdbjava: clean
+rocksdbjava:
 	OPT="-fPIC -DNDEBUG -O2" $(MAKE) $(LIBRARY) -j32
 	cd java;$(MAKE) java;
 	rm -f $(ROCKSDBJNILIB)
@@ -8,7 +8,7 @@ and Jeff Dean (jeff@google.com)

 This code is a library that forms the core building block for a fast
 key value server, especially suited for storing data on flash drives.
-It has an Log-Structured-Merge-Database (LSM) design with flexible tradeoffs
+It has a Log-Structured-Merge-Database (LSM) design with flexible tradeoffs
 between Write-Amplification-Factor (WAF), Read-Amplification-Factor (RAF)
 and Space-Amplification-Factor (SAF). It has multi-threaded compactions,
 making it specially suitable for storing multiple terabytes of data in a
@@ -16,8 +16,7 @@ single database.

 Start with example usage here: https://github.com/facebook/rocksdb/tree/master/examples

-See [doc/index.html](https://github.com/facebook/rocksdb/blob/master/doc/index.html) and
-[github wiki](https://github.com/facebook/rocksdb/wiki) for more explanation.
+See the [github wiki](https://github.com/facebook/rocksdb/wiki) for more explanation.

 The public interface is in `include/`.  Callers should not include or
 rely on the details of any other header files in this package.  Those
@@ -21,6 +21,7 @@
 #       -DLEVELDB_PLATFORM_NOATOMIC if it is not
 #       -DSNAPPY                    if the Snappy library is present
 #       -DLZ4                       if the LZ4 library is present
+#       -DNUMA                      if the NUMA library is present
 #
 # Using gflags in rocksdb:
 # Our project depends on gflags, which requires users to take some extra steps
@@ -47,13 +48,7 @@ COMMON_FLAGS="-DROCKSDB_PLATFORM_POSIX"
 if [ -d /mnt/gvfs/third-party -a -z "$CXX" ]; then
    FBCODE_BUILD="true"
    if [ -z "$USE_CLANG" ]; then
-        CENTOS_VERSION=`rpm -q --qf "%{VERSION}" \
-          $(rpm -q --whatprovides redhat-release)`
-        if [ "$CENTOS_VERSION" = "6" ]; then
-          source "$PWD/build_tools/fbcode.gcc481.sh"
-        else
-          source "$PWD/build_tools/fbcode.gcc471.sh"
-        fi
+        source $PWD/build_tools/fbcode_config.sh
    else
        source "$PWD/build_tools/fbcode.clang31.sh"
    fi
@@ -237,7 +232,7 @@ EOF
      int main() {}
 EOF
    if [ "$?" = 0 ]; then
-        COMMON_FLAGS="$COMMON_FLAGS -DGFLAGS=google"
+        COMMON_FLAGS="$COMMON_FLAGS -DGFLAGS=gflags"
        PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -lgflags"
    fi

@@ -272,6 +267,17 @@ EOF
        PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -llz4"
    fi

+    # Test whether numa is available
+    $CXX $CFLAGS -x c++ - -o /dev/null -lnuma 2>/dev/null  <<EOF
+      #include <numa.h>
+      #inlcude <numaif.h>
+      int main() {}
+EOF
+    if [ "$?" = 0 ]; then
+        COMMON_FLAGS="$COMMON_FLAGS -DNUMA"
+        PLATFORM_LDFLAGS="$PLATFORM_LDFLAGS -lnuma"
+    fi
+
    # Test whether tcmalloc is available
    $CXX $CFLAGS -x c++ - -o /dev/null -ltcmalloc 2>/dev/null  <<EOF
      int main() {}
@@ -0,0 +1,19 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.
+GCC_BASE=/mnt/gvfs/third-party2/gcc/7331085db891a2ef4a88a48a751d834e8d68f4cb/7.x/centos7-native/b2ef2b6
+CLANG_BASE=/mnt/gvfs/third-party2/llvm-fb/963d9aeda70cc4779885b1277484fe7544a04e3e/9.0.0/platform007/9e92d53/
+LIBGCC_BASE=/mnt/gvfs/third-party2/libgcc/6ace84e956873d53638c738b6f65f3f469cca74c/7.x/platform007/5620abc
+GLIBC_BASE=/mnt/gvfs/third-party2/glibc/192b0f42d63dcf6210d6ceae387b49af049e6e0c/2.26/platform007/f259413
+SNAPPY_BASE=/mnt/gvfs/third-party2/snappy/7f9bdaada18f59bc27ec2b0871eb8a6144343aef/1.1.3/platform007/ca4da3d
+ZLIB_BASE=/mnt/gvfs/third-party2/zlib/2d9f0b9a4274cc21f61272a9e89bdb859bce8f1f/1.2.8/platform007/ca4da3d
+BZIP2_BASE=/mnt/gvfs/third-party2/bzip2/dc49a21c5fceec6456a7a28a94dcd16690af1337/1.0.6/platform007/ca4da3d
+LZ4_BASE=/mnt/gvfs/third-party2/lz4/0f607f8fc442ea7d6b876931b1898bb573d5e5da/1.9.1/platform007/ca4da3d
+ZSTD_BASE=/mnt/gvfs/third-party2/zstd/ca22bc441a4eb709e9e0b1f9fec9750fed7b31c5/1.4.x/platform007/15a3614
+GFLAGS_BASE=/mnt/gvfs/third-party2/gflags/0b9929d2588991c65a57168bf88aff2db87c5d48/2.2.0/platform007/ca4da3d
+JEMALLOC_BASE=/mnt/gvfs/third-party2/jemalloc/c26f08f47ac35fc31da2633b7da92d6b863246eb/master/platform007/c26c002
+NUMA_BASE=/mnt/gvfs/third-party2/numa/3f3fb57a5ccc5fd21c66416c0b83e0aa76a05376/2.0.11/platform007/ca4da3d
+LIBUNWIND_BASE=/mnt/gvfs/third-party2/libunwind/40c73d874898b386a71847f1b99115d93822d11f/1.4/platform007/6f3e0a9
+TBB_BASE=/mnt/gvfs/third-party2/tbb/4ce8e8dba77cdbd81b75d6f0c32fd7a1b76a11ec/2018_U5/platform007/ca4da3d
+KERNEL_HEADERS_BASE=/mnt/gvfs/third-party2/kernel-headers/fb251ecd2f5ae16f8671f7014c246e52a748fe0b/fb/platform007/da39a3e
+BINUTILS_BASE=/mnt/gvfs/third-party2/binutils/ab9f09bba370e7066cafd4eb59752db93f2e8312/2.29.1/platform007/15a3614
+VALGRIND_BASE=/mnt/gvfs/third-party2/valgrind/d42d152a15636529b0861ec493927200ebebca8e/3.15.0/platform007/ca4da3d
+LUA_BASE=/mnt/gvfs/third-party2/lua/f0cd714433206d5139df61659eb7b28b1dea6683/5.3.4/platform007/5007832
@@ -55,22 +55,27 @@ GFLAGS_LIBS=" $TOOLCHAIN_LIB_BASE/gflags/gflags-1.6/c3f970a/lib/libgflags.a"
 JEMALLOC_INCLUDE=" -I $TOOLCHAIN_LIB_BASE/jemalloc/jemalloc-3.4.1/4d53c6f/include/"
 JEMALLOC_LIB=" -Wl,--whole-archive $TOOLCHAIN_LIB_BASE/jemalloc/jemalloc-3.4.1/4d53c6f/lib/libjemalloc.a"

+# location of numa
+NUMA_REV=829d10dac0230f99cd7e1778869d2adf3da24b65
+NUMA_INCLUDE=" -I /mnt/gvfs/third-party2/numa/$NUMA_REV/2.0.8/gcc-4.8.1-glibc-2.17/c3f970a/include/"
+NUMA_LIB=" /mnt/gvfs/third-party2/numa/$NUMA_REV/2.0.8/gcc-4.8.1-glibc-2.17/c3f970a/lib/libnuma.a"
+
 # use Intel SSE support for checksum calculations
 export USE_SSE=" -msse -msse4.2 "

 CC="$TOOLCHAIN_EXECUTABLES/gcc/gcc-4.8.1/cc6c9dc/bin/gcc"
-CXX="$TOOLCHAIN_EXECUTABLES/gcc/gcc-4.8.1/cc6c9dc/bin/g++ $JINCLUDE $SNAPPY_INCLUDE $ZLIB_INCLUDE $BZIP_INCLUDE $LZ4_INCLUDE $GFLAGS_INCLUDE"
+CXX="$TOOLCHAIN_EXECUTABLES/gcc/gcc-4.8.1/cc6c9dc/bin/g++ $JINCLUDE $SNAPPY_INCLUDE $ZLIB_INCLUDE $BZIP_INCLUDE $LZ4_INCLUDE $GFLAGS_INCLUDE $NUMA_INCLUDE"
 AR=$TOOLCHAIN_EXECUTABLES/binutils/binutils-2.21.1/da39a3e/bin/ar
 RANLIB=$TOOLCHAIN_EXECUTABLES/binutils/binutils-2.21.1/da39a3e/bin/ranlib

 CFLAGS="-B$TOOLCHAIN_EXECUTABLES/binutils/binutils-2.21.1/da39a3e/bin/gold -m64 -mtune=generic"
 CFLAGS+=" $LIBGCC_INCLUDE $GLIBC_INCLUDE"
 CFLAGS+=" -DROCKSDB_PLATFORM_POSIX -DROCKSDB_ATOMIC_PRESENT -DROCKSDB_FALLOCATE_PRESENT"
-CFLAGS+=" -DSNAPPY -DGFLAGS=google -DZLIB -DBZIP2 -DLZ4"
+CFLAGS+=" -DSNAPPY -DGFLAGS=google -DZLIB -DBZIP2 -DLZ4 -DNUMA"

 EXEC_LDFLAGS="-Wl,--dynamic-linker,/usr/local/fbcode/gcc-4.8.1-glibc-2.17/lib/ld.so"
 EXEC_LDFLAGS+=" -Wl,--no-whole-archive $TOOLCHAIN_LIB_BASE/libunwind/libunwind-1.0.1/675d945/lib/libunwind.a"
-EXEC_LDFLAGS+=" $HDFSLIB $SNAPPY_LIBS $ZLIB_LIBS $BZIP_LIBS $LZ4_LIBS $GFLAGS_LIBS"
+EXEC_LDFLAGS+=" $HDFSLIB $SNAPPY_LIBS $ZLIB_LIBS $BZIP_LIBS $LZ4_LIBS $GFLAGS_LIBS $NUMA_LIB"

 PLATFORM_LDFLAGS="$LIBGCC_LIBS $GLIBC_LIBS "

@@ -0,0 +1,134 @@
+#!/bin/sh
+#
+# Set environment variables so that we can compile rocksdb using
+# fbcode settings.  It uses the latest g++ and clang compilers and also
+# uses jemalloc
+# Environment variables that change the behavior of this script:
+# PIC_BUILD -- if true, it will only take pic versions of libraries from fbcode. libraries that don't have pic variant will not be included
+
+
+BASEDIR=`dirname $BASH_SOURCE`
+source "$BASEDIR/dependencies.sh"
+
+CFLAGS=""
+
+# libgcc
+LIBGCC_INCLUDE="$LIBGCC_BASE/include/c++/7.3.0"
+LIBGCC_LIBS=" -L $LIBGCC_BASE/lib"
+
+# glibc
+GLIBC_INCLUDE="$GLIBC_BASE/include"
+GLIBC_LIBS=" -L $GLIBC_BASE/lib"
+
+# snappy
+SNAPPY_INCLUDE=" -I $SNAPPY_BASE/include/"
+if test -z $PIC_BUILD; then
+  SNAPPY_LIBS=" $SNAPPY_BASE/lib/libsnappy.a"
+else
+  SNAPPY_LIBS=" $SNAPPY_BASE/lib/libsnappy_pic.a"
+fi
+CFLAGS+=" -DSNAPPY"
+
+if test -z $PIC_BUILD; then
+  # location of zlib headers and libraries
+  ZLIB_INCLUDE=" -I $ZLIB_BASE/include/"
+  ZLIB_LIBS=" $ZLIB_BASE/lib/libz.a"
+  CFLAGS+=" -DZLIB"
+
+  LZ4_INCLUDE=" -I $LZ4_BASE/include/"
+  LZ4_LIBS=" $LZ4_BASE/lib/liblz4.a"
+  CFLAGS+=" -DLZ4"
+fi
+
+# location of gflags headers and libraries
+GFLAGS_INCLUDE=" -I $GFLAGS_BASE/include/"
+if test -z $PIC_BUILD; then
+  GFLAGS_LIBS=" $GFLAGS_BASE/lib/libgflags.a"
+else
+  GFLAGS_LIBS=" $GFLAGS_BASE/lib/libgflags_pic.a"
+fi
+CFLAGS+=" -DGFLAGS=gflags"
+
+# location of jemalloc
+JEMALLOC_INCLUDE=" -I $JEMALLOC_BASE/include/"
+JEMALLOC_LIB=" $JEMALLOC_BASE/lib/libjemalloc.a"
+
+if test -z $PIC_BUILD; then
+  # location of libunwind
+  LIBUNWIND="$LIBUNWIND_BASE/lib/libunwind.a"
+fi
+
+# location of TBB
+TBB_INCLUDE=" -isystem $TBB_BASE/include/"
+if test -z $PIC_BUILD; then
+  TBB_LIBS="$TBB_BASE/lib/libtbb.a"
+else
+  TBB_LIBS="$TBB_BASE/lib/libtbb_pic.a"
+fi
+CFLAGS+=" -DTBB"
+
+# use Intel SSE support for checksum calculations
+export USE_SSE=" -msse -msse4.2 "
+
+BINUTILS="$BINUTILS_BASE/bin"
+AR="$BINUTILS/ar"
+
+DEPS_INCLUDE="$SNAPPY_INCLUDE $ZLIB_INCLUDE $LZ4_INCLUDE $GFLAGS_INCLUDE"
+
+STDLIBS="-L $GCC_BASE/lib64"
+
+CLANG_BIN="$CLANG_BASE/bin"
+CLANG_LIB="$CLANG_BASE/lib"
+CLANG_SRC="$CLANG_BASE/../../src"
+
+CLANG_ANALYZER="$CLANG_BIN/clang++"
+CLANG_SCAN_BUILD="$CLANG_SRC/llvm/tools/clang/tools/scan-build/bin/scan-build"
+
+if [ -z "$USE_CLANG" ]; then
+  # gcc
+  CC="$GCC_BASE/bin/gcc"
+  CXX="$GCC_BASE/bin/g++"
+
+  CFLAGS+=" -B$BINUTILS/gold"
+  CFLAGS+=" -isystem $LIBGCC_INCLUDE"
+  CFLAGS+=" -isystem $GLIBC_INCLUDE"
+  JEMALLOC=1
+else
+  # clang
+  CLANG_INCLUDE="$CLANG_LIB/clang/stable/include"
+  CC="$CLANG_BIN/clang"
+  CXX="$CLANG_BIN/clang++"
+
+  KERNEL_HEADERS_INCLUDE="$KERNEL_HEADERS_BASE/include"
+
+  CFLAGS+=" -B$BINUTILS/gold -nostdinc -nostdlib"
+  CFLAGS+=" -isystem $LIBGCC_BASE/include/c++/7.x "
+  CFLAGS+=" -isystem $LIBGCC_BASE/include/c++/7.x/x86_64-facebook-linux "
+  CFLAGS+=" -isystem $GLIBC_INCLUDE"
+  CFLAGS+=" -isystem $LIBGCC_INCLUDE"
+  CFLAGS+=" -isystem $CLANG_INCLUDE"
+  CFLAGS+=" -isystem $KERNEL_HEADERS_INCLUDE/linux "
+  CFLAGS+=" -isystem $KERNEL_HEADERS_INCLUDE "
+  CFLAGS+=" -Wno-expansion-to-defined "
+  CXXFLAGS="-nostdinc++"
+fi
+
+CFLAGS+=" $DEPS_INCLUDE"
+CFLAGS+=" -DROCKSDB_PLATFORM_POSIX -DROCKSDB_LIB_IO_POSIX -DROCKSDB_FALLOCATE_PRESENT -DROCKSDB_MALLOC_USABLE_SIZE -DROCKSDB_RANGESYNC_PRESENT -DROCKSDB_SCHED_GETCPU_PRESENT -DROCKSDB_SUPPORT_THREAD_LOCAL -DHAVE_SSE42"
+CXXFLAGS+=" $CFLAGS"
+
+EXEC_LDFLAGS=" $SNAPPY_LIBS $ZLIB_LIBS $LZ4_LIBS $ZSTD_LIBS $GFLAGS_LIBS $NUMA_LIB $TBB_LIBS"
+EXEC_LDFLAGS+=" -B$BINUTILS/gold"
+EXEC_LDFLAGS+=" -Wl,--dynamic-linker,/usr/local/fbcode/platform007/lib/ld.so"
+EXEC_LDFLAGS+=" $LIBUNWIND"
+EXEC_LDFLAGS+=" -Wl,-rpath=/usr/local/fbcode/platform007/lib"
+# required by libtbb
+EXEC_LDFLAGS+=" -ldl"
+
+PLATFORM_LDFLAGS="$LIBGCC_LIBS $GLIBC_LIBS $STDLIBS -lgcc -lstdc++"
+
+EXEC_LDFLAGS_SHARED="$SNAPPY_LIBS $ZLIB_LIBS $LZ4_LIBS $ZSTD_LIBS $GFLAGS_LIBS $TBB_LIBS"
+
+VALGRIND_VER="$VALGRIND_BASE/bin/"
+
+export CC CXX AR CFLAGS CXXFLAGS EXEC_LDFLAGS EXEC_LDFLAGS_SHARED VALGRIND_VER JEMALLOC_LIB JEMALLOC_INCLUDE CLANG_ANALYZER CLANG_SCAN_BUILD 
@@ -22,4 +22,4 @@ echo ""
 echo "-----------------------------------------------------------------------------"
 echo "|                         Installation Completed                            |"
 echo "-----------------------------------------------------------------------------"
-echo "Please run `. ~/bash_profile` to be able to compile with gflags"
+echo "Please run \`. ~/.bash_profile\` to be able to compile with gflags"
@@ -109,7 +109,6 @@ make release
    --cache_numshardbits=6 \
    --table_cache_numshardbits=4 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --statistics=1 \
    --histogram=1 \
    --disable_data_sync=1 \
@@ -129,7 +128,6 @@ make release
    --cache_numshardbits=6 \
    --table_cache_numshardbits=4 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --use_tailing_iterator=1 \
    --statistics=1 \
    --histogram=1 \
@@ -150,7 +148,6 @@ make release
    --cache_numshardbits=6 \
    --table_cache_numshardbits=4 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --statistics=1 \
    --histogram=1 \
    --disable_data_sync=1 \
@@ -172,7 +169,6 @@ make release
    --table_cache_numshardbits=4 \
    --write_buffer_size=1000000000 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --statistics=1 \
    --histogram=1 \
    --disable_data_sync=1 \
@@ -231,7 +227,6 @@ make release
    --cache_numshardbits=6 \
    --table_cache_numshardbits=4 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --disable_auto_compactions=1 \
    --statistics=1 \
    --histogram=1 \
@@ -254,7 +249,6 @@ make release
    --cache_numshardbits=6 \
    --table_cache_numshardbits=4 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --statistics=1 \
    --histogram=1 \
    --disable_data_sync=1 \
@@ -274,7 +268,6 @@ make release
    --table_cache_numshardbits=4 \
    --write_buffer_size=1000000000 \
    --open_files=55000 \
-    --disable_seek_compaction=1 \
    --statistics=1 \
    --histogram=1 \
    --disable_data_sync=1 \
@@ -283,6 +276,75 @@ make release
    --value_size=10 \
    --threads=16 > ${STAT_FILE}.memtablefillreadrandom

+common_in_mem_args="--db=/dev/shm/rocksdb \
+    --num_levels=6 \
+    --key_size=20 \
+    --prefix_size=12 \
+    --keys_per_prefix=10 \
+    --value_size=100 \
+    --compression_type=none \
+    --compression_ratio=1 \
+    --hard_rate_limit=2 \
+    --write_buffer_size=134217728 \
+    --max_write_buffer_number=4 \
+    --level0_file_num_compaction_trigger=8 \
+    --level0_slowdown_writes_trigger=16 \
+    --level0_stop_writes_trigger=24 \
+    --target_file_size_base=134217728 \
+    --max_bytes_for_level_base=1073741824 \
+    --disable_wal=0 \
+    --wal_dir=/dev/shm/rocksdb \
+    --sync=0 \
+    --disable_data_sync=1 \
+    --verify_checksum=1 \
+    --delete_obsolete_files_period_micros=314572800 \
+    --max_grandparent_overlap_factor=10 \
+    --use_plain_table=1 \
+    --open_files=-1 \
+    --mmap_read=1 \
+    --mmap_write=0 \
+    --memtablerep=prefix_hash \
+    --bloom_bits=10 \
+    --bloom_locality=1 \
+    --perf_level=0"
+
+# prepare a in-memory DB with 50M keys, total DB size is ~6G
+./db_bench \
+    $common_in_mem_args \
+    --statistics=0 \
+    --max_background_compactions=16 \
+    --max_background_flushes=16 \
+    --benchmarks=filluniquerandom \
+    --use_existing_db=0 \
+    --num=52428800 \
+    --threads=1 > /dev/null
+
+# Readwhilewriting
+./db_bench \
+    $common_in_mem_args \
+    --statistics=1 \
+    --max_background_compactions=4 \
+    --max_background_flushes=0 \
+    --benchmarks=readwhilewriting\
+    --use_existing_db=1 \
+    --duration=600 \
+    --threads=32 \
+    --writes_per_second=81920 > ${STAT_FILE}.readwhilewriting_in_ram
+
+# Seekrandomwhilewriting
+./db_bench \
+    $common_in_mem_args \
+    --statistics=1 \
+    --max_background_compactions=4 \
+    --max_background_flushes=0 \
+    --benchmarks=seekrandomwhilewriting \
+    --use_existing_db=1 \
+    --use_tailing_iterator=1 \
+    --duration=600 \
+    --threads=32 \
+    --writes_per_second=81920 > ${STAT_FILE}.seekwhilewriting_in_ram
+
+
 # send data to ods
 function send_to_ods {
  key="$1"
@@ -308,9 +370,9 @@ function send_benchmark_to_ods {
  file="$3"

  QPS=$(grep $bench $file | awk '{print $5}')
-  P50_MICROS=$(grep $bench $file -A 4 | tail -n1 | awk '{print $3}' )
-  P75_MICROS=$(grep $bench $file -A 4 | tail -n1 | awk '{print $5}' )
-  P99_MICROS=$(grep $bench $file -A 4 | tail -n1 | awk '{print $7}' )
+  P50_MICROS=$(grep $bench $file -A 6 | grep "Percentiles" | awk '{print $3}' )
+  P75_MICROS=$(grep $bench $file -A 6 | grep "Percentiles" | awk '{print $5}' )
+  P99_MICROS=$(grep $bench $file -A 6 | grep "Percentiles" | awk '{print $7}' )

  send_to_ods rocksdb.build.$bench_key.qps $QPS
  send_to_ods rocksdb.build.$bench_key.p50_micros $P50_MICROS
@@ -328,3 +390,5 @@ send_benchmark_to_ods readrandom readrandom_fillunique_random $STAT_FILE.readran
 send_benchmark_to_ods fillrandom memtablefillrandom $STAT_FILE.memtablefillreadrandom
 send_benchmark_to_ods readrandom memtablereadrandom $STAT_FILE.memtablefillreadrandom
 send_benchmark_to_ods readwhilewriting readwhilewriting $STAT_FILE.readwhilewriting
+send_benchmark_to_ods readwhilewriting readwhilewriting_in_ram ${STAT_FILE}.readwhilewriting_in_ram
+send_benchmark_to_ods seekrandomwhilewriting seekwhilewriting_in_ram ${STAT_FILE}.seekwhilewriting_in_ram
@@ -40,9 +40,10 @@ Status BuildTable(const std::string& dbname, Env* env, const Options& options,
                  const InternalKeyComparator& internal_comparator,
                  const SequenceNumber newest_snapshot,
                  const SequenceNumber earliest_seqno_in_memtable,
-                  const CompressionType compression) {
+                  const CompressionType compression,
+                  const Env::IOPriority io_priority) {
  Status s;
-  meta->file_size = 0;
+  meta->fd.file_size = 0;
  meta->smallest_seqno = meta->largest_seqno = 0;
  iter->SeekToFirst();

@@ -54,13 +55,15 @@ Status BuildTable(const std::string& dbname, Env* env, const Options& options,
    purge = false;
  }

-  std::string fname = TableFileName(dbname, meta->number);
+  std::string fname = TableFileName(options.db_paths, meta->fd.GetNumber(),
+                                    meta->fd.GetPathId());
  if (iter->Valid()) {
    unique_ptr<WritableFile> file;
    s = env->NewWritableFile(fname, &file, soptions);
    if (!s.ok()) {
      return s;
    }
+    file->SetIOPriority(io_priority);

    TableBuilder* builder =
        NewTableBuilder(options, internal_comparator, file.get(), compression);
@@ -177,8 +180,8 @@ Status BuildTable(const std::string& dbname, Env* env, const Options& options,
    if (s.ok()) {
      s = builder->Finish();
      if (s.ok()) {
-        meta->file_size = builder->FileSize();
-        assert(meta->file_size > 0);
+        meta->fd.file_size = builder->FileSize();
+        assert(meta->fd.GetFileSize() > 0);
      }
    } else {
      builder->Abandon();
@@ -202,7 +205,7 @@ Status BuildTable(const std::string& dbname, Env* env, const Options& options,
    if (s.ok()) {
      // Verify that the table is usable
      Iterator* it = table_cache->NewIterator(ReadOptions(), soptions,
-                                              internal_comparator, *meta);
+                                              internal_comparator, meta->fd);
      s = it->status();
      delete it;
    }
@@ -213,7 +216,7 @@ Status BuildTable(const std::string& dbname, Env* env, const Options& options,
    s = iter->status();
  }

-  if (s.ok() && meta->file_size > 0) {
+  if (s.ok() && meta->fd.GetFileSize() > 0) {
    // Keep it
  } else {
    env->DeleteFile(fname);
@@ -7,6 +7,7 @@
 // found in the LICENSE file. See the AUTHORS file for names of contributors.
 #pragma once
 #include "rocksdb/comparator.h"
+#include "rocksdb/env.h"
 #include "rocksdb/status.h"
 #include "rocksdb/types.h"
 #include "rocksdb/options.h"
@@ -29,7 +30,7 @@ extern TableBuilder* NewTableBuilder(
    WritableFile* file, CompressionType compression_type);

 // Build a Table file from the contents of *iter.  The generated file
-// will be named according to meta->number.  On success, the rest of
+// will be named according to number specified in meta. On success, the rest of
 // *meta will be filled with metadata about the generated table.
 // If no data is present in *iter, meta->file_size will be set to
 // zero, and no Table file will be produced.
@@ -40,6 +41,7 @@ extern Status BuildTable(const std::string& dbname, Env* env,
                         const InternalKeyComparator& internal_comparator,
                         const SequenceNumber newest_snapshot,
                         const SequenceNumber earliest_seqno_in_memtable,
-                         const CompressionType compression);
+                         const CompressionType compression,
+                         const Env::IOPriority io_priority = Env::IO_HIGH);

 }  // namespace rocksdb
@@ -14,6 +14,7 @@
 #include <stdlib.h>
 #include <unistd.h>
 #include "rocksdb/cache.h"
+#include "rocksdb/compaction_filter.h"
 #include "rocksdb/comparator.h"
 #include "rocksdb/db.h"
 #include "rocksdb/env.h"
@@ -30,9 +31,16 @@
 #include "rocksdb/table.h"

 using rocksdb::Cache;
+using rocksdb::ColumnFamilyDescriptor;
+using rocksdb::ColumnFamilyHandle;
+using rocksdb::ColumnFamilyOptions;
+using rocksdb::CompactionFilter;
+using rocksdb::CompactionFilterFactory;
+using rocksdb::CompactionOptionsFIFO;
 using rocksdb::Comparator;
 using rocksdb::CompressionType;
 using rocksdb::DB;
+using rocksdb::DBOptions;
 using rocksdb::Env;
 using rocksdb::InfoLogLevel;
 using rocksdb::FileLock;
@@ -66,6 +74,7 @@ struct rocksdb_iterator_t        { Iterator*         rep; };
 struct rocksdb_writebatch_t      { WriteBatch        rep; };
 struct rocksdb_snapshot_t        { const Snapshot*   rep; };
 struct rocksdb_flushoptions_t    { FlushOptions      rep; };
+struct rocksdb_fifo_compaction_options_t { CompactionOptionsFIFO rep; };
 struct rocksdb_readoptions_t     { ReadOptions       rep; };
 struct rocksdb_writeoptions_t    { WriteOptions      rep; };
 struct rocksdb_options_t         { Options           rep; };
@@ -76,6 +85,74 @@ struct rocksdb_filelock_t        { FileLock*         rep; };
 struct rocksdb_logger_t          { shared_ptr<Logger>  rep; };
 struct rocksdb_cache_t           { shared_ptr<Cache>   rep; };
 struct rocksdb_livefiles_t       { std::vector<LiveFileMetaData> rep; };
+struct rocksdb_column_family_handle_t  { ColumnFamilyHandle* rep; };
+
+struct rocksdb_compactionfiltercontext_t {
+  CompactionFilter::Context rep;
+};
+
+struct rocksdb_compactionfilter_t : public CompactionFilter {
+  void* state_;
+  void (*destructor_)(void*);
+  unsigned char (*filter_)(
+      void*,
+      int level,
+      const char* key, size_t key_length,
+      const char* existing_value, size_t value_length,
+      char** new_value, size_t *new_value_length,
+      unsigned char* value_changed);
+  const char* (*name_)(void*);
+
+  virtual ~rocksdb_compactionfilter_t() {
+    (*destructor_)(state_);
+  }
+
+  virtual bool Filter(
+      int level,
+      const Slice& key,
+      const Slice& existing_value,
+      std::string* new_value,
+      bool* value_changed) const {
+    char* c_new_value = NULL;
+    size_t new_value_length = 0;
+    unsigned char c_value_changed = 0;
+    unsigned char result = (*filter_)(
+        state_,
+        level,
+        key.data(), key.size(),
+        existing_value.data(), existing_value.size(),
+        &c_new_value, &new_value_length, &c_value_changed);
+    if (c_value_changed) {
+      new_value->assign(c_new_value, new_value_length);
+      *value_changed = true;
+    }
+    return result;
+  }
+
+  virtual const char* Name() const {
+    return (*name_)(state_);
+  }
+};
+
+struct rocksdb_compactionfilterfactory_t : public CompactionFilterFactory {
+  void* state_;
+  void (*destructor_)(void*);
+  rocksdb_compactionfilter_t* (*create_compaction_filter_)(
+      void*, rocksdb_compactionfiltercontext_t* context);
+  const char* (*name_)(void*);
+
+  virtual ~rocksdb_compactionfilterfactory_t() { (*destructor_)(state_); }
+
+  virtual std::unique_ptr<CompactionFilter> CreateCompactionFilter(
+      const CompactionFilter::Context& context) {
+    rocksdb_compactionfiltercontext_t ccontext;
+    ccontext.rep = context;
+    CompactionFilter* cf = (*create_compaction_filter_)(state_, &ccontext);
+    return std::unique_ptr<CompactionFilter>(cf);
+  }
+
+  virtual const char* Name() const { return (*name_)(state_); }
+};

 struct rocksdb_comparator_t : public Comparator {
  void* state_;
@@ -349,6 +426,120 @@ void rocksdb_close(rocksdb_t* db) {
  delete db;
 }

+rocksdb_t* rocksdb_open_column_families(
+    const rocksdb_options_t* db_options,
+    const char* name,
+    int num_column_families,
+    const char** column_family_names,
+    const rocksdb_options_t** column_family_options,
+    rocksdb_column_family_handle_t** column_family_handles,
+    char** errptr) {
+  std::vector<ColumnFamilyDescriptor> column_families;
+  for (int i = 0; i < num_column_families; i++) {
+    column_families.push_back(ColumnFamilyDescriptor(
+        std::string(column_family_names[i]),
+        ColumnFamilyOptions(column_family_options[i]->rep)));
+  }
+
+  DB* db;
+  std::vector<ColumnFamilyHandle*> handles;
+  if (SaveError(errptr, DB::Open(DBOptions(db_options->rep),
+          std::string(name), column_families, &handles, &db))) {
+    return nullptr;
+  }
+
+  for (size_t i = 0; i < handles.size(); i++) {
+    rocksdb_column_family_handle_t* c_handle = new rocksdb_column_family_handle_t;
+    c_handle->rep = handles[i];
+    column_family_handles[i] = c_handle;
+  }
+  rocksdb_t* result = new rocksdb_t;
+  result->rep = db;
+  return result;
+}
+
+rocksdb_t* rocksdb_open_for_read_only_column_families(
+    const rocksdb_options_t* db_options,
+    const char* name,
+    int num_column_families,
+    const char** column_family_names,
+    const rocksdb_options_t** column_family_options,
+    rocksdb_column_family_handle_t** column_family_handles,
+    unsigned char error_if_log_file_exist,
+    char** errptr) {
+  std::vector<ColumnFamilyDescriptor> column_families;
+  for (int i = 0; i < num_column_families; i++) {
+    column_families.push_back(ColumnFamilyDescriptor(
+        std::string(column_family_names[i]),
+        ColumnFamilyOptions(column_family_options[i]->rep)));
+  }
+
+  DB* db;
+  std::vector<ColumnFamilyHandle*> handles;
+  if (SaveError(errptr, DB::OpenForReadOnly(DBOptions(db_options->rep),
+          std::string(name), column_families, &handles, &db, error_if_log_file_exist))) {
+    return nullptr;
+  }
+
+  for (size_t i = 0; i < handles.size(); i++) {
+    rocksdb_column_family_handle_t* c_handle = new rocksdb_column_family_handle_t;
+    c_handle->rep = handles[i];
+    column_family_handles[i] = c_handle;
+  }
+  rocksdb_t* result = new rocksdb_t;
+  result->rep = db;
+  return result;
+}
+
+char** rocksdb_list_column_families(
+    const rocksdb_options_t* options,
+    const char* name,
+    size_t* lencfs,
+    char** errptr) {
+  std::vector<std::string> fams;
+  SaveError(errptr,
+      DB::ListColumnFamilies(DBOptions(options->rep),
+        std::string(name), &fams));
+
+  *lencfs = fams.size();
+  char** column_families = static_cast<char**>(malloc(sizeof(char*) * fams.size()));
+  for (size_t i = 0; i < fams.size(); i++) {
+    column_families[i] = strdup(fams[i].c_str());
+  }
+  return column_families;
+}
+
+void rocksdb_list_column_families_destroy(char** list, size_t len) {
+  for (size_t i = 0; i < len; ++i) {
+    free(list[i]);
+  }
+  free(list);
+}
+
+rocksdb_column_family_handle_t* rocksdb_create_column_family(
+    rocksdb_t* db,
+    const rocksdb_options_t* column_family_options,
+    const char* column_family_name,
+    char** errptr) {
+  rocksdb_column_family_handle_t* handle = new rocksdb_column_family_handle_t;
+  SaveError(errptr,
+      db->rep->CreateColumnFamily(ColumnFamilyOptions(column_family_options->rep),
+        std::string(column_family_name), &(handle->rep)));
+  return handle;
+}
+
+void rocksdb_drop_column_family(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* handle,
+    char** errptr) {
+  SaveError(errptr, db->rep->DropColumnFamily(handle->rep));
+}
+
+void rocksdb_column_family_handle_destroy(rocksdb_column_family_handle_t* handle) {
+  delete handle->rep;
+  delete handle;
+}
+
 void rocksdb_put(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -359,6 +550,18 @@ void rocksdb_put(
            db->rep->Put(options->rep, Slice(key, keylen), Slice(val, vallen)));
 }

+void rocksdb_put_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    const char* val, size_t vallen,
+    char** errptr) {
+  SaveError(errptr,
+            db->rep->Put(options->rep, column_family->rep,
+              Slice(key, keylen), Slice(val, vallen)));
+}
+
 void rocksdb_delete(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -367,6 +570,16 @@ void rocksdb_delete(
  SaveError(errptr, db->rep->Delete(options->rep, Slice(key, keylen)));
 }

+void rocksdb_delete_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    char** errptr) {
+  SaveError(errptr, db->rep->Delete(options->rep, column_family->rep,
+        Slice(key, keylen)));
+}
+
 void rocksdb_merge(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -377,6 +590,18 @@ void rocksdb_merge(
            db->rep->Merge(options->rep, Slice(key, keylen), Slice(val, vallen)));
 }

+void rocksdb_merge_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    const char* val, size_t vallen,
+    char** errptr) {
+  SaveError(errptr,
+            db->rep->Merge(options->rep, column_family->rep,
+              Slice(key, keylen), Slice(val, vallen)));
+}
+
 void rocksdb_write(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -406,6 +631,29 @@ char* rocksdb_get(
  return result;
 }

+char* rocksdb_get_cf(
+    rocksdb_t* db,
+    const rocksdb_readoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    size_t* vallen,
+    char** errptr) {
+  char* result = nullptr;
+  std::string tmp;
+  Status s = db->rep->Get(options->rep, column_family->rep,
+      Slice(key, keylen), &tmp);
+  if (s.ok()) {
+    *vallen = tmp.size();
+    result = CopyString(tmp);
+  } else {
+    *vallen = 0;
+    if (!s.IsNotFound()) {
+      SaveError(errptr, s);
+    }
+  }
+  return result;
+}
+
 rocksdb_iterator_t* rocksdb_create_iterator(
    rocksdb_t* db,
    const rocksdb_readoptions_t* options) {
@@ -414,6 +662,15 @@ rocksdb_iterator_t* rocksdb_create_iterator(
  return result;
 }

+rocksdb_iterator_t* rocksdb_create_iterator_cf(
+    rocksdb_t* db,
+    const rocksdb_readoptions_t* options,
+    rocksdb_column_family_handle_t* column_family) {
+  rocksdb_iterator_t* result = new rocksdb_iterator_t;
+  result->rep = db->rep->NewIterator(options->rep, column_family->rep);
+  return result;
+}
+
 const rocksdb_snapshot_t* rocksdb_create_snapshot(
    rocksdb_t* db) {
  rocksdb_snapshot_t* result = new rocksdb_snapshot_t;
@@ -440,6 +697,19 @@ char* rocksdb_property_value(
  }
 }

+char* rocksdb_property_value_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    const char* propname) {
+  std::string tmp;
+  if (db->rep->GetProperty(column_family->rep, Slice(propname), &tmp)) {
+    // We use strdup() since we expect human readable output.
+    return strdup(tmp.c_str());
+  } else {
+    return nullptr;
+  }
+}
+
 void rocksdb_approximate_sizes(
    rocksdb_t* db,
    int num_ranges,
@@ -455,6 +725,22 @@ void rocksdb_approximate_sizes(
  delete[] ranges;
 }

+void rocksdb_approximate_sizes_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    int num_ranges,
+    const char* const* range_start_key, const size_t* range_start_key_len,
+    const char* const* range_limit_key, const size_t* range_limit_key_len,
+    uint64_t* sizes) {
+  Range* ranges = new Range[num_ranges];
+  for (int i = 0; i < num_ranges; i++) {
+    ranges[i].start = Slice(range_start_key[i], range_start_key_len[i]);
+    ranges[i].limit = Slice(range_limit_key[i], range_limit_key_len[i]);
+  }
+  db->rep->GetApproximateSizes(column_family->rep, ranges, num_ranges, sizes);
+  delete[] ranges;
+}
+
 void rocksdb_delete_file(
    rocksdb_t* db,
    const char* name) {
@@ -479,6 +765,19 @@ void rocksdb_compact_range(
      (limit_key ? (b = Slice(limit_key, limit_key_len), &b) : nullptr));
 }

+void rocksdb_compact_range_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    const char* start_key, size_t start_key_len,
+    const char* limit_key, size_t limit_key_len) {
+  Slice a, b;
+  db->rep->CompactRange(
+      column_family->rep,
+      // Pass nullptr Slice if corresponding "const char*" is nullptr
+      (start_key ? (a = Slice(start_key, start_key_len), &a) : nullptr),
+      (limit_key ? (b = Slice(limit_key, limit_key_len), &b) : nullptr));
+}
+
 void rocksdb_flush(
    rocksdb_t* db,
    const rocksdb_flushoptions_t* options,
@@ -562,6 +861,13 @@ rocksdb_writebatch_t* rocksdb_writebatch_create() {
  return new rocksdb_writebatch_t;
 }

+rocksdb_writebatch_t* rocksdb_writebatch_create_from(const char* rep,
+                                                     size_t size) {
+  rocksdb_writebatch_t* b = new rocksdb_writebatch_t;
+  b->rep = WriteBatch(std::string(rep, size));
+  return b;
+}
+
 void rocksdb_writebatch_destroy(rocksdb_writebatch_t* b) {
  delete b;
 }
@@ -581,6 +887,14 @@ void rocksdb_writebatch_put(
  b->rep.Put(Slice(key, klen), Slice(val, vlen));
 }

+void rocksdb_writebatch_put_cf(
+    rocksdb_writebatch_t* b,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen,
+    const char* val, size_t vlen) {
+  b->rep.Put(column_family->rep, Slice(key, klen), Slice(val, vlen));
+}
+
 void rocksdb_writebatch_merge(
    rocksdb_writebatch_t* b,
    const char* key, size_t klen,
@@ -588,12 +902,27 @@ void rocksdb_writebatch_merge(
  b->rep.Merge(Slice(key, klen), Slice(val, vlen));
 }

+void rocksdb_writebatch_merge_cf(
+    rocksdb_writebatch_t* b,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen,
+    const char* val, size_t vlen) {
+  b->rep.Merge(column_family->rep, Slice(key, klen), Slice(val, vlen));
+}
+
 void rocksdb_writebatch_delete(
    rocksdb_writebatch_t* b,
    const char* key, size_t klen) {
  b->rep.Delete(Slice(key, klen));
 }

+void rocksdb_writebatch_delete_cf(
+    rocksdb_writebatch_t* b,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen) {
+  b->rep.Delete(column_family->rep, Slice(key, klen));
+}
+
 void rocksdb_writebatch_iterate(
    rocksdb_writebatch_t* b,
    void* state,
@@ -631,6 +960,38 @@ void rocksdb_options_destroy(rocksdb_options_t* options) {
  delete options;
 }

+void rocksdb_options_increase_parallelism(
+    rocksdb_options_t* opt, int total_threads) {
+  opt->rep.IncreaseParallelism(total_threads);
+}
+
+void rocksdb_options_optimize_for_point_lookup(
+    rocksdb_options_t* opt) {
+  opt->rep.OptimizeForPointLookup();
+}
+
+void rocksdb_options_optimize_level_style_compaction(
+    rocksdb_options_t* opt, uint64_t memtable_memory_budget) {
+  opt->rep.OptimizeLevelStyleCompaction(memtable_memory_budget);
+}
+
+void rocksdb_options_optimize_universal_style_compaction(
+    rocksdb_options_t* opt, uint64_t memtable_memory_budget) {
+  opt->rep.OptimizeUniversalStyleCompaction(memtable_memory_budget);
+}
+
+void rocksdb_options_set_compaction_filter(
+    rocksdb_options_t* opt,
+    rocksdb_compactionfilter_t* filter) {
+  opt->rep.compaction_filter = filter;
+}
+
+void rocksdb_options_set_compaction_filter_factory(
+    rocksdb_options_t* opt, rocksdb_compactionfilterfactory_t* factory) {
+  opt->rep.compaction_filter_factory =
+      std::shared_ptr<CompactionFilterFactory>(factory);
+}
+
 void rocksdb_options_set_comparator(
    rocksdb_options_t* opt,
    rocksdb_comparator_t* cmp) {
@@ -654,6 +1015,11 @@ void rocksdb_options_set_create_if_missing(
  opt->rep.create_if_missing = v;
 }

+void rocksdb_options_set_create_missing_column_families(
+    rocksdb_options_t* opt, unsigned char v) {
+  opt->rep.create_missing_column_families = v;
+}
+
 void rocksdb_options_set_error_if_exists(
    rocksdb_options_t* opt, unsigned char v) {
  opt->rep.error_if_exists = v;
@@ -1107,6 +1473,12 @@ void rocksdb_options_set_universal_compaction_options(rocksdb_options_t *opt, ro
  opt->rep.compaction_options_universal = *(uco->rep);
 }

+void rocksdb_options_set_fifo_compaction_options(
+    rocksdb_options_t* opt,
+    rocksdb_fifo_compaction_options_t* fifo) {
+  opt->rep.compaction_options_fifo = fifo->rep;
+}
+
 /*
 TODO:
 DB::OpenForReadOnly
@@ -1119,10 +1491,61 @@ DB::GetUpdatesSince
 DB::GetDbIdentity
 DB::RunManualCompaction
 custom cache
-compaction_filter
 table_properties_collectors
 */

+rocksdb_compactionfilter_t* rocksdb_compactionfilter_create(
+    void* state,
+    void (*destructor)(void*),
+    unsigned char (*filter)(
+        void*,
+        int level,
+        const char* key, size_t key_length,
+        const char* existing_value, size_t value_length,
+        char** new_value, size_t *new_value_length,
+        unsigned char* value_changed),
+    const char* (*name)(void*)) {
+  rocksdb_compactionfilter_t* result = new rocksdb_compactionfilter_t;
+  result->state_ = state;
+  result->destructor_ = destructor;
+  result->filter_ = filter;
+  result->name_ = name;
+  return result;
+}
+
+void rocksdb_compactionfilter_destroy(rocksdb_compactionfilter_t* filter) {
+  delete filter;
+}
+
+unsigned char rocksdb_compactionfiltercontext_is_full_compaction(
+    rocksdb_compactionfiltercontext_t* context) {
+  return context->rep.is_full_compaction;
+}
+
+unsigned char rocksdb_compactionfiltercontext_is_manual_compaction(
+    rocksdb_compactionfiltercontext_t* context) {
+  return context->rep.is_manual_compaction;
+}
+
+rocksdb_compactionfilterfactory_t* rocksdb_compactionfilterfactory_create(
+    void* state, void (*destructor)(void*),
+    rocksdb_compactionfilter_t* (*create_compaction_filter)(
+        void*, rocksdb_compactionfiltercontext_t* context),
+    const char* (*name)(void*)) {
+  rocksdb_compactionfilterfactory_t* result =
+      new rocksdb_compactionfilterfactory_t;
+  result->state_ = state;
+  result->destructor_ = destructor;
+  result->create_compaction_filter_ = create_compaction_filter;
+  result->name_ = name;
+  return result;
+}
+
+void rocksdb_compactionfilterfactory_destroy(
+    rocksdb_compactionfilterfactory_t* factory) {
+  delete factory;
+}
+
 rocksdb_comparator_t* rocksdb_comparator_create(
    void* state,
    void (*destructor)(void*),
@@ -1414,6 +1837,22 @@ void rocksdb_universal_compaction_options_destroy(
  delete uco;
 }

+rocksdb_fifo_compaction_options_t* rocksdb_fifo_compaction_options_create() {
+  rocksdb_fifo_compaction_options_t* result = new rocksdb_fifo_compaction_options_t;
+  result->rep =  CompactionOptionsFIFO();
+  return result;
+}
+
+void rocksdb_fifo_compaction_options_set_max_table_files_size(
+    rocksdb_fifo_compaction_options_t* fifo_opts, uint64_t size) {
+  fifo_opts->rep.max_table_files_size = size;
+}
+
+void rocksdb_fifo_compaction_options_destroy(
+    rocksdb_fifo_compaction_options_t* fifo_opts) {
+  delete fifo_opts;
+}
+
 void rocksdb_options_set_min_level_to_compress(rocksdb_options_t* opt, int level) {
  if (level >= 0) {
    assert(level <= opt->rep.num_levels);
@@ -75,6 +75,22 @@ static void CheckGet(
  Free(&val);
 }

+static void CheckGetCF(
+    rocksdb_t* db,
+    const rocksdb_readoptions_t* options,
+    rocksdb_column_family_handle_t* handle,
+    const char* key,
+    const char* expected) {
+  char* err = NULL;
+  size_t val_len;
+  char* val;
+  val = rocksdb_get_cf(db, options, handle, key, strlen(key), &val_len, &err);
+  CheckNoError(err);
+  CheckEqual(expected, val, val_len);
+  Free(&val);
+}
+
+
 static void CheckIter(rocksdb_iterator_t* iter,
                      const char* key, const char* val) {
  size_t len;
@@ -154,6 +170,61 @@ static unsigned char FilterKeyMatch(
  return fake_filter_result;
 }

+// Custom compaction filter
+static void CFilterDestroy(void* arg) {}
+static const char* CFilterName(void* arg) { return "foo"; }
+static unsigned char CFilterFilter(void* arg, int level, const char* key,
+                                   size_t key_length,
+                                   const char* existing_value,
+                                   size_t value_length, char** new_value,
+                                   size_t* new_value_length,
+                                   unsigned char* value_changed) {
+  if (key_length == 3) {
+    if (memcmp(key, "bar", key_length) == 0) {
+      return 1;
+    } else if (memcmp(key, "baz", key_length) == 0) {
+      *value_changed = 1;
+      *new_value = "newbazvalue";
+      *new_value_length = 11;
+      return 0;
+    }
+  }
+  return 0;
+}
+
+static void CFilterFactoryDestroy(void* arg) {}
+static const char* CFilterFactoryName(void* arg) { return "foo"; }
+static rocksdb_compactionfilter_t* CFilterCreate(
+    void* arg, rocksdb_compactionfiltercontext_t* context) {
+  return rocksdb_compactionfilter_create(NULL, CFilterDestroy, CFilterFilter,
+                                         CFilterName);
+}
+
+static rocksdb_t* CheckCompaction(rocksdb_t* db, rocksdb_options_t* options,
+                                  rocksdb_readoptions_t* roptions,
+                                  rocksdb_writeoptions_t* woptions) {
+  char* err = NULL;
+  db = rocksdb_open(options, dbname, &err);
+  CheckNoError(err);
+  rocksdb_put(db, woptions, "foo", 3, "foovalue", 8, &err);
+  CheckNoError(err);
+  CheckGet(db, roptions, "foo", "foovalue");
+  rocksdb_put(db, woptions, "bar", 3, "barvalue", 8, &err);
+  CheckNoError(err);
+  CheckGet(db, roptions, "bar", "barvalue");
+  rocksdb_put(db, woptions, "baz", 3, "bazvalue", 8, &err);
+  CheckNoError(err);
+  CheckGet(db, roptions, "baz", "bazvalue");
+
+  // Force compaction
+  rocksdb_compact_range(db, NULL, 0, NULL, 0);
+  // should have filtered bar, but not foo
+  CheckGet(db, roptions, "foo", "foovalue");
+  CheckGet(db, roptions, "bar", NULL);
+  CheckGet(db, roptions, "baz", "newbazvalue");
+  return db;
+}
+
 // Custom merge operator
 static void MergeOperatorDestroy(void* arg) { }
 static const char* MergeOperatorName(void* arg) {
@@ -277,6 +348,24 @@ int main(int argc, char** argv) {
    rocksdb_writebatch_destroy(wb);
  }

+  StartPhase("writebatch_rep");
+  {
+    rocksdb_writebatch_t* wb1 = rocksdb_writebatch_create();
+    rocksdb_writebatch_put(wb1, "baz", 3, "d", 1);
+    rocksdb_writebatch_put(wb1, "quux", 4, "e", 1);
+    rocksdb_writebatch_delete(wb1, "quux", 4);
+    size_t repsize1 = 0;
+    const char* rep = rocksdb_writebatch_data(wb1, &repsize1);
+    rocksdb_writebatch_t* wb2 = rocksdb_writebatch_create_from(rep, repsize1);
+    CheckCondition(rocksdb_writebatch_count(wb1) ==
+                   rocksdb_writebatch_count(wb2));
+    size_t repsize2 = 0;
+    CheckCondition(
+        memcmp(rep, rocksdb_writebatch_data(wb2, &repsize2), repsize1) == 0);
+    rocksdb_writebatch_destroy(wb1);
+    rocksdb_writebatch_destroy(wb2);
+  }
+
  StartPhase("iter");
  {
    rocksdb_iterator_t* iter = rocksdb_create_iterator(db, roptions);
@@ -407,6 +496,41 @@ int main(int argc, char** argv) {
    rocksdb_filterpolicy_destroy(policy);
  }

+  StartPhase("compaction_filter");
+  {
+    rocksdb_options_t* options = rocksdb_options_create();
+    rocksdb_options_set_create_if_missing(options, 1);
+    rocksdb_compactionfilter_t* cfilter;
+    cfilter = rocksdb_compactionfilter_create(NULL, CFilterDestroy,
+                                              CFilterFilter, CFilterName);
+    // Create new database
+    rocksdb_close(db);
+    rocksdb_destroy_db(options, dbname, &err);
+    rocksdb_options_set_compaction_filter(options, cfilter);
+    db = CheckCompaction(db, options, roptions, woptions);
+
+    rocksdb_options_set_compaction_filter(options, NULL);
+    rocksdb_compactionfilter_destroy(cfilter);
+    rocksdb_options_destroy(options);
+  }
+
+  StartPhase("compaction_filter_factory");
+  {
+    rocksdb_options_t* options = rocksdb_options_create();
+    rocksdb_options_set_create_if_missing(options, 1);
+    rocksdb_compactionfilterfactory_t* factory;
+    factory = rocksdb_compactionfilterfactory_create(
+        NULL, CFilterFactoryDestroy, CFilterCreate, CFilterFactoryName);
+    // Create new database
+    rocksdb_close(db);
+    rocksdb_destroy_db(options, dbname, &err);
+    rocksdb_options_set_compaction_filter_factory(options, factory);
+    db = CheckCompaction(db, options, roptions, woptions);
+
+    rocksdb_options_set_compaction_filter_factory(options, NULL);
+    rocksdb_options_destroy(options);
+  }
+
  StartPhase("merge_operator");
  {
    rocksdb_mergeoperator_t* merge_operator;
@@ -433,16 +557,93 @@ int main(int argc, char** argv) {

  }

+  StartPhase("columnfamilies");
+  {
+    rocksdb_close(db);
+    rocksdb_destroy_db(options, dbname, &err);
+    CheckNoError(err)
+
+    rocksdb_options_t* db_options = rocksdb_options_create();
+    rocksdb_options_set_create_if_missing(db_options, 1);
+    db = rocksdb_open(db_options, dbname, &err);
+    CheckNoError(err)
+    rocksdb_column_family_handle_t* cfh;
+    cfh = rocksdb_create_column_family(db, db_options, "cf1", &err);
+    rocksdb_column_family_handle_destroy(cfh);
+    CheckNoError(err);
+    rocksdb_close(db);
+
+    size_t cflen;
+    char** column_fams = rocksdb_list_column_families(db_options, dbname, &cflen, &err);
+    CheckNoError(err);
+    CheckEqual("default", column_fams[0], 7);
+    CheckEqual("cf1", column_fams[1], 3);
+    CheckCondition(cflen == 2);
+    rocksdb_list_column_families_destroy(column_fams, cflen);
+
+    rocksdb_options_t* cf_options = rocksdb_options_create();
+
+    const char* cf_names[2] = {"default", "cf1"};
+    const rocksdb_options_t* cf_opts[2] = {cf_options, cf_options};
+    rocksdb_column_family_handle_t* handles[2];
+    db = rocksdb_open_column_families(db_options, dbname, 2, cf_names, cf_opts, handles, &err);
+    CheckNoError(err);
+
+    rocksdb_put_cf(db, woptions, handles[1], "foo", 3, "hello", 5, &err);
+    CheckNoError(err);
+
+    CheckGetCF(db, roptions, handles[1], "foo", "hello");
+
+    rocksdb_delete_cf(db, woptions, handles[1], "foo", 3, &err);
+    CheckNoError(err);
+
+    CheckGetCF(db, roptions, handles[1], "foo", NULL);
+
+    rocksdb_writebatch_t* wb = rocksdb_writebatch_create();
+    rocksdb_writebatch_put_cf(wb, handles[1], "baz", 3, "a", 1);
+    rocksdb_writebatch_clear(wb);
+    rocksdb_writebatch_put_cf(wb, handles[1], "bar", 3, "b", 1);
+    rocksdb_writebatch_put_cf(wb, handles[1], "box", 3, "c", 1);
+    rocksdb_writebatch_delete_cf(wb, handles[1], "bar", 3);
+    rocksdb_write(db, woptions, wb, &err);
+    CheckNoError(err);
+    CheckGetCF(db, roptions, handles[1], "baz", NULL);
+    CheckGetCF(db, roptions, handles[1], "bar", NULL);
+    CheckGetCF(db, roptions, handles[1], "box", "c");
+    rocksdb_writebatch_destroy(wb);
+
+    rocksdb_iterator_t* iter = rocksdb_create_iterator_cf(db, roptions, handles[1]);
+    CheckCondition(!rocksdb_iter_valid(iter));
+    rocksdb_iter_seek_to_first(iter);
+    CheckCondition(rocksdb_iter_valid(iter));
+
+    int i;
+    for (i = 0; rocksdb_iter_valid(iter) != 0; rocksdb_iter_next(iter)) {
+      i++;
+    }
+    CheckCondition(i == 1);
+    rocksdb_iter_get_error(iter, &err);
+    CheckNoError(err);
+    rocksdb_iter_destroy(iter);
+
+    rocksdb_drop_column_family(db, handles[1], &err);
+    CheckNoError(err);
+    for (i = 0; i < 2; i++) {
+      rocksdb_column_family_handle_destroy(handles[i]);
+    }
+    rocksdb_close(db);
+    rocksdb_destroy_db(options, dbname, &err);
+    rocksdb_options_destroy(db_options);
+    rocksdb_options_destroy(cf_options);
+  }
+
  StartPhase("prefix");
  {
    // Create new database
-    rocksdb_close(db);
-    rocksdb_destroy_db(options, dbname, &err);
-
    rocksdb_filterpolicy_t* policy = rocksdb_filterpolicy_create_bloom(10);
    rocksdb_options_set_filter_policy(options, policy);
    rocksdb_options_set_prefix_extractor(options, rocksdb_slicetransform_create_fixed_prefix(3));
-    rocksdb_options_set_hash_skip_list_rep(options, 50000, 4, 4);
+    rocksdb_options_set_hash_skip_list_rep(options, 5000, 4, 4);
    rocksdb_options_set_plain_table_factory(options, 4, 10, 0.75, 16);

    db = rocksdb_open(options, dbname, &err);
@@ -480,6 +681,7 @@ int main(int argc, char** argv) {
    rocksdb_filterpolicy_destroy(policy);
  }

+
  StartPhase("cleanup");
  rocksdb_close(db);
  rocksdb_options_destroy(options);
@@ -92,6 +92,9 @@ ColumnFamilyOptions SanitizeOptions(const InternalKeyComparator* icmp,
  if (result.soft_rate_limit > result.hard_rate_limit) {
    result.soft_rate_limit = result.hard_rate_limit;
  }
+  if (result.max_write_buffer_number < 2) {
+    result.max_write_buffer_number = 2;
+  }
  if (!result.prefix_extractor) {
    assert(result.memtable_factory);
    Slice name = result.memtable_factory->Name();
@@ -221,8 +224,7 @@ ColumnFamilyData::ColumnFamilyData(const std::string& dbname, uint32_t id,
  if (dummy_versions != nullptr) {
    internal_stats_.reset(new InternalStats(
        options_.num_levels, db_options->env, db_options->statistics.get()));
-    table_cache_.reset(
-        new TableCache(dbname, &options_, storage_options, table_cache));
+    table_cache_.reset(new TableCache(&options_, storage_options, table_cache));
    if (options_.compaction_style == kCompactionStyleUniversal) {
      compaction_picker_.reset(
          new UniversalCompactionPicker(&options_, &internal_comparator_));
@@ -240,6 +242,8 @@ ColumnFamilyData::ColumnFamilyData(const std::string& dbname, uint32_t id,
    const ColumnFamilyOptions* cf_options = &options_;
    cf_options->Dump(options_.info_log.get());
  }
+
+  RecalculateWriteStallConditions();
 }

 // DB mutex held
@@ -292,6 +296,35 @@ ColumnFamilyData::~ColumnFamilyData() {
  }
 }

+void ColumnFamilyData::RecalculateWriteStallConditions() {
+  need_wait_for_num_memtables_ =
+    (imm()->size() == options()->max_write_buffer_number - 1);
+
+  if (current_ != nullptr) {
+    need_wait_for_num_level0_files_ =
+      (current_->NumLevelFiles(0) >= options()->level0_stop_writes_trigger);
+  } else {
+    need_wait_for_num_level0_files_ = false;
+  }
+
+  RecalculateWriteStallRateLimitsConditions();
+}
+
+void ColumnFamilyData::RecalculateWriteStallRateLimitsConditions() {
+  if (current_ != nullptr) {
+    exceeds_hard_rate_limit_ =
+        (options()->hard_rate_limit > 1.0 &&
+         current_->MaxCompactionScore() > options()->hard_rate_limit);
+
+    exceeds_soft_rate_limit_ =
+        (options()->soft_rate_limit > 0.0 &&
+         current_->MaxCompactionScore() > options()->soft_rate_limit);
+  } else {
+    exceeds_hard_rate_limit_ = false;
+    exceeds_soft_rate_limit_ = false;
+  }
+}
+
 const EnvOptions* ColumnFamilyData::soptions() const {
  return &(column_family_set_->storage_options_);
 }
@@ -313,7 +346,9 @@ void ColumnFamilyData::CreateNewMemtable() {
 }

 Compaction* ColumnFamilyData::PickCompaction(LogBuffer* log_buffer) {
-  return compaction_picker_->PickCompaction(current_, log_buffer);
+  auto result = compaction_picker_->PickCompaction(current_, log_buffer);
+  RecalculateWriteStallRateLimitsConditions();
+  return result;
 }

 Compaction* ColumnFamilyData::CompactRange(int input_level, int output_level,
@@ -417,6 +452,9 @@ SuperVersion* ColumnFamilyData::InstallSuperVersion(
  if (column_family_set_->db_options_->allow_thread_local) {
    ResetThreadLocalSuperVersions();
  }
+
+  RecalculateWriteStallConditions();
+
  if (old_superversion != nullptr && old_superversion->Unref()) {
    old_superversion->Cleanup();
    return old_superversion;  // will let caller delete outside of mutex
@@ -506,6 +544,10 @@ void ColumnFamilySet::UpdateMaxColumnFamily(uint32_t new_max_column_family) {
  max_column_family_ = std::max(new_max_column_family, max_column_family_);
 }

+size_t ColumnFamilySet::NumberOfColumnFamilies() const {
+  return column_families_.size();
+}
+
 // under a DB mutex
 ColumnFamilyData* ColumnFamilySet::CreateColumnFamily(
    const std::string& name, uint32_t id, Version* dummy_versions,
@@ -180,7 +180,7 @@ class ColumnFamilyData {
  void SetCurrent(Version* current);
  void CreateNewMemtable();

-  TableCache* table_cache() { return table_cache_.get(); }
+  TableCache* table_cache() const { return table_cache_.get(); }

  // See documentation in compaction_picker.h
  Compaction* PickCompaction(LogBuffer* log_buffer);
@@ -229,6 +229,22 @@ class ColumnFamilyData {
    return need_slowdown_for_num_level0_files_;
  }

+  bool NeedWaitForNumLevel0Files() const {
+    return need_wait_for_num_level0_files_;
+  }
+
+  bool NeedWaitForNumMemtables() const {
+    return need_wait_for_num_memtables_;
+  }
+
+  bool ExceedsSoftRateLimit() const {
+    return exceeds_soft_rate_limit_;
+  }
+
+  bool ExceedsHardRateLimit() const {
+    return exceeds_hard_rate_limit_;
+  }
+
 private:
  friend class ColumnFamilySet;
  ColumnFamilyData(const std::string& dbname, uint32_t id,
@@ -238,6 +254,14 @@ class ColumnFamilyData {
                   const EnvOptions& storage_options,
                   ColumnFamilySet* column_family_set);

+  // Recalculate some small conditions, which are changed only during
+  // compaction, adding new memtable and/or
+  // recalculation of compaction score. These values are used in
+  // DBImpl::MakeRoomForWrite function to decide, if it need to make
+  // a write stall
+  void RecalculateWriteStallConditions();
+  void RecalculateWriteStallRateLimitsConditions();
+
  uint32_t id_;
  const std::string name_;
  Version* dummy_versions_;  // Head of circular doubly-linked list of versions.
@@ -282,6 +306,22 @@ class ColumnFamilyData {
  // we have too many level 0 files
  bool need_slowdown_for_num_level0_files_;

+  // These 4 variables are updated only after compaction,
+  // adding new memtable, flushing memtables to files
+  // and/or add recalculation of compaction score.
+  // That's why theirs values are cached in ColumnFamilyData.
+  // Recalculation is made by RecalculateWriteStallConditions and
+  // RecalculateWriteStallRateLimitsConditions function. They are used
+  // in DBImpl::MakeRoomForWrite function to decide, if it need
+  // to sleep during write operation
+  bool need_wait_for_num_memtables_;
+
+  bool need_wait_for_num_level0_files_;
+
+  bool exceeds_hard_rate_limit_;
+
+  bool exceeds_soft_rate_limit_;
+
  // An object that keeps all the compaction stats
  // and picks the next compaction
  std::unique_ptr<CompactionPicker> compaction_picker_;
@@ -302,8 +342,8 @@ class ColumnFamilyData {
 // family might get dropped when the DB mutex is released
 // * GetDefault() -- thread safe
 // * GetColumnFamily() -- either inside of DB mutex or call Lock() <-> Unlock()
-// * GetNextColumnFamilyID(), GetMaxColumnFamily(), UpdateMaxColumnFamily() --
-// inside of DB mutex
+// * GetNextColumnFamilyID(), GetMaxColumnFamily(), UpdateMaxColumnFamily(),
+// NumberOfColumnFamilies -- inside of DB mutex
 class ColumnFamilySet {
 public:
  // ColumnFamilySet supports iteration
@@ -342,6 +382,7 @@ class ColumnFamilySet {
  uint32_t GetNextColumnFamilyID();
  uint32_t GetMaxColumnFamily();
  void UpdateMaxColumnFamily(uint32_t new_max_column_family);
+  size_t NumberOfColumnFamilies() const;

  ColumnFamilyData* CreateColumnFamily(const std::string& name, uint32_t id,
                                       Version* dummy_version,
@@ -749,7 +749,6 @@ TEST(ColumnFamilyTest, DifferentCompactionStyles) {
  default_cf.filter_policy = nullptr;
  default_cf.no_block_cache = true;
  default_cf.source_compaction_factor = 100;
-  default_cf.disable_seek_compaction = false;

  one.compaction_style = kCompactionStyleUniversal;
  // trigger compaction if there are >= 4 files
@@ -764,33 +763,6 @@ TEST(ColumnFamilyTest, DifferentCompactionStyles) {

  Reopen({default_cf, one, two});

-  // SETUP column family "default" - test read compaction
-  ASSERT_EQ("", FilesPerLevel(0));
-  PutRandomData(0, 1, 4096);
-  ASSERT_OK(Flush(0));
-  ASSERT_EQ("0,0,1", FilesPerLevel(0));
-  // write 8MB
-  PutRandomData(0, 2000, 4096);
-  ASSERT_OK(Flush(0));
-  // clear levels 0 and 1
-  dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[0]);
-  dbfull()->TEST_CompactRange(1, nullptr, nullptr, handles_[0]);
-  ASSERT_EQ(NumTableFilesAtLevel(0, 0), 0);
-  ASSERT_EQ(NumTableFilesAtLevel(1, 0), 0);
-  // write some new keys into level 0 and 1
-  PutRandomData(0, 1024, 512);
-  ASSERT_OK(Flush(0));
-  WaitForCompaction();
-  PutRandomData(0, 10, 512);
-  ASSERT_OK(Flush(0));
-  // remember number of files in each level
-  int l1 = NumTableFilesAtLevel(0, 0);
-  int l2 = NumTableFilesAtLevel(1, 0);
-  int l3 = NumTableFilesAtLevel(2, 0);
-  ASSERT_NE(l1, 0);
-  ASSERT_NE(l2, 0);
-  ASSERT_NE(l3, 0);
-
  // SETUP column family "one" -- universal style
  for (int i = 0; i < one.level0_file_num_compaction_trigger - 1; ++i) {
    PutRandomData(1, 11, 10000);
@@ -805,12 +777,6 @@ TEST(ColumnFamilyTest, DifferentCompactionStyles) {
    ASSERT_EQ(std::to_string(i + 1), FilesPerLevel(2));
  }

-  // TRIGGER compaction "default"
-  // read a bunch of times, trigger read compaction
-  for (int i = 0; i < 200000; ++i) {
-    Get(0, std::to_string(i));
-  }
-
  // TRIGGER compaction "one"
  PutRandomData(1, 12, 10000);

@@ -820,13 +786,6 @@ TEST(ColumnFamilyTest, DifferentCompactionStyles) {
  // WAIT for compactions
  WaitForCompaction();

-  // VERIFY compaction "default"
-  // verify that the number of files have decreased
-  // in some level, indicating that there was a compaction
-  ASSERT_TRUE(NumTableFilesAtLevel(0, 0) < l1 ||
-              NumTableFilesAtLevel(1, 0) < l2 ||
-              NumTableFilesAtLevel(2, 0) < l3);
-
  // VERIFY compaction "one"
  ASSERT_EQ("1", FilesPerLevel(1));

@@ -970,6 +929,15 @@ TEST(ColumnFamilyTest, FlushStaleColumnFamilies) {
  Close();
 }

+TEST(ColumnFamilyTest, CreateMissingColumnFamilies) {
+  Status s = TryOpen({"one", "two"});
+  ASSERT_TRUE(!s.ok());
+  db_options_.create_missing_column_families = true;
+  s = TryOpen({"default", "one", "two"});
+  ASSERT_TRUE(s.ok());
+  Close();
+}
+
 }  // namespace rocksdb

 int main(int argc, char** argv) {
@@ -21,7 +21,7 @@ namespace rocksdb {
 static uint64_t TotalFileSize(const std::vector<FileMetaData*>& files) {
  uint64_t sum = 0;
  for (size_t i = 0; i < files.size() && files[i]; i++) {
-    sum += files[i]->file_size;
+    sum += files[i]->fd.GetFileSize();
  }
  return sum;
 }
@@ -29,7 +29,8 @@ static uint64_t TotalFileSize(const std::vector<FileMetaData*>& files) {
 Compaction::Compaction(Version* input_version, int level, int out_level,
                       uint64_t target_file_size,
                       uint64_t max_grandparent_overlap_bytes,
-                       bool seek_compaction, bool enable_compression,
+                       uint32_t output_path_id,
+                       CompressionType output_compression, bool seek_compaction,
                       bool deletion_compaction)
    : level_(level),
      out_level_(out_level),
@@ -38,8 +39,9 @@ Compaction::Compaction(Version* input_version, int level, int out_level,
      input_version_(input_version),
      number_levels_(input_version_->NumberLevels()),
      cfd_(input_version_->cfd_),
+      output_path_id_(output_path_id),
+      output_compression_(output_compression),
      seek_compaction_(seek_compaction),
-      enable_compression_(enable_compression),
      deletion_compaction_(deletion_compaction),
      grandparent_index_(0),
      seen_key_(false),
@@ -90,7 +92,7 @@ bool Compaction::IsDeletionCompaction() const { return deletion_compaction_; }
 void Compaction::AddInputDeletions(VersionEdit* edit) {
  for (int which = 0; which < 2; which++) {
    for (size_t i = 0; i < inputs_[which].size(); i++) {
-      edit->DeleteFile(level_ + which, inputs_[which][i]->number);
+      edit->DeleteFile(level_ + which, inputs_[which][i]->fd.GetNumber());
    }
  }
 }
@@ -127,7 +129,7 @@ bool Compaction::ShouldStopBefore(const Slice& internal_key) {
      icmp->Compare(internal_key,
                    grandparents_[grandparent_index_]->largest.Encode()) > 0) {
    if (seen_key_) {
-      overlapped_bytes_ += grandparents_[grandparent_index_]->file_size;
+      overlapped_bytes_ += grandparents_[grandparent_index_]->fd.GetFileSize();
    }
    assert(grandparent_index_ + 1 >= grandparents_.size() ||
           icmp->Compare(grandparents_[grandparent_index_]->largest.Encode(),
@@ -212,9 +214,9 @@ int InputSummary(const std::vector<FileMetaData*>& files, char* output,
    int sz = len - write;
    int ret;
    char sztxt[16];
-    AppendHumanBytes(files.at(i)->file_size, sztxt, 16);
-    ret = snprintf(output + write, sz, "%" PRIu64 "(%s) ", files.at(i)->number,
-                   sztxt);
+    AppendHumanBytes(files.at(i)->fd.GetFileSize(), sztxt, 16);
+    ret = snprintf(output + write, sz, "%" PRIu64 "(%s) ",
+                   files.at(i)->fd.GetNumber(), sztxt);
    if (ret < 0 || ret >= sz) break;
    write += ret;
  }
@@ -250,4 +252,20 @@ void Compaction::Summary(char* output, int len) {
  snprintf(output + write, len - write, "]");
 }

+uint64_t Compaction::OutputFilePreallocationSize() {
+  uint64_t preallocation_size = 0;
+
+  if (cfd_->options()->compaction_style == kCompactionStyleLevel) {
+    preallocation_size =
+        cfd_->compaction_picker()->MaxFileSizeForLevel(output_level());
+  } else {
+    for (const auto& f : inputs_[0]) {
+      preallocation_size += f->fd.GetFileSize();
+    }
+  }
+  // Over-estimate slightly so we don't end up just barely crossing
+  // the threshold
+  return preallocation_size * 1.1;
+}
+
 }  // namespace rocksdb
@@ -47,10 +47,13 @@ class Compaction {
  // Maximum size of files to build during this compaction.
  uint64_t MaxOutputFileSize() const { return max_output_file_size_; }

-  // Whether compression will be enabled for compaction outputs
-  bool enable_compression() const { return enable_compression_; }
+  // What compression for output
+  CompressionType OutputCompressionType() const { return output_compression_; }

-  // Is this a trivial compaction that can be implemented by just
+  // Whether need to write output file to second DB path.
+  uint32_t GetOutputPathId() const { return output_path_id_; }
+
+    // Is this a trivial compaction that can be implemented by just
  // moving a single input file to the next level (no merging or splitting)
  bool IsTrivialMove() const;

@@ -91,6 +94,11 @@ class Compaction {
  // Was this compaction triggered manually by the client?
  bool IsManualCompaction() { return is_manual_compaction_; }

+  // Returns a number of byte that the output file should be preallocated to
+  // In level compaction, that is max_file_size_. In universal compaction, that
+  // is the sum of all input file sizes
+  uint64_t OutputFilePreallocationSize();
+
 private:
  friend class CompactionPicker;
  friend class UniversalCompactionPicker;
@@ -99,8 +107,8 @@ class Compaction {

  Compaction(Version* input_version, int level, int out_level,
             uint64_t target_file_size, uint64_t max_grandparent_overlap_bytes,
-             bool seek_compaction = false, bool enable_compression = true,
-             bool deletion_compaction = false);
+             uint32_t output_path_id, CompressionType output_compression,
+             bool seek_compaction = false, bool deletion_compaction = false);

  int level_;
  int out_level_; // levels to which output files are stored
@@ -111,8 +119,9 @@ class Compaction {
  int number_levels_;
  ColumnFamilyData* cfd_;

+  uint32_t output_path_id_;
+  CompressionType output_compression_;
  bool seek_compaction_;
-  bool enable_compression_;
  // if true, just delete files in inputs_[0]
  bool deletion_compaction_;

@@ -12,17 +12,43 @@
 #define __STDC_FORMAT_MACROS
 #include <inttypes.h>
 #include <limits>
+#include "db/filename.h"
 #include "util/log_buffer.h"
 #include "util/statistics.h"

 namespace rocksdb {

 namespace {
+// Determine compression type, based on user options, level of the output
+// file and whether compression is disabled.
+// If enable_compression is false, then compression is always disabled no
+// matter what the values of the other two parameters are.
+// Otherwise, the compression type is determined based on options and level.
+CompressionType GetCompressionType(const Options& options, int level,
+                                   const bool enable_compression = true) {
+  if (!enable_compression) {
+    // disable compression
+    return kNoCompression;
+  }
+  // If the use has specified a different compression level for each level,
+  // then pick the compresison for that level.
+  if (!options.compression_per_level.empty()) {
+    const int n = options.compression_per_level.size() - 1;
+    // It is possible for level_ to be -1; in that case, we use level
+    // 0's compression.  This occurs mostly in backwards compatibility
+    // situations when the builder doesn't know what level the file
+    // belongs to.  Likewise, if level_ is beyond the end of the
+    // specified compression levels, use the last value.
+    return options.compression_per_level[std::max(0, std::min(level, n))];
+  } else {
+    return options.compression;
+  }
+}

-uint64_t TotalFileSize(const std::vector<FileMetaData*>& files) {
+uint64_t TotalCompensatedFileSize(const std::vector<FileMetaData*>& files) {
  uint64_t sum = 0;
  for (size_t i = 0; i < files.size() && files[i]; i++) {
-    sum += files[i]->file_size;
+    sum += files[i]->compensated_file_size;
  }
  return sum;
 }
@@ -80,7 +106,7 @@ void CompactionPicker::SizeBeingCompacted(std::vector<uint64_t>& sizes) {
    for (auto c : compactions_in_progress_[level]) {
      assert(c->level() == level);
      for (int i = 0; i < c->num_input_files(0); i++) {
-        total += c->input(0,i)->file_size;
+        total += c->input(0, i)->compensated_file_size;
      }
    }
    sizes[level] = total;
@@ -261,9 +287,9 @@ void CompactionPicker::SetupOtherInputs(Compaction* c) {
    std::vector<FileMetaData*> expanded0;
    c->input_version_->GetOverlappingInputs(
        level, &all_start, &all_limit, &expanded0, c->base_index_, nullptr);
-    const uint64_t inputs0_size = TotalFileSize(c->inputs_[0]);
-    const uint64_t inputs1_size = TotalFileSize(c->inputs_[1]);
-    const uint64_t expanded0_size = TotalFileSize(expanded0);
+    const uint64_t inputs0_size = TotalCompensatedFileSize(c->inputs_[0]);
+    const uint64_t inputs1_size = TotalCompensatedFileSize(c->inputs_[1]);
+    const uint64_t expanded0_size = TotalCompensatedFileSize(expanded0);
    uint64_t limit = ExpandedCompactionByteSizeLimit(level);
    if (expanded0.size() > c->inputs_[0].size() &&
        inputs1_size + expanded0_size < limit &&
@@ -278,14 +304,12 @@ void CompactionPicker::SetupOtherInputs(Compaction* c) {
      if (expanded1.size() == c->inputs_[1].size() &&
          !FilesInCompaction(expanded1)) {
        Log(options_->info_log,
-            "[%s] Expanding@%lu %lu+%lu (%lu+%lu bytes) to %lu+%lu (%lu+%lu "
-            "bytes)\n",
-            c->column_family_data()->GetName().c_str(), (unsigned long)level,
-            (unsigned long)(c->inputs_[0].size()),
-            (unsigned long)(c->inputs_[1].size()), (unsigned long)inputs0_size,
-            (unsigned long)inputs1_size, (unsigned long)(expanded0.size()),
-            (unsigned long)(expanded1.size()), (unsigned long)expanded0_size,
-            (unsigned long)inputs1_size);
+            "[%s] Expanding@%d %zu+%zu (%" PRIu64 "+%" PRIu64
+            " bytes) to %zu+%zu (%" PRIu64 "+%" PRIu64 "bytes)\n",
+            c->column_family_data()->GetName().c_str(), level,
+            c->inputs_[0].size(), c->inputs_[1].size(), inputs0_size,
+            inputs1_size, expanded0.size(), expanded1.size(), expanded0_size,
+            inputs1_size);
        smallest = new_start;
        largest = new_limit;
        c->inputs_[0] = expanded0;
@@ -335,7 +359,7 @@ Compaction* CompactionPicker::CompactRange(Version* version, int input_level,
        MaxFileSizeForLevel(input_level) * options_->source_compaction_factor;
    uint64_t total = 0;
    for (size_t i = 0; i + 1 < inputs.size(); ++i) {
-      uint64_t s = inputs[i]->file_size;
+      uint64_t s = inputs[i]->compensated_file_size;
      total += s;
      if (total >= limit) {
        **compaction_end = inputs[i + 1]->smallest;
@@ -347,7 +371,8 @@ Compaction* CompactionPicker::CompactRange(Version* version, int input_level,
  }
  Compaction* c = new Compaction(version, input_level, output_level,
                                 MaxFileSizeForLevel(output_level),
-                                 MaxGrandParentOverlapBytes(input_level));
+                                 MaxGrandParentOverlapBytes(input_level), 0,
+                                 GetCompressionType(*options_, output_level));

  c->inputs_[0] = inputs;
  if (ExpandWhileOverlapping(c) == false) {
@@ -407,31 +432,6 @@ Compaction* LevelCompactionPicker::PickCompaction(Version* version,
    }
  }

-  // Find compactions needed by seeks
-  FileMetaData* f = version->file_to_compact_;
-  if (c == nullptr && f != nullptr && !f->being_compacted) {
-
-    level = version->file_to_compact_level_;
-    int parent_index = -1;
-
-    // Only allow one level 0 compaction at a time.
-    // Do not pick this file if its parents at level+1 are being compacted.
-    if (level != 0 || compactions_in_progress_[0].empty()) {
-      if (!ParentRangeInCompaction(version, &f->smallest, &f->largest, level,
-                                   &parent_index)) {
-        c = new Compaction(version, level, level + 1,
-                           MaxFileSizeForLevel(level + 1),
-                           MaxGrandParentOverlapBytes(level), true);
-        c->inputs_[0].push_back(f);
-        c->parent_index_ = parent_index;
-        c->input_version_->file_to_compact_ = nullptr;
-        if (ExpandWhileOverlapping(c) == false) {
-          return nullptr;
-        }
-      }
-    }
-  }
-
  if (c == nullptr) {
    return nullptr;
  }
@@ -492,7 +492,8 @@ Compaction* LevelCompactionPicker::PickCompactionBySize(Version* version,
  assert(level >= 0);
  assert(level + 1 < NumberLevels());
  c = new Compaction(version, level, level + 1, MaxFileSizeForLevel(level + 1),
-                     MaxGrandParentOverlapBytes(level));
+                     MaxGrandParentOverlapBytes(level), 0,
+                     GetCompressionType(*options_, level + 1));
  c->score_ = score;

  // Pick the largest file in this level that is not already
@@ -507,11 +508,12 @@ Compaction* LevelCompactionPicker::PickCompactionBySize(Version* version,
    int index = file_size[i];
    FileMetaData* f = c->input_version_->files_[level][index];

-    // check to verify files are arranged in descending size
+    // Check to verify files are arranged in descending compensated size.
    assert((i == file_size.size() - 1) ||
           (i >= Version::number_of_files_to_sort_ - 1) ||
-           (f->file_size >=
-            c->input_version_->files_[level][file_size[i + 1]]->file_size));
+           (f->compensated_file_size >=
+            c->input_version_->files_[level][file_size[i + 1]]->
+                compensated_file_size));

    // do not pick a file to compact if it is being compacted
    // from n-1 level.
@@ -611,15 +613,9 @@ Compaction* UniversalCompactionPicker::PickCompaction(Version* version,
    newerfile = f;
  }

-  // The files are sorted from newest first to oldest last.
-  std::vector<int>& file_by_time = c->input_version_->files_by_size_[level];
-
  // Is the earliest file part of this compaction?
-  int last_index = file_by_time[file_by_time.size()-1];
-  FileMetaData* last_file = c->input_version_->files_[level][last_index];
-  if (c->inputs_[0][c->inputs_[0].size()-1] == last_file) {
-    c->bottommost_level_ = true;
-  }
+  FileMetaData* last_file = c->input_version_->files_[level].back();
+  c->bottommost_level_ = c->inputs_[0].back() == last_file;

  // update statistics
  MeasureTime(options_->statistics.get(), NUM_FILES_IN_SINGLE_COMPACTION,
@@ -654,12 +650,12 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(
    options_->compaction_options_universal.max_merge_width;

  // The files are sorted from newest first to oldest last.
-  std::vector<int>& file_by_time = version->files_by_size_[level];
+  const auto& files = version->files_[level];
+
  FileMetaData* f = nullptr;
  bool done = false;
  int start_index = 0;
  unsigned int candidate_count = 0;
-  assert(file_by_time.size() == version->files_[level].size());

  unsigned int max_files_to_compact = std::min(max_merge_width,
                                       max_number_of_files_to_compact);
@@ -667,40 +663,38 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(

  // Considers a candidate file only if it is smaller than the
  // total size accumulated so far.
-  for (unsigned int loop = 0; loop < file_by_time.size(); loop++) {
+  for (unsigned int loop = 0; loop < files.size(); loop++) {

    candidate_count = 0;

    // Skip files that are already being compacted
-    for (f = nullptr; loop < file_by_time.size(); loop++) {
-      int index = file_by_time[loop];
-      f = version->files_[level][index];
+    for (f = nullptr; loop < files.size(); loop++) {
+      f = files[loop];

      if (!f->being_compacted) {
        candidate_count = 1;
        break;
      }
-      LogToBuffer(
-          log_buffer, "[%s] Universal: file %lu[%d] being compacted, skipping",
-          version->cfd_->GetName().c_str(), (unsigned long)f->number, loop);
+      LogToBuffer(log_buffer, "[%s] Universal: file %" PRIu64
+                              "[%d] being compacted, skipping",
+                  version->cfd_->GetName().c_str(), f->fd.GetNumber(), loop);
      f = nullptr;
    }

    // This file is not being compacted. Consider it as the
    // first candidate to be compacted.
-    uint64_t candidate_size =  f != nullptr? f->file_size : 0;
+    uint64_t candidate_size =  f != nullptr? f->compensated_file_size : 0;
    if (f != nullptr) {
      LogToBuffer(
-          log_buffer, "[%s] Universal: Possible candidate file %lu[%d].",
-          version->cfd_->GetName().c_str(), (unsigned long)f->number, loop);
+          log_buffer, "[%s] Universal: Possible candidate file %s[%d].",
+          version->cfd_->GetName().c_str(),
+          FormatFileNumber(f->fd.GetNumber(), f->fd.GetPathId()).c_str(), loop);
    }

    // Check if the suceeding files need compaction.
-    for (unsigned int i = loop+1;
-         candidate_count < max_files_to_compact && i < file_by_time.size();
-         i++) {
-      int index = file_by_time[i];
-      FileMetaData* f = version->files_[level][index];
+    for (unsigned int i = loop + 1;
+         candidate_count < max_files_to_compact && i < files.size(); i++) {
+      FileMetaData* f = files[i];
      if (f->being_compacted) {
        break;
      }
@@ -711,13 +705,13 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(
      // kCompactionStopStyleSimilarSize, it's simply the size of the last
      // picked file.
      uint64_t sz = (candidate_size * (100L + ratio)) /100;
-      if (sz < f->file_size) {
+      if (sz < f->fd.GetFileSize()) {
        break;
      }
      if (options_->compaction_options_universal.stop_style == kCompactionStopStyleSimilarSize) {
        // Similar-size stopping rule: also check the last picked file isn't
        // far larger than the next candidate file.
-        sz = (f->file_size * (100L + ratio)) / 100;
+        sz = (f->fd.GetFileSize() * (100L + ratio)) / 100;
        if (sz < candidate_size) {
          // If the small file we've encountered begins a run of similar-size
          // files, we'll pick them up on a future iteration of the outer
@@ -725,9 +719,9 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(
          // by the last-resort read amp strategy which disregards size ratios.
          break;
        }
-        candidate_size = f->file_size;
+        candidate_size = f->compensated_file_size;
      } else { // default kCompactionStopStyleTotalSize
-        candidate_size += f->file_size;
+        candidate_size += f->compensated_file_size;
      }
      candidate_count++;
    }
@@ -739,13 +733,14 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(
      break;
    } else {
      for (unsigned int i = loop;
-           i < loop + candidate_count && i < file_by_time.size(); i++) {
-       int index = file_by_time[i];
-       FileMetaData* f = version->files_[level][index];
-       LogToBuffer(log_buffer,
-                   "[%s] Universal: Skipping file %lu[%d] with size %lu %d\n",
-                   version->cfd_->GetName().c_str(), (unsigned long)f->number,
-                   i, (unsigned long)f->file_size, f->being_compacted);
+           i < loop + candidate_count && i < files.size(); i++) {
+        FileMetaData* f = files[i];
+        LogToBuffer(log_buffer, "[%s] Universal: Skipping file %" PRIu64
+                                "[%d] with size %" PRIu64
+                                " (compensated size %" PRIu64 ") %d\n",
+                    version->cfd_->GetName().c_str(), f->fd.GetNumber(), i,
+                    f->fd.GetFileSize(), f->compensated_file_size,
+                    f->being_compacted);
      }
    }
  }
@@ -761,28 +756,29 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalReadAmp(
  if (ratio_to_compress >= 0) {
    uint64_t total_size = version->NumLevelBytes(level);
    uint64_t older_file_size = 0;
-    for (unsigned int i = file_by_time.size() - 1; i >= first_index_after;
-        i--) {
-      older_file_size += version->files_[level][file_by_time[i]]->file_size;
+    for (unsigned int i = files.size() - 1;
+         i >= first_index_after; i--) {
+      older_file_size += files[i]->fd.GetFileSize();
      if (older_file_size * 100L >= total_size * (long) ratio_to_compress) {
        enable_compression = false;
        break;
      }
    }
  }
-  Compaction* c =
-      new Compaction(version, level, level, MaxFileSizeForLevel(level),
-                     LLONG_MAX, false, enable_compression);
+  Compaction* c = new Compaction(
+      version, level, level, MaxFileSizeForLevel(level), LLONG_MAX, 0,
+      GetCompressionType(*options_, level, enable_compression));
  c->score_ = score;

  for (unsigned int i = start_index; i < first_index_after; i++) {
-    int index = file_by_time[i];
-    FileMetaData* f = c->input_version_->files_[level][index];
+    FileMetaData* f = c->input_version_->files_[level][i];
    c->inputs_[0].push_back(f);
    LogToBuffer(log_buffer,
-                "[%s] Universal: Picking file %lu[%d] with size %lu\n",
-                version->cfd_->GetName().c_str(), (unsigned long)f->number, i,
-                (unsigned long)f->file_size);
+                "[%s] Universal: Picking file %s[%d] "
+                "with size %" PRIu64 " (compensated size %" PRIu64 ")\n",
+                version->cfd_->GetName().c_str(),
+                FormatFileNumber(f->fd.GetNumber(), f->fd.GetPathId()).c_str(),
+                i, f->fd.GetFileSize(), f->compensated_file_size);
  }
  return c;
 }
@@ -802,8 +798,7 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalSizeAmp(
                     max_size_amplification_percent;

  // The files are sorted from newest first to oldest last.
-  std::vector<int>& file_by_time = version->files_by_size_[level];
-  assert(file_by_time.size() == version->files_[level].size());
+  const auto& files = version->files_[level];

  unsigned int candidate_count = 0;
  uint64_t candidate_size = 0;
@@ -811,16 +806,15 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalSizeAmp(
  FileMetaData* f = nullptr;

  // Skip files that are already being compacted
-  for (unsigned int loop = 0; loop < file_by_time.size() - 1; loop++) {
-    int index = file_by_time[loop];
-    f = version->files_[level][index];
+  for (unsigned int loop = 0; loop < files.size() - 1; loop++) {
+    f = files[loop];
    if (!f->being_compacted) {
      start_index = loop;         // Consider this as the first candidate.
      break;
    }
-    LogToBuffer(log_buffer,
-                "[%s] Universal: skipping file %lu[%d] compacted %s",
-                version->cfd_->GetName().c_str(), (unsigned long)f->number,
+    LogToBuffer(log_buffer, "[%s] Universal: skipping file %s[%d] compacted %s",
+                version->cfd_->GetName().c_str(),
+                FormatFileNumber(f->fd.GetNumber(), f->fd.GetPathId()).c_str(),
                loop, " cannot be a candidate to reduce size amp.\n");
    f = nullptr;
  }
@@ -828,23 +822,23 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalSizeAmp(
    return nullptr;             // no candidate files
  }

-  LogToBuffer(log_buffer, "[%s] Universal: First candidate file %lu[%d] %s",
-              version->cfd_->GetName().c_str(), (unsigned long)f->number,
+  LogToBuffer(log_buffer, "[%s] Universal: First candidate file %s[%d] %s",
+              version->cfd_->GetName().c_str(),
+              FormatFileNumber(f->fd.GetNumber(), f->fd.GetPathId()).c_str(),
              start_index, " to reduce size amp.\n");

  // keep adding up all the remaining files
-  for (unsigned int loop = start_index; loop < file_by_time.size() - 1;
-       loop++) {
-    int index = file_by_time[loop];
-    f = version->files_[level][index];
+  for (unsigned int loop = start_index; loop < files.size() - 1; loop++) {
+    f = files[loop];
    if (f->being_compacted) {
      LogToBuffer(
-          log_buffer, "[%s] Universal: Possible candidate file %lu[%d] %s.",
-          version->cfd_->GetName().c_str(), (unsigned long)f->number, loop,
+          log_buffer, "[%s] Universal: Possible candidate file %s[%d] %s.",
+          version->cfd_->GetName().c_str(),
+          FormatFileNumber(f->fd.GetNumber(), f->fd.GetPathId()).c_str(), loop,
          " is already being compacted. No size amp reduction possible.\n");
      return nullptr;
    }
-    candidate_size += f->file_size;
+    candidate_size += f->compensated_file_size;
    candidate_count++;
  }
  if (candidate_count == 0) {
@@ -852,41 +846,40 @@ Compaction* UniversalCompactionPicker::PickCompactionUniversalSizeAmp(
  }

  // size of earliest file
-  int index = file_by_time[file_by_time.size() - 1];
-  uint64_t earliest_file_size = version->files_[level][index]->file_size;
+  uint64_t earliest_file_size = files.back()->fd.GetFileSize();

  // size amplification = percentage of additional size
  if (candidate_size * 100 < ratio * earliest_file_size) {
    LogToBuffer(
        log_buffer,
-        "[%s] Universal: size amp not needed. newer-files-total-size %lu "
-        "earliest-file-size %lu",
-        version->cfd_->GetName().c_str(), (unsigned long)candidate_size,
-        (unsigned long)earliest_file_size);
+        "[%s] Universal: size amp not needed. newer-files-total-size %" PRIu64
+        "earliest-file-size %" PRIu64,
+        version->cfd_->GetName().c_str(), candidate_size, earliest_file_size);
    return nullptr;
  } else {
-    LogToBuffer(log_buffer,
-                "[%s] Universal: size amp needed. newer-files-total-size %lu "
-                "earliest-file-size %lu",
-                version->cfd_->GetName().c_str(), (unsigned long)candidate_size,
-                (unsigned long)earliest_file_size);
+    LogToBuffer(
+        log_buffer,
+        "[%s] Universal: size amp needed. newer-files-total-size %" PRIu64
+        "earliest-file-size %" PRIu64,
+        version->cfd_->GetName().c_str(), candidate_size, earliest_file_size);
  }
-  assert(start_index >= 0 && start_index < file_by_time.size() - 1);
+  assert(start_index >= 0 && start_index < files.size() - 1);

  // create a compaction request
  // We always compact all the files, so always compress.
  Compaction* c =
      new Compaction(version, level, level, MaxFileSizeForLevel(level),
-                     LLONG_MAX, false, true);
+                     LLONG_MAX, 0, GetCompressionType(*options_, level));
  c->score_ = score;
-  for (unsigned int loop = start_index; loop < file_by_time.size(); loop++) {
-    int index = file_by_time[loop];
-    f = c->input_version_->files_[level][index];
+  for (unsigned int loop = start_index; loop < files.size(); loop++) {
+    f = c->input_version_->files_[level][loop];
    c->inputs_[0].push_back(f);
    LogToBuffer(log_buffer,
-                "[%s] Universal: size amp picking file %lu[%d] with size %lu",
-                version->cfd_->GetName().c_str(), (unsigned long)f->number,
-                index, (unsigned long)f->file_size);
+        "[%s] Universal: size amp picking file %" PRIu64 "[%d] "
+        "with size %" PRIu64 " (compensated size %" PRIu64 ")",
+        version->cfd_->GetName().c_str(),
+        f->fd.GetNumber(), loop,
+        f->fd.GetFileSize(), f->compensated_file_size);
  }
  return c;
 }
@@ -896,7 +889,7 @@ Compaction* FIFOCompactionPicker::PickCompaction(Version* version,
  assert(version->NumberLevels() == 1);
  uint64_t total_size = 0;
  for (const auto& file : version->files_[0]) {
-    total_size += file->file_size;
+    total_size += file->compensated_file_size;
  }

  if (total_size <= options_->compaction_options_fifo.max_table_files_size ||
@@ -918,19 +911,19 @@ Compaction* FIFOCompactionPicker::PickCompaction(Version* version,
    return nullptr;
  }

-  Compaction* c = new Compaction(version, 0, 0, 0, 0, false, false,
+  Compaction* c = new Compaction(version, 0, 0, 0, 0, 0, kNoCompression, false,
                                 true /* is deletion compaction */);
  // delete old files (FIFO)
  for (auto ritr = version->files_[0].rbegin();
       ritr != version->files_[0].rend(); ++ritr) {
    auto f = *ritr;
-    total_size -= f->file_size;
+    total_size -= f->compensated_file_size;
    c->inputs_[0].push_back(f);
    char tmp_fsize[16];
-    AppendHumanBytes(f->file_size, tmp_fsize, sizeof(tmp_fsize));
+    AppendHumanBytes(f->fd.GetFileSize(), tmp_fsize, sizeof(tmp_fsize));
    LogToBuffer(log_buffer, "[%s] FIFO compaction: picking file %" PRIu64
                            " with size %s for deletion",
-                version->cfd_->GetName().c_str(), f->number, tmp_fsize);
+                version->cfd_->GetName().c_str(), f->fd.GetNumber(), tmp_fsize);
    if (total_size <= options_->compaction_options_fifo.max_table_files_size) {
      break;
    }
@@ -17,6 +17,11 @@ int main() {
 }
 #else

+#ifdef NUMA
+#include <numa.h>
+#include <numaif.h>
+#endif
+
 #include <inttypes.h>
 #include <cstddef>
 #include <sys/types.h>
@@ -173,6 +178,14 @@ DEFINE_double(compression_ratio, 0.5, "Arrange to generate values that shrink"

 DEFINE_bool(histogram, false, "Print histogram of operation timings");

+DEFINE_bool(enable_numa, false,
+            "Make operations aware of NUMA architecture and bind memory "
+            "and cpus corresponding to nodes together. In NUMA, memory "
+            "in same node as CPUs are closer when compared to memory in "
+            "other nodes. Reads can be faster when the process is bound to "
+            "CPU and memory of same node. Use \"$numactl --hardware\" command "
+            "to see NUMA memory architecture.");
+
 DEFINE_int64(write_buffer_size, rocksdb::Options().write_buffer_size,
             "Number of bytes to buffer in memtable before compacting");

@@ -333,9 +346,6 @@ DEFINE_int32(deletepercent, 2, "Percentage of deletes out of reads/writes/"
             "deletepercent), so deletepercent must be smaller than (100 - "
             "FLAGS_readwritepercent)");

-DEFINE_int32(disable_seek_compaction, false, "Option to disable compaction"
-             " triggered by read.");
-
 DEFINE_uint64(delete_obsolete_files_period_micros, 0, "Option to delete "
              "obsolete files periodically. 0 means that obsolete files are"
              " deleted after every compaction run.");
@@ -465,6 +475,8 @@ static auto FLAGS_compaction_fadvice_e =

 DEFINE_bool(use_tailing_iterator, false,
            "Use tailing iterator to access a series of keys instead of get");
+DEFINE_int64(iter_refresh_interval_us, -1,
+             "How often to refresh iterators. Disable refresh when -1");

 DEFINE_bool(use_adaptive_mutex, rocksdb::Options().use_adaptive_mutex,
            "Use adaptive mutex");
@@ -551,6 +563,9 @@ static const bool FLAGS_cache_numshardbits_dummy __attribute__((unused)) =
 static const bool FLAGS_readwritepercent_dummy __attribute__((unused)) =
    RegisterFlagValidator(&FLAGS_readwritepercent, &ValidateInt32Percent);

+DEFINE_int32(disable_seek_compaction, false,
+             "Not used, left here for backwards compatibility");
+
 static const bool FLAGS_deletepercent_dummy __attribute__((unused)) =
    RegisterFlagValidator(&FLAGS_deletepercent, &ValidateInt32Percent);
 static const bool FLAGS_table_cache_numshardbits_dummy __attribute__((unused)) =
@@ -582,9 +597,9 @@ class RandomGenerator {
  }

  Slice Generate(unsigned int len) {
+    assert(len <= data_.size());
    if (pos_ + len > data_.size()) {
      pos_ = 0;
-      assert(len < data_.size());
    }
    pos_ += len;
    return Slice(data_.data() + pos_ - len, len);
@@ -662,7 +677,7 @@ class Stats {
  void SetId(int id) { id_ = id; }
  void SetExcludeFromMerge() { exclude_from_merge_ = true; }

-  void FinishedSingleOp(DB* db) {
+  void FinishedOps(DB* db, int64_t num_ops) {
    if (FLAGS_histogram) {
      double now = FLAGS_env->NowMicros();
      double micros = now - last_op_finish_;
@@ -674,7 +689,7 @@ class Stats {
      last_op_finish_ = now;
    }

-    done_++;
+    done_ += num_ops;
    if (done_ >= next_report_) {
      if (!FLAGS_stats_interval) {
        if      (next_report_ < 1000)   next_report_ += 100;
@@ -720,7 +735,7 @@ class Stats {

  void Report(const Slice& name) {
    // Pretend at least one op was done in case we are running a benchmark
-    // that does not call FinishedSingleOp().
+    // that does not call FinishedOps().
    if (done_ < 1) done_ = 1;

    std::string extra;
@@ -835,6 +850,15 @@ class Benchmark {
  int64_t writes_;
  int64_t readwrites_;
  int64_t merge_keys_;
+
+  bool SanityCheck() {
+    if (FLAGS_compression_ratio > 1) {
+      fprintf(stderr, "compression_ratio should be between 0 and 1\n");
+      return false;
+    }
+    return true;
+  }
+
  void PrintHeader() {
    PrintEnvironment();
    fprintf(stdout, "Keys:       %d bytes each\n", FLAGS_key_size);
@@ -852,6 +876,18 @@ class Benchmark {
              * num_)
             / 1048576.0));
    fprintf(stdout, "Write rate limit: %d\n", FLAGS_writes_per_second);
+    if (FLAGS_enable_numa) {
+      fprintf(stderr, "Running in NUMA enabled mode.\n");
+#ifndef NUMA
+      fprintf(stderr, "NUMA is not defined in the system.\n");
+      exit(1);
+#else
+      if (numa_available() == -1) {
+        fprintf(stderr, "NUMA is not supported by the system.\n");
+        exit(1);
+      }
+#endif
+    }
    switch (FLAGS_compression_type_e) {
      case rocksdb::kNoCompression:
        fprintf(stdout, "Compression: none\n");
@@ -1114,6 +1150,9 @@ class Benchmark {
  }

  void Run() {
+    if (!SanityCheck()) {
+      exit(1);
+    }
    PrintHeader();
    Open();
    const char* benchmarks = FLAGS_benchmarks.c_str();
@@ -1334,7 +1373,25 @@ class Benchmark {
    shared.start = false;

    ThreadArg* arg = new ThreadArg[n];
+
    for (int i = 0; i < n; i++) {
+#ifdef NUMA
+      if (FLAGS_enable_numa) {
+        // Performs a local allocation of memory to threads in numa node.
+        int n_nodes = numa_num_task_nodes();  // Number of nodes in NUMA.
+        numa_exit_on_error = 1;
+        int numa_node = i % n_nodes;
+        bitmask* nodes = numa_allocate_nodemask();
+        numa_bitmask_clearall(nodes);
+        numa_bitmask_setbit(nodes, numa_node);
+        // numa_bind() call binds the process to the node and these
+        // properties are passed on to the thread that is created in
+        // StartThread method called later in the loop.
+        numa_bind(nodes);
+        numa_set_strict(1);
+        numa_free_nodemask(nodes);
+      }
+#endif
      arg[i].bm = this;
      arg[i].method = method;
      arg[i].shared = &shared;
@@ -1377,7 +1434,7 @@ class Benchmark {
    uint32_t crc = 0;
    while (bytes < 500 * 1048576) {
      crc = crc32c::Value(data.data(), size);
-      thread->stats.FinishedSingleOp(nullptr);
+      thread->stats.FinishedOps(nullptr, 1);
      bytes += size;
    }
    // Print so result is not dead
@@ -1396,7 +1453,7 @@ class Benchmark {
    unsigned int xxh32 = 0;
    while (bytes < 500 * 1048576) {
      xxh32 = XXH32(data.data(), size, 0);
-      thread->stats.FinishedSingleOp(nullptr);
+      thread->stats.FinishedOps(nullptr, 1);
      bytes += size;
    }
    // Print so result is not dead
@@ -1417,7 +1474,7 @@ class Benchmark {
        ptr = ap.Acquire_Load();
      }
      count++;
-      thread->stats.FinishedSingleOp(nullptr);
+      thread->stats.FinishedOps(nullptr, 1);
    }
    if (ptr == nullptr) exit(1); // Disable unused variable warning.
  }
@@ -1458,7 +1515,7 @@ class Benchmark {
      }
      produced += compressed.size();
      bytes += input.size();
-      thread->stats.FinishedSingleOp(nullptr);
+      thread->stats.FinishedOps(nullptr, 1);
    }

    if (!ok) {
@@ -1539,7 +1596,7 @@ class Benchmark {
      }
      delete[] uncompressed;
      bytes += input.size();
-      thread->stats.FinishedSingleOp(nullptr);
+      thread->stats.FinishedOps(nullptr, 1);
    }

    if (!ok) {
@@ -1661,7 +1718,6 @@ class Benchmark {
        options.compression_per_level[i] = FLAGS_compression_type_e;
      }
    }
-    options.disable_seek_compaction = FLAGS_disable_seek_compaction;
    options.delete_obsolete_files_period_micros =
      FLAGS_delete_obsolete_files_period_micros;
    options.soft_rate_limit = FLAGS_soft_rate_limit;
@@ -1849,9 +1905,9 @@ class Benchmark {
        GenerateKeyFromInt(key_gens[id]->Next(), FLAGS_num, &key);
        batch.Put(key, gen.Generate(value_size_));
        bytes += value_size_ + key_size_;
-        thread->stats.FinishedSingleOp(db_to_write);
      }
      s = db_to_write->Write(write_options_, &batch);
+      thread->stats.FinishedOps(db_to_write, entries_per_batch_);
      if (!s.ok()) {
        fprintf(stderr, "put error: %s\n", s.ToString().c_str());
        exit(1);
@@ -1876,7 +1932,7 @@ class Benchmark {
    int64_t bytes = 0;
    for (iter->SeekToFirst(); i < reads_ && iter->Valid(); iter->Next()) {
      bytes += iter->key().size() + iter->value().size();
-      thread->stats.FinishedSingleOp(db);
+      thread->stats.FinishedOps(db, 1);
      ++i;
    }
    delete iter;
@@ -1899,7 +1955,7 @@ class Benchmark {
    int64_t bytes = 0;
    for (iter->SeekToLast(); i < reads_ && iter->Valid(); iter->Prev()) {
      bytes += iter->key().size() + iter->value().size();
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
      ++i;
    }
    delete iter;
@@ -1922,11 +1978,11 @@ class Benchmark {
      if (db->Get(options, key, &value).ok()) {
        found++;
      }
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }

    char msg[100];
-    snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)",
+    snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)\n",
             found, read);

    thread->stats.AddMessage(msg);
@@ -1982,7 +2038,7 @@ class Benchmark {
      DB* db = SelectDB(thread);
      Iterator* iter = db->NewIterator(options);
      delete iter;
-      thread->stats.FinishedSingleOp(db);
+      thread->stats.FinishedOps(db, 1);
    }
  }

@@ -2009,12 +2065,31 @@ class Benchmark {
        multi_iters.push_back(db->NewIterator(options));
      }
    }
+    uint64_t last_refresh = FLAGS_env->NowMicros();

    Slice key = AllocateKey();
    std::unique_ptr<const char[]> key_guard(key.data());

    Duration duration(FLAGS_duration, reads_);
    while (!duration.Done(1)) {
+      if (!FLAGS_use_tailing_iterator && FLAGS_iter_refresh_interval_us >= 0) {
+        uint64_t now = FLAGS_env->NowMicros();
+        if (now - last_refresh > (uint64_t)FLAGS_iter_refresh_interval_us) {
+          if (db_ != nullptr) {
+            delete single_iter;
+            single_iter = db_->NewIterator(options);
+          } else {
+            for (auto iter : multi_iters) {
+              delete iter;
+            }
+            multi_iters.clear();
+            for (DB* db : multi_dbs_) {
+              multi_iters.push_back(db->NewIterator(options));
+            }
+          }
+        }
+        last_refresh = now;
+      }
      // Pick a Iterator to use
      Iterator* iter_to_use = single_iter;
      if (single_iter == nullptr) {
@@ -2027,7 +2102,7 @@ class Benchmark {
      if (iter_to_use->Valid() && iter_to_use->key().compare(key) == 0) {
        found++;
      }
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }
    delete single_iter;
    for (auto iter : multi_iters) {
@@ -2035,9 +2110,12 @@ class Benchmark {
    }

    char msg[100];
-    snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)",
+    snprintf(msg, sizeof(msg), "(%" PRIu64 " of %" PRIu64 " found)\n",
             found, read);
    thread->stats.AddMessage(msg);
+    if (FLAGS_perf_level > 0) {
+      thread->stats.AddMessage(perf_context.ToString());
+    }
  }

  void SeekRandomWhileWriting(ThreadState* thread) {
@@ -2062,9 +2140,9 @@ class Benchmark {
        const int64_t k = seq ? i + j : (thread->rand.Next() % FLAGS_num);
        GenerateKeyFromInt(k, FLAGS_num, &key);
        batch.Delete(key);
-        thread->stats.FinishedSingleOp(db);
      }
      auto s = db->Write(write_options_, &batch);
+      thread->stats.FinishedOps(db, entries_per_batch_);
      if (!s.ok()) {
        fprintf(stderr, "del error: %s\n", s.ToString().c_str());
        exit(1);
@@ -2124,7 +2202,7 @@ class Benchmark {
        fprintf(stderr, "put error: %s\n", s.ToString().c_str());
        exit(1);
      }
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);

      ++num_writes;
      if (writes_per_second_by_10 && num_writes >= writes_per_second_by_10) {
@@ -2284,7 +2362,7 @@ class Benchmark {
        deletes_done++;
      }

-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }
    char msg[100];
    snprintf(msg, sizeof(msg),
@@ -2342,7 +2420,7 @@ class Benchmark {
        put_weight--;
        writes_done++;
      }
-      thread->stats.FinishedSingleOp(db);
+      thread->stats.FinishedOps(db, 1);
    }
    char msg[100];
    snprintf(msg, sizeof(msg), "( reads:%" PRIu64 " writes:%" PRIu64 \
@@ -2376,7 +2454,7 @@ class Benchmark {
        fprintf(stderr, "put error: %s\n", s.ToString().c_str());
        exit(1);
      }
-      thread->stats.FinishedSingleOp(db);
+      thread->stats.FinishedOps(db, 1);
    }
    char msg[100];
    snprintf(msg, sizeof(msg),
@@ -2423,7 +2501,7 @@ class Benchmark {
        fprintf(stderr, "put error: %s\n", s.ToString().c_str());
        exit(1);
      }
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }

    char msg[100];
@@ -2459,7 +2537,7 @@ class Benchmark {
        fprintf(stderr, "merge error: %s\n", s.ToString().c_str());
        exit(1);
      }
-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }

    // Print some statistics
@@ -2520,7 +2598,7 @@ class Benchmark {

      }

-      thread->stats.FinishedSingleOp(db_);
+      thread->stats.FinishedOps(db_, 1);
    }

    char msg[100];
@@ -98,7 +98,7 @@ Status DBImpl::GetLiveFiles(std::vector<std::string>& ret,
  }

  // Make a set of all of the live *.sst files
-  std::set<uint64_t> live;
+  std::vector<FileDescriptor> live;
  for (auto cfd : *versions_->GetColumnFamilySet()) {
    cfd->current()->AddLiveFiles(&live);
  }
@@ -109,7 +109,7 @@ Status DBImpl::GetLiveFiles(std::vector<std::string>& ret,
  // create names of the live files. The names are not absolute
  // paths, instead they are relative to dbname_;
  for (auto live_file : live) {
-    ret.push_back(TableFileName("", live_file));
+    ret.push_back(MakeTableFileName("", live_file.GetNumber()));
  }

  ret.push_back(CurrentFileName(""));
@@ -10,6 +10,7 @@

 #include <atomic>
 #include <deque>
+#include <limits>
 #include <set>
 #include <utility>
 #include <vector>
@@ -28,6 +29,7 @@
 #include "rocksdb/transaction_log.h"
 #include "util/autovector.h"
 #include "util/stats_logger.h"
+#include "util/stop_watch.h"
 #include "util/thread_local.h"
 #include "db/internal_stats.h"

@@ -39,6 +41,7 @@ class Version;
 class VersionEdit;
 class VersionSet;
 class CompactionFilterV2;
+class Arena;

 class DBImpl : public DB {
 public:
@@ -197,6 +200,17 @@ class DBImpl : public DB {
  Status TEST_ReadFirstLine(const std::string& fname, SequenceNumber* sequence);
 #endif  // NDEBUG

+  // Structure to store information for candidate files to delete.
+  struct CandidateFileInfo {
+    std::string file_name;
+    uint32_t path_id;
+    CandidateFileInfo(std::string name, uint32_t path)
+        : file_name(name), path_id(path) {}
+    bool operator==(const CandidateFileInfo& other) const {
+      return file_name == other.file_name && path_id == other.path_id;
+    }
+  };
+
  // needed for CleanupIteratorState
  struct DeletionState {
    inline bool HaveSomethingToDelete() const {
@@ -208,10 +222,10 @@ class DBImpl : public DB {
    // a list of all files that we'll consider deleting
    // (every once in a while this is filled up with all files
    // in the DB directory)
-    std::vector<std::string> candidate_files;
+    std::vector<CandidateFileInfo> candidate_files;

    // the list of all live sst files that cannot be deleted
-    std::vector<uint64_t> sst_live;
+    std::vector<FileDescriptor> sst_live;

    // a list of sst files that we need to delete
    std::vector<FileMetaData*> sst_delete_files;
@@ -278,13 +292,15 @@ class DBImpl : public DB {
  const DBOptions options_;

  Iterator* NewInternalIterator(const ReadOptions&, ColumnFamilyData* cfd,
-                                SuperVersion* super_version);
+                                SuperVersion* super_version,
+                                Arena* arena = nullptr);

 private:
  friend class DB;
  friend class InternalStats;
 #ifndef ROCKSDB_LITE
  friend class TailingIterator;
+  friend class ForwardIterator;
 #endif
  friend struct SuperVersion;
  struct CompactionState;
@@ -331,7 +347,8 @@ class DBImpl : public DB {
  Status MakeRoomForWrite(ColumnFamilyData* cfd,
                          bool force /* flush even if there is room? */,
                          autovector<SuperVersion*>* superversions_to_free,
-                          autovector<log::Writer*>* logs_to_free);
+                          autovector<log::Writer*>* logs_to_free,
+                          uint64_t expiration_time);

  void BuildBatchGroup(Writer** last_writer,
                       autovector<WriteBatch*>* write_batch_group);
@@ -342,6 +359,9 @@ class DBImpl : public DB {
  // Wait for memtable flushed
  Status WaitForFlushMemTable(ColumnFamilyData* cfd);

+  void RecordFlushIOStats();
+  void RecordCompactionIOStats();
+
  void MaybeScheduleLogDBDeployStats();

 #ifndef ROCKSDB_LITE
@@ -373,6 +393,7 @@ class DBImpl : public DB {
  // Call compaction filter if is_compaction_v2 is not true. Then iterate
  // through input and compact the kv-pairs
  Status ProcessKeyValueCompaction(
+    bool is_snapshot_supported,
    SequenceNumber visible_at_tip,
    SequenceNumber earliest_snapshot,
    SequenceNumber latest_snapshot,
@@ -449,7 +470,16 @@ class DBImpl : public DB {
  // State below is protected by mutex_
  port::Mutex mutex_;
  port::AtomicPointer shutting_down_;
-  port::CondVar bg_cv_;          // Signalled when background work finishes
+  // This condition variable is signaled on these conditions:
+  // * whenever bg_compaction_scheduled_ goes down to 0
+  // * if bg_manual_only_ > 0, whenever a compaction finishes, even if it hasn't
+  // made any progress
+  // * whenever a compaction made any progress
+  // * whenever bg_flush_scheduled_ value decreases (i.e. whenever a flush is
+  // done, even if it didn't make any progress)
+  // * whenever there is an error in background flush or compaction
+  // * whenever bg_logstats_scheduled_ turns to false
+  port::CondVar bg_cv_;
  uint64_t logfile_number_;
  unique_ptr<log::Writer> log_;
  bool log_empty_;
@@ -468,6 +498,9 @@ class DBImpl : public DB {
  // only used for dynamically adjusting max_total_wal_size. it is a sum of
  // [write_buffer_size * max_write_buffer_number] over all column families
  uint64_t max_total_in_memory_state_;
+  // If true, we have only one (default) column family. We use this to optimize
+  // some code-paths
+  bool single_column_family_mode_;

  std::string host_name_;

@@ -485,7 +518,8 @@ class DBImpl : public DB {

  // Set of table files to protect from deletion because they are
  // part of ongoing compactions.
-  std::set<uint64_t> pending_outputs_;
+  // map from pending file number ID to their path IDs.
+  FileNumToPathIdMap pending_outputs_;

  // At least one compaction or flush job is pending but not yet scheduled
  // because of the max background thread limit.
@@ -550,6 +584,7 @@ class DBImpl : public DB {
  bool flush_on_destroy_; // Used when disableWAL is true.

  static const int KEEP_LOG_FILE_NUM = 1000;
+  static const uint64_t kNoTimeOut = std::numeric_limits<uint64_t>::max();
  std::string db_absolute_path_;

  // count of the number of contiguous delaying writes
@@ -609,15 +644,4 @@ extern Options SanitizeOptions(const std::string& db,
                               const Options& src);
 extern DBOptions SanitizeOptions(const std::string& db, const DBOptions& src);

-// Determine compression type, based on user options, level of the output
-// file and whether compression is disabled.
-// If enable_compression is false, then compression is always disabled no
-// matter what the values of the other two parameters are.
-// Otherwise, the compression type is determined based on options and level.
-CompressionType GetCompressionType(const Options& options, int level,
-                                   const bool enable_compression);
-
-// Determine compression type for L0 file written by memtable flush.
-CompressionType GetCompressionFlush(const Options& options);
-
 }  // namespace rocksdb
@@ -65,9 +65,7 @@ Status DBImplReadOnly::Get(const ReadOptions& options,
  if (super_version->mem->Get(lkey, value, &s, merge_context,
                              *cfd->options())) {
  } else {
-    Version::GetStats stats;
-    super_version->current->Get(options, lkey, value, &s, &merge_context,
-                                &stats);
+    super_version->current->Get(options, lkey, value, &s, &merge_context);
  }
  return s;
 }
@@ -18,6 +18,7 @@
 #include "rocksdb/iterator.h"
 #include "rocksdb/merge_operator.h"
 #include "port/port.h"
+#include "util/arena.h"
 #include "util/logging.h"
 #include "util/mutexlock.h"
 #include "util/perf_context_imp.h"
@@ -37,8 +38,6 @@ static void DumpInternalIter(Iterator* iter) {
 }
 #endif

-namespace {
-
 // Memtables and sstables that make the DB representation contain
 // (userkey,seq,type) => uservalue entries.  DBIter
 // combines multiple entries for the same userkey found in the DB
@@ -57,9 +56,10 @@ class DBIter: public Iterator {
    kReverse
  };

-  DBIter(Env* env, const Options& options,
-         const Comparator* cmp, Iterator* iter, SequenceNumber s)
-      : env_(env),
+  DBIter(Env* env, const Options& options, const Comparator* cmp,
+         Iterator* iter, SequenceNumber s, bool arena_mode)
+      : arena_mode_(arena_mode),
+        env_(env),
        logger_(options.info_log.get()),
        user_comparator_(cmp),
        user_merge_operator_(options.merge_operator.get()),
@@ -74,7 +74,15 @@ class DBIter: public Iterator {
  }
  virtual ~DBIter() {
    RecordTick(statistics_, NO_ITERATORS, -1);
-    delete iter_;
+    if (!arena_mode_) {
+      delete iter_;
+    } else {
+      iter_->~Iterator();
+    }
+  }
+  virtual void SetIter(Iterator* iter) {
+    assert(iter_ == nullptr);
+    iter_ = iter;
  }
  virtual bool Valid() const { return valid_; }
  virtual Slice key() const {
@@ -116,11 +124,12 @@ class DBIter: public Iterator {
    }
  }

+  bool arena_mode_;
  Env* const env_;
  Logger* logger_;
  const Comparator* const user_comparator_;
  const MergeOperator* const user_merge_operator_;
-  Iterator* const iter_;
+  Iterator* iter_;
  SequenceNumber const sequence_;

  Status status_;
@@ -211,18 +220,18 @@ void DBIter::FindNextUserEntryInternal(bool skipping) {
          case kTypeDeletion:
            // Arrange to skip all upcoming entries for this key since
            // they are hidden by this deletion.
-            saved_key_.SetUserKey(ikey.user_key);
+            saved_key_.SetKey(ikey.user_key);
            skipping = true;
            num_skipped = 0;
            PERF_COUNTER_ADD(internal_delete_skipped_count, 1);
            break;
          case kTypeValue:
            valid_ = true;
-            saved_key_.SetUserKey(ikey.user_key);
+            saved_key_.SetKey(ikey.user_key);
            return;
          case kTypeMerge:
            // By now, we are sure the current ikey is going to yield a value
-            saved_key_.SetUserKey(ikey.user_key);
+            saved_key_.SetKey(ikey.user_key);
            current_entry_is_merged_ = true;
            valid_ = true;
            MergeValuesNewToOld();  // Go to a different state machine
@@ -331,7 +340,7 @@ void DBIter::Prev() {
    // iter_ is pointing at the current entry.  Scan backwards until
    // the key changes so we can use the normal reverse scanning code.
    assert(iter_->Valid());  // Otherwise valid_ would have been false
-    saved_key_.SetUserKey(ExtractUserKey(iter_->key()));
+    saved_key_.SetKey(ExtractUserKey(iter_->key()));
    while (true) {
      iter_->Prev();
      if (!iter_->Valid()) {
@@ -377,7 +386,7 @@ void DBIter::FindPrevUserEntry() {
            std::string empty;
            swap(empty, saved_value_);
          }
-          saved_key_.SetUserKey(ExtractUserKey(iter_->key()));
+          saved_key_.SetKey(ExtractUserKey(iter_->key()));
          saved_value_.assign(raw_value.data(), raw_value.size());
        }
      } else {
@@ -461,16 +470,48 @@ void DBIter::SeekToLast() {
  FindPrevUserEntry();
 }

-}  // anonymous namespace
+Iterator* NewDBIterator(Env* env, const Options& options,
+                        const Comparator* user_key_comparator,
+                        Iterator* internal_iter,
+                        const SequenceNumber& sequence) {
+  return new DBIter(env, options, user_key_comparator, internal_iter, sequence,
+                    false);
+}

-Iterator* NewDBIterator(
-    Env* env,
-    const Options& options,
-    const Comparator *user_key_comparator,
-    Iterator* internal_iter,
+ArenaWrappedDBIter::~ArenaWrappedDBIter() { db_iter_->~DBIter(); }
+
+void ArenaWrappedDBIter::SetDBIter(DBIter* iter) { db_iter_ = iter; }
+
+void ArenaWrappedDBIter::SetIterUnderDBIter(Iterator* iter) {
+  static_cast<DBIter*>(db_iter_)->SetIter(iter);
+}
+
+inline bool ArenaWrappedDBIter::Valid() const { return db_iter_->Valid(); }
+inline void ArenaWrappedDBIter::SeekToFirst() { db_iter_->SeekToFirst(); }
+inline void ArenaWrappedDBIter::SeekToLast() { db_iter_->SeekToLast(); }
+inline void ArenaWrappedDBIter::Seek(const Slice& target) {
+  db_iter_->Seek(target);
+}
+inline void ArenaWrappedDBIter::Next() { db_iter_->Next(); }
+inline void ArenaWrappedDBIter::Prev() { db_iter_->Prev(); }
+inline Slice ArenaWrappedDBIter::key() const { return db_iter_->key(); }
+inline Slice ArenaWrappedDBIter::value() const { return db_iter_->value(); }
+inline Status ArenaWrappedDBIter::status() const { return db_iter_->status(); }
+void ArenaWrappedDBIter::RegisterCleanup(CleanupFunction function, void* arg1,
+                                         void* arg2) {
+  db_iter_->RegisterCleanup(function, arg1, arg2);
+}
+
+ArenaWrappedDBIter* NewArenaWrappedDbIterator(
+    Env* env, const Options& options, const Comparator* user_key_comparator,
    const SequenceNumber& sequence) {
-  return new DBIter(env, options, user_key_comparator,
-                    internal_iter, sequence);
+  ArenaWrappedDBIter* iter = new ArenaWrappedDBIter();
+  Arena* arena = iter->GetArena();
+  auto mem = arena->AllocateAligned(sizeof(DBIter));
+  DBIter* db_iter = new (mem)
+      DBIter(env, options, user_key_comparator, nullptr, sequence, true);
+  iter->SetDBIter(db_iter);
+  return iter;
 }

 }  // namespace rocksdb
@@ -11,9 +11,14 @@
 #include <stdint.h>
 #include "rocksdb/db.h"
 #include "db/dbformat.h"
+#include "util/arena.h"
+#include "util/autovector.h"

 namespace rocksdb {

+class Arena;
+class DBIter;
+
 // Return a new iterator that converts internal keys (yielded by
 // "*internal_iter") that were live at the specified "sequence" number
 // into appropriate user keys.
@@ -24,4 +29,45 @@ extern Iterator* NewDBIterator(
    Iterator* internal_iter,
    const SequenceNumber& sequence);

+// A wrapper iterator which wraps DB Iterator and the arena, with which the DB
+// iterator is supposed be allocated. This class is used as an entry point of
+// a iterator hierarchy whose memory can be allocated inline. In that way,
+// accessing the iterator tree can be more cache friendly. It is also faster
+// to allocate.
+class ArenaWrappedDBIter : public Iterator {
+ public:
+  virtual ~ArenaWrappedDBIter();
+
+  // Get the arena to be used to allocate memory for DBIter to be wrapped,
+  // as well as child iterators in it.
+  virtual Arena* GetArena() { return &arena_; }
+
+  // Set the DB Iterator to be wrapped
+
+  virtual void SetDBIter(DBIter* iter);
+
+  // Set the internal iterator wrapped inside the DB Iterator. Usually it is
+  // a merging iterator.
+  virtual void SetIterUnderDBIter(Iterator* iter);
+  virtual bool Valid() const override;
+  virtual void SeekToFirst() override;
+  virtual void SeekToLast() override;
+  virtual void Seek(const Slice& target) override;
+  virtual void Next() override;
+  virtual void Prev() override;
+  virtual Slice key() const override;
+  virtual Slice value() const override;
+  virtual Status status() const override;
+  void RegisterCleanup(CleanupFunction function, void* arg1, void* arg2);
+
+ private:
+  DBIter* db_iter_;
+  Arena arena_;
+};
+
+// Generate the arena wrapped iterator class.
+extern ArenaWrappedDBIter* NewArenaWrappedDbIterator(
+    Env* env, const Options& options, const Comparator* user_key_comparator,
+    const SequenceNumber& sequence);
+
 }  // namespace rocksdb
@@ -27,6 +27,7 @@
 #include "rocksdb/slice.h"
 #include "rocksdb/slice_transform.h"
 #include "rocksdb/table.h"
+#include "rocksdb/options.h"
 #include "rocksdb/table_properties.h"
 #include "table/block_based_table_factory.h"
 #include "table/plain_table_factory.h"
@@ -35,6 +36,7 @@
 #include "utilities/merge_operators.h"
 #include "util/logging.h"
 #include "util/mutexlock.h"
+#include "util/rate_limiter.h"
 #include "util/statistics.h"
 #include "util/testharness.h"
 #include "util/sync_point.h"
@@ -98,7 +100,12 @@ class AtomicCounter {
    count_ = 0;
  }
 };
+}  // namespace anon

+static std::string Key(int i) {
+  char buf[100];
+  snprintf(buf, sizeof(buf), "key%06d", i);
+  return std::string(buf);
 }

 // Special Env used to delay background operations
@@ -130,6 +137,8 @@ class SpecialEnv : public EnvWrapper {

  anon::AtomicCounter sleep_counter_;

+  std::atomic<int64_t> bytes_written_;
+
  explicit SpecialEnv(Env* base) : EnvWrapper(base) {
    delay_sstable_sync_.Release_Store(nullptr);
    no_space_.Release_Store(nullptr);
@@ -139,7 +148,8 @@ class SpecialEnv : public EnvWrapper {
    manifest_sync_error_.Release_Store(nullptr);
    manifest_write_error_.Release_Store(nullptr);
    log_write_error_.Release_Store(nullptr);
-   }
+    bytes_written_ = 0;
+  }

  Status NewWritableFile(const std::string& f, unique_ptr<WritableFile>* r,
                         const EnvOptions& soptions) {
@@ -158,6 +168,7 @@ class SpecialEnv : public EnvWrapper {
          // Drop writes on the floor
          return Status::OK();
        } else {
+          env_->bytes_written_ += data.size();
          return base_->Append(data);
        }
      }
@@ -169,6 +180,9 @@ class SpecialEnv : public EnvWrapper {
        }
        return base_->Sync();
      }
+      void SetIOPriority(Env::IOPriority pri) {
+        base_->SetIOPriority(pri);
+      }
    };
    class ManifestFile : public WritableFile {
     private:
@@ -355,7 +369,10 @@ class DBTest {

  ~DBTest() {
    Close();
-    ASSERT_OK(DestroyDB(dbname_, Options()));
+    Options options;
+    options.db_paths.push_back(dbname_);
+    options.db_paths.push_back(dbname_ + "_2");
+    ASSERT_OK(DestroyDB(dbname_, options));
    delete env_;
    delete filter_policy_;
  }
@@ -436,7 +453,8 @@ class DBTest {
    switch (option_config_) {
      case kHashSkipList:
        options.prefix_extractor.reset(NewFixedPrefixTransform(1));
-        options.memtable_factory.reset(NewHashSkipListRepFactory());
+        options.memtable_factory.reset(
+            NewHashSkipListRepFactory(16));
        break;
      case kPlainTableFirstBytePrefix:
        options.table_factory.reset(new PlainTableFactory());
@@ -466,7 +484,7 @@ class DBTest {
        options.db_log_dir = test::TmpDir();
        break;
      case kWalDir:
-        options.wal_dir = "/tmp/wal";
+        options.wal_dir = test::TmpDir() + "/wal";
        break;
      case kManifestFileSize:
        options.max_manifest_file_size = 50; // 50 bytes
@@ -487,7 +505,8 @@ class DBTest {
        break;
      case kHashLinkList:
        options.prefix_extractor.reset(NewFixedPrefixTransform(1));
-        options.memtable_factory.reset(NewHashLinkListRepFactory(4, 0));
+        options.memtable_factory.reset(
+            NewHashLinkListRepFactory(4, 0, 3, true, 4));
        break;
      case kHashCuckoo:
        options.memtable_factory.reset(
@@ -895,6 +914,30 @@ class DBTest {
    return property;
  }

+  int GetSstFileCount(std::string path) {
+    std::vector<std::string> files;
+    env_->GetChildren(path, &files);
+
+    int sst_count = 0;
+    uint64_t number;
+    FileType type;
+    for (size_t i = 0; i < files.size(); i++) {
+      if (ParseFileName(files[i], &number, &type) && type == kTableFile) {
+        sst_count++;
+      }
+    }
+    return sst_count;
+  }
+
+  void GenerateNewFile(Random* rnd, int* key_idx) {
+    for (int i = 0; i < 11; i++) {
+      ASSERT_OK(Put(Key(*key_idx), RandomString(rnd, (i == 10) ? 1 : 10000)));
+      (*key_idx)++;
+    }
+    dbfull()->TEST_WaitForFlushMemTable();
+    dbfull()->TEST_WaitForCompact();
+  }
+
  std::string IterStatus(Iterator* iter) {
    std::string result;
    if (iter->Valid()) {
@@ -1035,12 +1078,6 @@ class DBTest {

 };

-static std::string Key(int i) {
-  char buf[100];
-  snprintf(buf, sizeof(buf), "key%06d", i);
-  return std::string(buf);
-}
-
 static long TestGetTickerCount(const Options& options, Tickers ticker_type) {
  return options.statistics->getTickerCount(ticker_type);
 }
@@ -2318,6 +2355,7 @@ TEST(DBTest, NumImmutableMemTable) {
    std::string big_value(1000000 * 2, 'x');
    std::string num;
    SetPerfLevel(kEnableTime);;
+    ASSERT_TRUE(GetPerfLevel() == kEnableTime);

    ASSERT_OK(dbfull()->Put(writeOpt, handles_[1], "k1", big_value));
    ASSERT_TRUE(dbfull()->GetProperty(handles_[1],
@@ -2380,6 +2418,7 @@ TEST(DBTest, NumImmutableMemTable) {
    // break if we change the default skiplist implementation
    ASSERT_EQ(num, "200");
    SetPerfLevel(kDisable);
+    ASSERT_TRUE(GetPerfLevel() == kDisable);
  } while (ChangeCompactOptions());
 }

@@ -2726,6 +2765,119 @@ TEST(DBTest, CompactionTrigger) {
  ASSERT_EQ(NumTableFilesAtLevel(1, 1), 1);
 }

+namespace {
+static const int kCDTValueSize = 1000;
+static const int kCDTKeysPerBuffer = 4;
+static const int kCDTNumLevels = 8;
+Options DeletionTriggerOptions() {
+  Options options;
+  options.compression = kNoCompression;
+  options.write_buffer_size = kCDTKeysPerBuffer * (kCDTValueSize + 24);
+  options.min_write_buffer_number_to_merge = 1;
+  options.num_levels = kCDTNumLevels;
+  options.max_mem_compaction_level = 0;
+  options.level0_file_num_compaction_trigger = 1;
+  options.target_file_size_base = options.write_buffer_size * 2;
+  options.target_file_size_multiplier = 2;
+  options.max_bytes_for_level_base =
+      options.target_file_size_base * options.target_file_size_multiplier;
+  options.max_bytes_for_level_multiplier = 2;
+  options.disable_auto_compactions = false;
+  return options;
+}
+}  // anonymous namespace
+
+TEST(DBTest, CompactionDeletionTrigger) {
+  Options options = DeletionTriggerOptions();
+  options.create_if_missing = true;
+
+  for (int tid = 0; tid < 2; ++tid) {
+    uint64_t db_size[2];
+
+    DestroyAndReopen(&options);
+    Random rnd(301);
+
+    const int kTestSize = kCDTKeysPerBuffer * 512;
+    std::vector<std::string> values;
+    for (int k = 0; k < kTestSize; ++k) {
+      values.push_back(RandomString(&rnd, kCDTValueSize));
+      ASSERT_OK(Put(Key(k), values[k]));
+    }
+    dbfull()->TEST_WaitForFlushMemTable();
+    dbfull()->TEST_WaitForCompact();
+    db_size[0] = Size(Key(0), Key(kTestSize - 1));
+
+    for (int k = 0; k < kTestSize; ++k) {
+      ASSERT_OK(Delete(Key(k)));
+    }
+    dbfull()->TEST_WaitForFlushMemTable();
+    dbfull()->TEST_WaitForCompact();
+    db_size[1] = Size(Key(0), Key(kTestSize - 1));
+
+    // must have much smaller db size.
+    ASSERT_GT(db_size[0] / 3, db_size[1]);
+
+    // repeat the test with universal compaction
+    options.compaction_style = kCompactionStyleUniversal;
+    options.num_levels = 1;
+  }
+}
+
+TEST(DBTest, CompactionDeletionTriggerReopen) {
+  for (int tid = 0; tid < 2; ++tid) {
+    uint64_t db_size[3];
+    Options options = DeletionTriggerOptions();
+    options.create_if_missing = true;
+
+    DestroyAndReopen(&options);
+    Random rnd(301);
+
+    // round 1 --- insert key/value pairs.
+    const int kTestSize = kCDTKeysPerBuffer * 512;
+    std::vector<std::string> values;
+    for (int k = 0; k < kTestSize; ++k) {
+      values.push_back(RandomString(&rnd, kCDTValueSize));
+      ASSERT_OK(Put(Key(k), values[k]));
+    }
+    dbfull()->TEST_WaitForFlushMemTable();
+    dbfull()->TEST_WaitForCompact();
+    db_size[0] = Size(Key(0), Key(kTestSize - 1));
+    Close();
+
+    // round 2 --- disable auto-compactions and issue deletions.
+    options.create_if_missing = false;
+    options.disable_auto_compactions = true;
+    Reopen(&options);
+
+    for (int k = 0; k < kTestSize; ++k) {
+      ASSERT_OK(Delete(Key(k)));
+    }
+    db_size[1] = Size(Key(0), Key(kTestSize - 1));
+    Close();
+    // as auto_compaction is off, we shouldn't see too much reduce
+    // in db size.
+    ASSERT_LT(db_size[0] / 3, db_size[1]);
+
+    // round 3 --- reopen db with auto_compaction on and see if
+    // deletion compensation still work.
+    options.disable_auto_compactions = false;
+    Reopen(&options);
+    // insert relatively small amount of data to trigger auto compaction.
+    for (int k = 0; k < kTestSize / 10; ++k) {
+      ASSERT_OK(Put(Key(k), values[k]));
+    }
+    dbfull()->TEST_WaitForFlushMemTable();
+    dbfull()->TEST_WaitForCompact();
+    db_size[2] = Size(Key(0), Key(kTestSize - 1));
+    // this time we're expecting significant drop in size.
+    ASSERT_GT(db_size[0] / 3, db_size[2]);
+
+    // repeat the test with universal compaction
+    options.compaction_style = kCompactionStyleUniversal;
+    options.num_levels = 1;
+  }
+}
+
 // This is a static filter used for filtering
 // kvs during the compaction process.
 static int cfilter_count;
@@ -3319,6 +3471,13 @@ TEST(DBTest, UniversalCompactionCompressRatio2) {
  ASSERT_LT((int)dbfull()->TEST_GetLevel0TotalSize(),
            120000 * 12 * 0.8 + 120000 * 2);
 }
+
+TEST(DBTest, FailMoreDbPaths) {
+  Options options;
+  options.db_paths.push_back(dbname_);
+  options.db_paths.push_back(dbname_ + "_2");
+  ASSERT_TRUE(TryReopen(&options).IsNotSupported());
+}
 #endif

 TEST(DBTest, ConvertCompactionStyle) {
@@ -5869,69 +6028,6 @@ TEST(DBTest, ReadFirstRecordCache) {
  ASSERT_EQ(env_->sequential_read_counter_.Read(), 1);
 }

-TEST(DBTest, ReadCompaction) {
-  std::string value(4096, '4'); // a string of size 4K
-  {
-    Options options = CurrentOptions();
-    options.create_if_missing = true;
-    options.max_open_files = 20; // only 10 file in file-cache
-    options.target_file_size_base = 512;
-    options.write_buffer_size = 64 * 1024;
-    options.filter_policy = nullptr;
-    options.block_size = 4096;
-    options.no_block_cache = true;
-    options.disable_seek_compaction = false;
-
-    CreateAndReopenWithCF({"pikachu"}, &options);
-
-    // Write 8MB (2000 values, each 4K)
-    ASSERT_EQ(NumTableFilesAtLevel(0, 1), 0);
-    std::vector<std::string> values;
-    for (int i = 0; i < 2000; i++) {
-      ASSERT_OK(Put(1, Key(i), value));
-    }
-
-    // clear level 0 and 1 if necessary.
-    Flush(1);
-    dbfull()->TEST_CompactRange(0, nullptr, nullptr, handles_[1]);
-    dbfull()->TEST_CompactRange(1, nullptr, nullptr, handles_[1]);
-    ASSERT_EQ(NumTableFilesAtLevel(0, 1), 0);
-    ASSERT_EQ(NumTableFilesAtLevel(1, 1), 0);
-
-    // write some new keys into level 0
-    for (int i = 0; i < 2000; i = i + 16) {
-      ASSERT_OK(Put(1, Key(i), value));
-    }
-    Flush(1);
-
-    // Wait for any write compaction to finish
-    dbfull()->TEST_WaitForCompact();
-
-    // remember number of files in each level
-    int l1 = NumTableFilesAtLevel(0, 1);
-    int l2 = NumTableFilesAtLevel(1, 1);
-    int l3 = NumTableFilesAtLevel(2, 1);
-    ASSERT_NE(NumTableFilesAtLevel(0, 1), 0);
-    ASSERT_NE(NumTableFilesAtLevel(1, 1), 0);
-    ASSERT_NE(NumTableFilesAtLevel(2, 1), 0);
-
-    // read a bunch of times, trigger read compaction
-    for (int j = 0; j < 100; j++) {
-      for (int i = 0; i < 2000; i++) {
-        Get(1, Key(i));
-      }
-    }
-    // wait for read compaction to finish
-    env_->SleepForMicroseconds(1000000);
-
-    // verify that the number of files have decreased
-    // in some level, indicating that there was a compaction
-    ASSERT_TRUE(NumTableFilesAtLevel(0, 1) < l1 ||
-                NumTableFilesAtLevel(1, 1) < l2 ||
-                NumTableFilesAtLevel(2, 1) < l3);
-  }
-}
-
 // Multi-threaded test:
 namespace {

@@ -6641,8 +6737,7 @@ TEST(DBTest, PrefixScan) {
  options.disable_auto_compactions = true;
  options.max_background_compactions = 2;
  options.create_if_missing = true;
-  options.disable_seek_compaction = true;
-  options.memtable_factory.reset(NewHashSkipListRepFactory());
+  options.memtable_factory.reset(NewHashSkipListRepFactory(16));

  // 11 RAND I/Os
  DestroyAndReopen(&options);
@@ -6704,6 +6799,53 @@ TEST(DBTest, TailingIteratorKeepAdding) {
  }
 }

+TEST(DBTest, TailingIteratorSeekToNext) {
+  CreateAndReopenWithCF({"pikachu"});
+  ReadOptions read_options;
+  read_options.tailing = true;
+
+  std::unique_ptr<Iterator> iter(db_->NewIterator(read_options, handles_[1]));
+  std::string value(1024, 'a');
+
+  const int num_records = 1000;
+  for (int i = 1; i < num_records; ++i) {
+    char buf1[32];
+    char buf2[32];
+    snprintf(buf1, sizeof(buf1), "00a0%016d", i * 5);
+
+    Slice key(buf1, 20);
+    ASSERT_OK(Put(1, key, value));
+
+    if (i % 100 == 99) {
+      ASSERT_OK(Flush(1));
+    }
+
+    snprintf(buf2, sizeof(buf2), "00a0%016d", i * 5 - 2);
+    Slice target(buf2, 20);
+    iter->Seek(target);
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(iter->key().compare(key), 0);
+  }
+  for (int i = 2 * num_records; i > 0; --i) {
+    char buf1[32];
+    char buf2[32];
+    snprintf(buf1, sizeof(buf1), "00a0%016d", i * 5);
+
+    Slice key(buf1, 20);
+    ASSERT_OK(Put(1, key, value));
+
+    if (i % 100 == 99) {
+      ASSERT_OK(Flush(1));
+    }
+
+    snprintf(buf2, sizeof(buf2), "00a0%016d", i * 5 - 2);
+    Slice target(buf2, 20);
+    iter->Seek(target);
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(iter->key().compare(key), 0);
+  }
+}
+
 TEST(DBTest, TailingIteratorDeletes) {
  CreateAndReopenWithCF({"pikachu"});
  ReadOptions read_options;
@@ -6752,7 +6894,7 @@ TEST(DBTest, TailingIteratorPrefixSeek) {
  options.create_if_missing = true;
  options.disable_auto_compactions = true;
  options.prefix_extractor.reset(NewFixedPrefixTransform(2));
-  options.memtable_factory.reset(NewHashSkipListRepFactory());
+  options.memtable_factory.reset(NewHashSkipListRepFactory(16));
  DestroyAndReopen(&options);
  CreateAndReopenWithCF({"pikachu"}, &options);

@@ -6775,6 +6917,53 @@ TEST(DBTest, TailingIteratorPrefixSeek) {
  ASSERT_TRUE(!iter->Valid());
 }

+TEST(DBTest, TailingIteratorIncomplete) {
+  CreateAndReopenWithCF({"pikachu"});
+  ReadOptions read_options;
+  read_options.tailing = true;
+  read_options.read_tier = kBlockCacheTier;
+
+  std::string key("key");
+  std::string value("value");
+
+  ASSERT_OK(db_->Put(WriteOptions(), key, value));
+
+  std::unique_ptr<Iterator> iter(db_->NewIterator(read_options));
+  iter->SeekToFirst();
+  // we either see the entry or it's not in cache
+  ASSERT_TRUE(iter->Valid() || iter->status().IsIncomplete());
+
+  ASSERT_OK(db_->CompactRange(nullptr, nullptr));
+  iter->SeekToFirst();
+  // should still be true after compaction
+  ASSERT_TRUE(iter->Valid() || iter->status().IsIncomplete());
+}
+
+TEST(DBTest, BlockBasedTablePrefixIndexTest) {
+  // create a DB with block prefix index
+  BlockBasedTableOptions table_options;
+  Options options = CurrentOptions();
+  table_options.index_type = BlockBasedTableOptions::kHashSearch;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  options.prefix_extractor.reset(NewFixedPrefixTransform(1));
+
+
+  Reopen(&options);
+  ASSERT_OK(Put("k1", "v1"));
+  Flush();
+  ASSERT_OK(Put("k2", "v2"));
+
+  // Reopen it without prefix extractor, make sure everything still works.
+  // RocksDB should just fall back to the binary index.
+  table_options.index_type = BlockBasedTableOptions::kBinarySearch;
+  options.table_factory.reset(NewBlockBasedTableFactory(table_options));
+  options.prefix_extractor.reset();
+
+  Reopen(&options);
+  ASSERT_EQ("v1", Get("k1"));
+  ASSERT_EQ("v2", Get("k2"));
+}
+
 TEST(DBTest, ChecksumTest) {
  BlockBasedTableOptions table_options;
  Options options = CurrentOptions();
@@ -6845,6 +7034,202 @@ TEST(DBTest, FIFOCompactionTest) {
    }
  }
 }
+
+TEST(DBTest, SimpleWriteTimeoutTest) {
+  Options options;
+  options.env = env_;
+  options.create_if_missing = true;
+  options.write_buffer_size = 100000;
+  options.max_background_flushes = 0;
+  options.max_write_buffer_number = 2;
+  options.min_write_buffer_number_to_merge = 3;
+  options.max_total_wal_size = std::numeric_limits<uint64_t>::max();
+  WriteOptions write_opt = WriteOptions();
+  write_opt.timeout_hint_us = 0;
+  DestroyAndReopen(&options);
+  // fill the two write buffer
+  ASSERT_OK(Put(Key(1), Key(1) + std::string(100000, 'v'), write_opt));
+  ASSERT_OK(Put(Key(2), Key(2) + std::string(100000, 'v'), write_opt));
+  // As the only two write buffers are full in this moment, the third
+  // Put is expected to be timed-out.
+  write_opt.timeout_hint_us = 300;
+  ASSERT_TRUE(
+      Put(Key(3), Key(3) + std::string(100000, 'v'), write_opt).IsTimedOut());
+}
+
+// Multi-threaded Timeout Test
+namespace {
+
+static const int kValueSize = 1000;
+static const int kWriteBufferSize = 100000;
+
+struct TimeoutWriterState {
+  int id;
+  DB* db;
+  std::atomic<bool> done;
+  std::map<int, std::string> success_kvs;
+};
+
+static void RandomTimeoutWriter(void* arg) {
+  TimeoutWriterState* state = reinterpret_cast<TimeoutWriterState*>(arg);
+  static const uint64_t kTimerBias = 50;
+  int thread_id = state->id;
+  DB* db = state->db;
+
+  Random rnd(1000 + thread_id);
+  WriteOptions write_opt = WriteOptions();
+  write_opt.timeout_hint_us = 500;
+  int timeout_count = 0;
+  int num_keys = kNumKeys * 5;
+
+  for (int k = 0; k < num_keys; ++k) {
+    int key = k + thread_id * num_keys;
+    std::string value = RandomString(&rnd, kValueSize);
+    // only the second-half is randomized
+    if (k > num_keys / 2) {
+      switch (rnd.Next() % 5) {
+        case 0:
+          write_opt.timeout_hint_us = 500 * thread_id;
+          break;
+        case 1:
+          write_opt.timeout_hint_us = num_keys - k;
+          break;
+        case 2:
+          write_opt.timeout_hint_us = 1;
+          break;
+        default:
+          write_opt.timeout_hint_us = 0;
+          state->success_kvs.insert({key, value});
+      }
+    }
+
+    uint64_t time_before_put = db->GetEnv()->NowMicros();
+    Status s = db->Put(write_opt, Key(key), value);
+    uint64_t put_duration = db->GetEnv()->NowMicros() - time_before_put;
+    if (write_opt.timeout_hint_us == 0 ||
+        put_duration + kTimerBias < write_opt.timeout_hint_us) {
+      ASSERT_OK(s);
+      std::string result;
+    }
+    if (s.IsTimedOut()) {
+      timeout_count++;
+      ASSERT_GT(put_duration + kTimerBias, write_opt.timeout_hint_us);
+    }
+  }
+
+  state->done = true;
+}
+
+TEST(DBTest, MTRandomTimeoutTest) {
+  Options options;
+  options.env = env_;
+  options.create_if_missing = true;
+  options.max_write_buffer_number = 2;
+  options.compression = kNoCompression;
+  options.level0_slowdown_writes_trigger = 10;
+  options.level0_stop_writes_trigger = 20;
+  options.write_buffer_size = kWriteBufferSize;
+  DestroyAndReopen(&options);
+
+  TimeoutWriterState thread_states[kNumThreads];
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    thread_states[tid].id = tid;
+    thread_states[tid].db = db_;
+    thread_states[tid].done = false;
+    env_->StartThread(RandomTimeoutWriter, &thread_states[tid]);
+  }
+
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    while (thread_states[tid].done == false) {
+      env_->SleepForMicroseconds(100000);
+    }
+  }
+
+  Flush();
+
+  for (int tid = 0; tid < kNumThreads; ++tid) {
+    auto& success_kvs = thread_states[tid].success_kvs;
+    for (auto it = success_kvs.begin(); it != success_kvs.end(); ++it) {
+      ASSERT_EQ(Get(Key(it->first)), it->second);
+    }
+  }
+}
+
+}  // anonymous namespace
+
+/*
+ * This test is not reliable enough as it heavily depends on disk behavior.
+ *
+TEST(DBTest, RateLimitingTest) {
+  Options options = CurrentOptions();
+  options.write_buffer_size = 1 << 20;         // 1MB
+  options.level0_file_num_compaction_trigger = 2;
+  options.target_file_size_base = 1 << 20;     // 1MB
+  options.max_bytes_for_level_base = 4 << 20;  // 4MB
+  options.max_bytes_for_level_multiplier = 4;
+  options.compression = kNoCompression;
+  options.create_if_missing = true;
+  options.env = env_;
+  options.IncreaseParallelism(4);
+  DestroyAndReopen(&options);
+
+  WriteOptions wo;
+  wo.disableWAL = true;
+
+  // # no rate limiting
+  Random rnd(301);
+  uint64_t start = env_->NowMicros();
+  // Write ~96M data
+  for (int64_t i = 0; i < (96 << 10); ++i) {
+    ASSERT_OK(Put(RandomString(&rnd, 32),
+                  RandomString(&rnd, (1 << 10) + 1), wo));
+  }
+  uint64_t elapsed = env_->NowMicros() - start;
+  double raw_rate = env_->bytes_written_ * 1000000 / elapsed;
+  Close();
+
+  // # rate limiting with 0.7 x threshold
+  options.rate_limiter.reset(
+      NewRateLimiter(static_cast<int64_t>(0.7 * raw_rate)));
+  env_->bytes_written_ = 0;
+  DestroyAndReopen(&options);
+
+  start = env_->NowMicros();
+  // Write ~96M data
+  for (int64_t i = 0; i < (96 << 10); ++i) {
+    ASSERT_OK(Put(RandomString(&rnd, 32),
+                  RandomString(&rnd, (1 << 10) + 1), wo));
+  }
+  elapsed = env_->NowMicros() - start;
+  Close();
+  ASSERT_TRUE(options.rate_limiter->GetTotalBytesThrough() ==
+              env_->bytes_written_);
+  double ratio = env_->bytes_written_ * 1000000 / elapsed / raw_rate;
+  fprintf(stderr, "write rate ratio = %.2lf, expected 0.7\n", ratio);
+  ASSERT_TRUE(ratio > 0.6 && ratio < 0.8);
+
+  // # rate limiting with half of the raw_rate
+  options.rate_limiter.reset(
+      NewRateLimiter(static_cast<int64_t>(raw_rate / 2)));
+  env_->bytes_written_ = 0;
+  DestroyAndReopen(&options);
+
+  start = env_->NowMicros();
+  // Write ~96M data
+  for (int64_t i = 0; i < (96 << 10); ++i) {
+    ASSERT_OK(Put(RandomString(&rnd, 32),
+                  RandomString(&rnd, (1 << 10) + 1), wo));
+  }
+  elapsed = env_->NowMicros() - start;
+  Close();
+  ASSERT_TRUE(options.rate_limiter->GetTotalBytesThrough() ==
+              env_->bytes_written_);
+  ratio = env_->bytes_written_ * 1000000 / elapsed / raw_rate;
+  fprintf(stderr, "write rate ratio = %.2lf, expected 0.5\n", ratio);
+  ASSERT_TRUE(ratio > 0.4 && ratio < 0.6);
+}
+*/
+
 }  // namespace rocksdb

 int main(int argc, char** argv) {
@@ -256,24 +256,52 @@ class IterKey {

  void Clear() { key_size_ = 0; }

-  void SetUserKey(const Slice& user_key) {
-    size_t size = user_key.size();
+  void SetKey(const Slice& key) {
+    size_t size = key.size();
    EnlargeBufferIfNeeded(size);
-    memcpy(key_, user_key.data(), size);
+    memcpy(key_, key.data(), size);
    key_size_ = size;
  }

+  void SetInternalKey(const Slice& key_prefix, const Slice& user_key,
+                      SequenceNumber s,
+                      ValueType value_type = kValueTypeForSeek) {
+    size_t psize = key_prefix.size();
+    size_t usize = user_key.size();
+    EnlargeBufferIfNeeded(psize + usize + sizeof(uint64_t));
+    if (psize > 0) {
+      memcpy(key_, key_prefix.data(), psize);
+    }
+    memcpy(key_ + psize, user_key.data(), usize);
+    EncodeFixed64(key_ + usize + psize, PackSequenceAndType(s, value_type));
+    key_size_ = psize + usize + sizeof(uint64_t);
+  }
+
  void SetInternalKey(const Slice& user_key, SequenceNumber s,
                      ValueType value_type = kValueTypeForSeek) {
-    size_t usize = user_key.size();
-    EnlargeBufferIfNeeded(usize + sizeof(uint64_t));
-    memcpy(key_, user_key.data(), usize);
-    EncodeFixed64(key_ + usize, PackSequenceAndType(s, value_type));
-    key_size_ = usize + sizeof(uint64_t);
+    SetInternalKey(Slice(), user_key, s, value_type);
+  }
+
+  void Reserve(size_t size) {
+    EnlargeBufferIfNeeded(size);
+    key_size_ = size;
  }

  void SetInternalKey(const ParsedInternalKey& parsed_key) {
-    SetInternalKey(parsed_key.user_key, parsed_key.sequence, parsed_key.type);
+    SetInternalKey(Slice(), parsed_key);
+  }
+
+  void SetInternalKey(const Slice& key_prefix,
+                      const ParsedInternalKey& parsed_key_suffix) {
+    SetInternalKey(key_prefix, parsed_key_suffix.user_key,
+                   parsed_key_suffix.sequence, parsed_key_suffix.type);
+  }
+
+  void EncodeLengthPrefixedKey(const Slice& key) {
+    auto size = key.size();
+    EnlargeBufferIfNeeded(size + VarintLength(size));
+    char* ptr = EncodeVarint32(key_, size);
+    memcpy(ptr, key.data(), size);
  }

 private:
@@ -11,6 +11,7 @@

 #include <ctype.h>
 #include <stdio.h>
+#include <vector>
 #include "db/dbformat.h"
 #include "rocksdb/env.h"
 #include "util/logging.h"
@@ -66,9 +67,28 @@ std::string ArchivedLogFileName(const std::string& name, uint64_t number) {
  return MakeFileName(name + "/" + ARCHIVAL_DIR, number, "log");
 }

-std::string TableFileName(const std::string& name, uint64_t number) {
+std::string MakeTableFileName(const std::string& path, uint64_t number) {
+  return MakeFileName(path, number, "sst");
+}
+
+std::string TableFileName(const std::vector<std::string> db_paths,
+                          uint64_t number, uint32_t path_id) {
  assert(number > 0);
-  return MakeFileName(name, number, "sst");
+  std::string path;
+  if (path_id >= db_paths.size()) {
+    path = db_paths.back();
+  } else {
+    path = db_paths[path_id];
+  }
+  return MakeTableFileName(path, number);
+}
+
+std::string FormatFileNumber(uint64_t number, uint32_t path_id) {
+  if (path_id == 0) {
+    return std::to_string(number);
+  } else {
+    return std::to_string(number) + "(path " + std::to_string(path_id) + ")";
+  }
 }

 std::string DescriptorFileName(const std::string& dbname, uint64_t number) {
@@ -11,7 +11,9 @@

 #pragma once
 #include <stdint.h>
+#include <unordered_map>
 #include <string>
+#include <vector>
 #include "rocksdb/slice.h"
 #include "rocksdb/status.h"
 #include "rocksdb/transaction_log.h"
@@ -34,6 +36,9 @@ enum FileType {
  kIdentityFile
 };

+// map from file number to path ID.
+typedef std::unordered_map<uint64_t, uint32_t> FileNumToPathIdMap;
+
 // Return the name of the log file with the specified number
 // in the db named by "dbname".  The result will be prefixed with
 // "dbname".
@@ -48,10 +53,15 @@ extern std::string ArchivalDirectory(const std::string& dbname);
 extern std::string ArchivedLogFileName(const std::string& dbname,
                                       uint64_t num);

+extern std::string MakeTableFileName(const std::string& name, uint64_t number);
+
 // Return the name of the sstable with the specified number
 // in the db named by "dbname".  The result will be prefixed with
 // "dbname".
-extern std::string TableFileName(const std::string& dbname, uint64_t number);
+extern std::string TableFileName(const std::vector<std::string> db_paths,
+                                 uint64_t number, uint32_t path_id);
+
+extern std::string FormatFileNumber(uint64_t number, uint32_t path_id);

 // Return the name of the descriptor file for the db named by
 // "dbname" and the specified incarnation number.  The result will be
@@ -108,7 +108,9 @@ TEST(FileNameTest, Construction) {
  ASSERT_EQ(192U, number);
  ASSERT_EQ(kLogFile, type);

-  fname = TableFileName("bar", 200);
+  fname = TableFileName({"bar"}, 200, 0);
+  std::string fname1 = TableFileName({"foo", "bar"}, 200, 1);
+  ASSERT_EQ(fname, fname1);
  ASSERT_EQ("bar/", std::string(fname.data(), 4));
  ASSERT_TRUE(ParseFileName(fname.c_str() + 4, &number, &type));
  ASSERT_EQ(200U, number);
@@ -0,0 +1,460 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+
+#ifndef ROCKSDB_LITE
+#include "db/forward_iterator.h"
+
+#include <string>
+#include <utility>
+#include <limits>
+#include "db/db_impl.h"
+#include "db/db_iter.h"
+#include "db/column_family.h"
+#include "rocksdb/env.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/slice_transform.h"
+#include "table/merger.h"
+#include "db/dbformat.h"
+
+namespace rocksdb {
+
+// Usage:
+//     LevelIterator iter;
+//     iter.SetFileIndex(file_index);
+//     iter.Seek(target);
+//     iter.Next()
+class LevelIterator : public Iterator {
+ public:
+  LevelIterator(const ColumnFamilyData* const cfd,
+      const ReadOptions& read_options,
+      const std::vector<FileMetaData*>& files)
+    : cfd_(cfd), read_options_(read_options), files_(files), valid_(false),
+      file_index_(std::numeric_limits<uint32_t>::max()) {}
+
+  void SetFileIndex(uint32_t file_index) {
+    assert(file_index < files_.size());
+    if (file_index != file_index_) {
+      file_index_ = file_index;
+      file_iter_.reset(cfd_->table_cache()->NewIterator(
+          read_options_, *(cfd_->soptions()), cfd_->internal_comparator(),
+          files_[file_index_]->fd, nullptr /* table_reader_ptr */, false));
+    }
+    valid_ = false;
+  }
+  void SeekToLast() override {
+    status_ = Status::NotSupported("LevelIterator::SeekToLast()");
+    valid_ = false;
+  }
+  void Prev() {
+    status_ = Status::NotSupported("LevelIterator::Prev()");
+    valid_ = false;
+  }
+  bool Valid() const override {
+    return valid_;
+  }
+  void SeekToFirst() override {
+    SetFileIndex(0);
+    file_iter_->SeekToFirst();
+    valid_ = file_iter_->Valid();
+  }
+  void Seek(const Slice& internal_key) override {
+    assert(file_iter_ != nullptr);
+    file_iter_->Seek(internal_key);
+    valid_ = file_iter_->Valid();
+    assert(valid_);
+  }
+  void Next() override {
+    assert(valid_);
+    file_iter_->Next();
+    while (!file_iter_->Valid()) {
+      if (file_index_ + 1 >= files_.size()) {
+        valid_ = false;
+        return;
+      }
+      SetFileIndex(file_index_ + 1);
+      file_iter_->SeekToFirst();
+    }
+    valid_ = file_iter_->Valid();
+  }
+  Slice key() const override {
+    assert(valid_);
+    return file_iter_->key();
+  }
+  Slice value() const override {
+    assert(valid_);
+    return file_iter_->value();
+  }
+  Status status() const override {
+    if (!status_.ok()) {
+      return status_;
+    } else if (file_iter_ && !file_iter_->status().ok()) {
+      return file_iter_->status();
+    }
+    return Status::OK();
+  }
+
+ private:
+  const ColumnFamilyData* const cfd_;
+  const ReadOptions& read_options_;
+  const std::vector<FileMetaData*>& files_;
+
+  bool valid_;
+  uint32_t file_index_;
+  Status status_;
+  std::unique_ptr<Iterator> file_iter_;
+};
+
+ForwardIterator::ForwardIterator(DBImpl* db, const ReadOptions& read_options,
+                                 ColumnFamilyData* cfd)
+    : db_(db),
+      read_options_(read_options),
+      cfd_(cfd),
+      prefix_extractor_(cfd->options()->prefix_extractor.get()),
+      user_comparator_(cfd->user_comparator()),
+      immutable_min_heap_(MinIterComparator(&cfd_->internal_comparator())),
+      sv_(nullptr),
+      mutable_iter_(nullptr),
+      current_(nullptr),
+      valid_(false),
+      is_prev_set_(false) {}
+
+ForwardIterator::~ForwardIterator() {
+  Cleanup();
+}
+
+void ForwardIterator::Cleanup() {
+  delete mutable_iter_;
+  for (auto* m : imm_iters_) {
+    delete m;
+  }
+  imm_iters_.clear();
+  for (auto* f : l0_iters_) {
+    delete f;
+  }
+  l0_iters_.clear();
+  for (auto* l : level_iters_) {
+    delete l;
+  }
+  level_iters_.clear();
+
+  if (sv_ != nullptr && sv_->Unref()) {
+    DBImpl::DeletionState deletion_state;
+    db_->mutex_.Lock();
+    sv_->Cleanup();
+    db_->FindObsoleteFiles(deletion_state, false, true);
+    db_->mutex_.Unlock();
+    delete sv_;
+    if (deletion_state.HaveSomethingToDelete()) {
+      db_->PurgeObsoleteFiles(deletion_state);
+    }
+  }
+}
+
+bool ForwardIterator::Valid() const {
+  return valid_;
+}
+
+void ForwardIterator::SeekToFirst() {
+  if (sv_ == nullptr ||
+      sv_ ->version_number != cfd_->GetSuperVersionNumber()) {
+    RebuildIterators();
+  }
+  SeekInternal(Slice(), true);
+}
+
+void ForwardIterator::Seek(const Slice& internal_key) {
+  if (sv_ == nullptr ||
+      sv_ ->version_number != cfd_->GetSuperVersionNumber()) {
+    RebuildIterators();
+  }
+  SeekInternal(internal_key, false);
+}
+
+void ForwardIterator::SeekInternal(const Slice& internal_key,
+                                   bool seek_to_first) {
+  // mutable
+  seek_to_first ? mutable_iter_->SeekToFirst() :
+                  mutable_iter_->Seek(internal_key);
+
+  // immutable
+  // TODO(ljin): NeedToSeekImmutable has negative impact on performance
+  // if it turns to need to seek immutable often. We probably want to have
+  // an option to turn it off.
+  if (seek_to_first || NeedToSeekImmutable(internal_key)) {
+    {
+      auto tmp = MinIterHeap(MinIterComparator(&cfd_->internal_comparator()));
+      immutable_min_heap_.swap(tmp);
+    }
+    for (auto* m : imm_iters_) {
+      seek_to_first ? m->SeekToFirst() : m->Seek(internal_key);
+      if (m->Valid()) {
+        immutable_min_heap_.push(m);
+      }
+    }
+
+    Slice user_key;
+    if (!seek_to_first) {
+      user_key = ExtractUserKey(internal_key);
+    }
+    auto* files = sv_->current->files_;
+    for (uint32_t i = 0; i < files[0].size(); ++i) {
+      if (seek_to_first) {
+        l0_iters_[i]->SeekToFirst();
+      } else {
+        // If the target key passes over the larget key, we are sure Next()
+        // won't go over this file.
+        if (user_comparator_->Compare(user_key,
+              files[0][i]->largest.user_key()) > 0) {
+          continue;
+        }
+        l0_iters_[i]->Seek(internal_key);
+      }
+      if (l0_iters_[i]->Valid()) {
+        immutable_min_heap_.push(l0_iters_[i]);
+      }
+    }
+
+    int32_t search_left_bound = 0;
+    int32_t search_right_bound = FileIndexer::kLevelMaxIndex;
+    for (int32_t level = 1; level < sv_->current->NumberLevels(); ++level) {
+      if (files[level].empty()) {
+        search_left_bound = 0;
+        search_right_bound = FileIndexer::kLevelMaxIndex;
+        continue;
+      }
+      assert(level_iters_[level - 1] != nullptr);
+      uint32_t f_idx = 0;
+      if (!seek_to_first) {
+        // TODO(ljin): remove before committing
+        // f_idx = FindFileInRange(
+        //    files[level], internal_key, 0, files[level].size());
+
+        if (search_left_bound == search_right_bound) {
+          f_idx = search_left_bound;
+        } else if (search_left_bound < search_right_bound) {
+          f_idx = FindFileInRange(
+              files[level], internal_key, search_left_bound,
+              search_right_bound == FileIndexer::kLevelMaxIndex ?
+                files[level].size() : search_right_bound);
+        } else {
+          // search_left_bound > search_right_bound
+          // There are only 2 cases this can happen:
+          // (1) target key is smaller than left most file
+          // (2) target key is larger than right most file
+          assert(search_left_bound == (int32_t)files[level].size() ||
+                 search_right_bound == -1);
+          if (search_right_bound == -1) {
+            assert(search_left_bound == 0);
+            f_idx = 0;
+          } else {
+            sv_->current->file_indexer_.GetNextLevelIndex(
+                level, files[level].size() - 1,
+                1, 1, &search_left_bound, &search_right_bound);
+            continue;
+          }
+        }
+
+        // Prepare hints for the next level
+        if (f_idx < files[level].size()) {
+          int cmp_smallest = user_comparator_->Compare(
+              user_key, files[level][f_idx]->smallest.user_key());
+          int cmp_largest = -1;
+          if (cmp_smallest >= 0) {
+            cmp_smallest = user_comparator_->Compare(
+                user_key, files[level][f_idx]->smallest.user_key());
+          }
+          sv_->current->file_indexer_.GetNextLevelIndex(level, f_idx,
+              cmp_smallest, cmp_largest,
+              &search_left_bound, &search_right_bound);
+        } else {
+          sv_->current->file_indexer_.GetNextLevelIndex(
+              level, files[level].size() - 1,
+              1, 1, &search_left_bound, &search_right_bound);
+        }
+      }
+
+      // Seek
+      if (f_idx < files[level].size()) {
+        level_iters_[level - 1]->SetFileIndex(f_idx);
+        seek_to_first ? level_iters_[level - 1]->SeekToFirst() :
+                        level_iters_[level - 1]->Seek(internal_key);
+        if (level_iters_[level - 1]->Valid()) {
+          immutable_min_heap_.push(level_iters_[level - 1]);
+        }
+      }
+    }
+
+    if (seek_to_first || immutable_min_heap_.empty()) {
+      is_prev_set_ = false;
+    } else {
+      prev_key_.SetKey(internal_key);
+      is_prev_set_ = true;
+    }
+  }
+
+  UpdateCurrent();
+}
+
+void ForwardIterator::Next() {
+  assert(valid_);
+
+  if (sv_ == nullptr ||
+      sv_ ->version_number != cfd_->GetSuperVersionNumber()) {
+    std::string current_key = key().ToString();
+    Slice old_key(current_key.data(), current_key.size());
+
+    RebuildIterators();
+    SeekInternal(old_key, false);
+    if (!valid_ || key().compare(old_key) != 0) {
+      return;
+    }
+  } else if (current_ != mutable_iter_) {
+    // It is going to advance immutable iterator
+    prev_key_.SetKey(current_->key());
+    is_prev_set_ = true;
+  }
+
+  current_->Next();
+  if (current_->Valid() && current_ != mutable_iter_) {
+    immutable_min_heap_.push(current_);
+  }
+  UpdateCurrent();
+}
+
+Slice ForwardIterator::key() const {
+  assert(valid_);
+  return current_->key();
+}
+
+Slice ForwardIterator::value() const {
+  assert(valid_);
+  return current_->value();
+}
+
+Status ForwardIterator::status() const {
+  if (!status_.ok()) {
+    return status_;
+  } else if (!mutable_iter_->status().ok()) {
+    return mutable_iter_->status();
+  }
+
+  for (auto *it : imm_iters_) {
+    if (it && !it->status().ok()) {
+      return it->status();
+    }
+  }
+  for (auto *it : l0_iters_) {
+    if (it && !it->status().ok()) {
+      return it->status();
+    }
+  }
+  for (auto *it : level_iters_) {
+    if (it && !it->status().ok()) {
+      return it->status();
+    }
+  }
+
+  return Status::OK();
+}
+
+void ForwardIterator::RebuildIterators() {
+  // Clean up
+  Cleanup();
+  // New
+  sv_ = cfd_->GetReferencedSuperVersion(&(db_->mutex_));
+  mutable_iter_ = sv_->mem->NewIterator(read_options_);
+  sv_->imm->AddIterators(read_options_, &imm_iters_);
+  const auto& l0_files = sv_->current->files_[0];
+  l0_iters_.reserve(l0_files.size());
+  for (const auto* l0 : l0_files) {
+    l0_iters_.push_back(cfd_->table_cache()->NewIterator(
+        read_options_, *cfd_->soptions(), cfd_->internal_comparator(), l0->fd));
+  }
+  level_iters_.reserve(sv_->current->NumberLevels() - 1);
+  for (int32_t level = 1; level < sv_->current->NumberLevels(); ++level) {
+    if (sv_->current->files_[level].empty()) {
+      level_iters_.push_back(nullptr);
+    } else {
+      level_iters_.push_back(new LevelIterator(cfd_, read_options_,
+          sv_->current->files_[level]));
+    }
+  }
+
+  current_ = nullptr;
+  is_prev_set_ = false;
+}
+
+void ForwardIterator::UpdateCurrent() {
+  if (immutable_min_heap_.empty() && !mutable_iter_->Valid()) {
+    current_ = nullptr;
+  } else if (immutable_min_heap_.empty()) {
+    current_ = mutable_iter_;
+  } else if (!mutable_iter_->Valid()) {
+    current_ = immutable_min_heap_.top();
+    immutable_min_heap_.pop();
+  } else {
+    current_ = immutable_min_heap_.top();
+    assert(current_ != nullptr);
+    assert(current_->Valid());
+    int cmp = cfd_->internal_comparator().InternalKeyComparator::Compare(
+        mutable_iter_->key(), current_->key());
+    assert(cmp != 0);
+    if (cmp > 0) {
+      immutable_min_heap_.pop();
+    } else {
+      current_ = mutable_iter_;
+    }
+  }
+  valid_ = (current_ != nullptr);
+  if (!status_.ok()) {
+    status_ = Status::OK();
+  }
+}
+
+bool ForwardIterator::NeedToSeekImmutable(const Slice& target) {
+  if (!is_prev_set_) {
+    return true;
+  }
+  Slice prev_key = prev_key_.GetKey();
+  if (prefix_extractor_ && prefix_extractor_->Transform(target).compare(
+    prefix_extractor_->Transform(prev_key)) != 0) {
+    return true;
+  }
+  if (cfd_->internal_comparator().InternalKeyComparator::Compare(
+        prev_key, target) >= 0) {
+    return true;
+  }
+  if (immutable_min_heap_.empty() ||
+      cfd_->internal_comparator().InternalKeyComparator::Compare(
+          target, current_ == mutable_iter_ ? immutable_min_heap_.top()->key()
+                                            : current_->key()) > 0) {
+    return true;
+  }
+  return false;
+}
+
+uint32_t ForwardIterator::FindFileInRange(
+    const std::vector<FileMetaData*>& files, const Slice& internal_key,
+    uint32_t left, uint32_t right) {
+  while (left < right) {
+    uint32_t mid = (left + right) / 2;
+    const FileMetaData* f = files[mid];
+    if (cfd_->internal_comparator().InternalKeyComparator::Compare(
+          f->largest.Encode(), internal_key) < 0) {
+      // Key at "mid.largest" is < "target".  Therefore all
+      // files at or before "mid" are uninteresting.
+      left = mid + 1;
+    } else {
+      // Key at "mid.largest" is >= "target".  Therefore all files
+      // after "mid" are uninteresting.
+      right = mid;
+    }
+  }
+  return right;
+}
+
+}  // namespace rocksdb
+
+#endif  // ROCKSDB_LITE
@@ -0,0 +1,105 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+#pragma once
+
+#ifndef ROCKSDB_LITE
+
+#include <string>
+#include <vector>
+#include <queue>
+
+#include "rocksdb/db.h"
+#include "rocksdb/iterator.h"
+#include "rocksdb/options.h"
+#include "db/dbformat.h"
+
+namespace rocksdb {
+
+class DBImpl;
+class Env;
+struct SuperVersion;
+class ColumnFamilyData;
+class LevelIterator;
+struct FileMetaData;
+
+class MinIterComparator {
+ public:
+  explicit MinIterComparator(const Comparator* comparator) :
+    comparator_(comparator) {}
+
+  bool operator()(Iterator* a, Iterator* b) {
+    return comparator_->Compare(a->key(), b->key()) > 0;
+  }
+ private:
+  const Comparator* comparator_;
+};
+
+typedef std::priority_queue<Iterator*,
+          std::vector<Iterator*>,
+          MinIterComparator> MinIterHeap;
+
+/**
+ * ForwardIterator is a special type of iterator that only supports Seek()
+ * and Next(). It is expected to perform better than TailingIterator by
+ * removing the encapsulation and making all information accessible within
+ * the iterator. At the current implementation, snapshot is taken at the
+ * time Seek() is called. The Next() followed do not see new values after.
+ */
+class ForwardIterator : public Iterator {
+ public:
+  ForwardIterator(DBImpl* db, const ReadOptions& read_options,
+                  ColumnFamilyData* cfd);
+  virtual ~ForwardIterator();
+
+  void SeekToLast() override {
+    status_ = Status::NotSupported("ForwardIterator::SeekToLast()");
+    valid_ = false;
+  }
+  void Prev() {
+    status_ = Status::NotSupported("ForwardIterator::Prev");
+    valid_ = false;
+  }
+
+  virtual bool Valid() const override;
+  void SeekToFirst() override;
+  virtual void Seek(const Slice& target) override;
+  virtual void Next() override;
+  virtual Slice key() const override;
+  virtual Slice value() const override;
+  virtual Status status() const override;
+
+ private:
+  void Cleanup();
+  void RebuildIterators();
+  void SeekInternal(const Slice& internal_key, bool seek_to_first);
+  void UpdateCurrent();
+  bool NeedToSeekImmutable(const Slice& internal_key);
+  uint32_t FindFileInRange(
+    const std::vector<FileMetaData*>& files, const Slice& internal_key,
+    uint32_t left, uint32_t right);
+
+  DBImpl* const db_;
+  const ReadOptions read_options_;
+  ColumnFamilyData* const cfd_;
+  const SliceTransform* const prefix_extractor_;
+  const Comparator* user_comparator_;
+  MinIterHeap immutable_min_heap_;
+
+  SuperVersion* sv_;
+  Iterator* mutable_iter_;
+  std::vector<Iterator*> imm_iters_;
+  std::vector<Iterator*> l0_iters_;
+  std::vector<LevelIterator*> level_iters_;
+  Iterator* current_;
+  // internal iterator status
+  Status status_;
+  bool valid_;
+
+  IterKey prev_key_;
+  bool is_prev_set_;
+};
+
+}  // namespace rocksdb
+#endif  // ROCKSDB_LITE
@@ -294,9 +294,9 @@ bool InternalStats::GetProperty(DBPropertyType property_type,
      snprintf(
          buf, sizeof(buf),
          "Amplification interval: %.1f write, %.1f compaction\n",
-          (double)(interval_bytes_written + wal_bytes) /
+          (double)(interval_bytes_written + interval_wal_bytes) /
              (interval_bytes_new + 1),
-          (double)(interval_bytes_written + interval_bytes_read + wal_bytes) /
+          (double)(interval_bytes_written + interval_bytes_read + interval_wal_bytes) /
              (interval_bytes_new + 1));
      value->append(buf);

@@ -118,12 +118,12 @@ class InternalStats {

  void RecordWriteStall(WriteStallType write_stall_type, uint64_t micros) {
    stall_micros_[write_stall_type] += micros;
-    stall_counts_[write_stall_type]++;
+    ++stall_counts_[write_stall_type];
  }

  void RecordLevelNSlowdown(int level, uint64_t micros) {
    stall_leveln_slowdown_[level] += micros;
-    stall_leveln_slowdown_count_[level] += micros;
+    ++stall_leveln_slowdown_count_[level];
  }

  uint64_t GetBackgroundErrorCount() const { return bg_error_count_; }
@@ -51,7 +51,7 @@ void BM_LogAndApply(int iters, int num_base_files) {
    for (int i = 0; i < num_base_files; i++) {
      InternalKey start(MakeKey(2 * fnum), 1, kTypeValue);
      InternalKey limit(MakeKey(2 * fnum + 1), 1, kTypeDeletion);
-      vbase.AddFile(2, ++fnum, 1 /* file size */, start, limit, 1, 1);
+      vbase.AddFile(2, ++fnum, 0, 1 /* file size */, start, limit, 1, 1);
    }
    ASSERT_OK(vset->LogAndApply(default_cfd, &vbase, &mu));
  }
@@ -61,7 +61,7 @@ void BM_LogAndApply(int iters, int num_base_files) {
    vedit.DeleteFile(2, fnum);
    InternalKey start(MakeKey(2 * fnum), 1, kTypeValue);
    InternalKey limit(MakeKey(2 * fnum + 1), 1, kTypeDeletion);
-    vedit.AddFile(2, ++fnum, 1 /* file size */, start, limit, 1, 1);
+    vedit.AddFile(2, ++fnum, 0, 1 /* file size */, start, limit, 1, 1);
    vset->LogAndApply(default_cfd, &vedit, &mu);
  }
 }
@@ -20,6 +20,7 @@
 #include "rocksdb/iterator.h"
 #include "rocksdb/merge_operator.h"
 #include "rocksdb/slice_transform.h"
+#include "table/merger.h"
 #include "util/arena.h"
 #include "util/coding.h"
 #include "util/murmurhash.h"
@@ -54,6 +55,7 @@ MemTable::MemTable(const InternalKeyComparator& cmp, const Options& options)
  assert(!should_flush_);
  if (prefix_extractor_ && options.memtable_prefix_bloom_bits > 0) {
    prefix_bloom_.reset(new DynamicBloom(
+        &arena_,
        options.memtable_prefix_bloom_bits, options.bloom_locality,
        options.memtable_prefix_bloom_probes, nullptr,
        options.memtable_prefix_bloom_huge_page_tlb_size,
@@ -173,15 +175,24 @@ const char* EncodeKey(std::string* scratch, const Slice& target) {
 class MemTableIterator: public Iterator {
 public:
  MemTableIterator(const MemTable& mem, const ReadOptions& options,
-                   bool enforce_total_order)
+                   bool enforce_total_order, Arena* arena)
      : bloom_(nullptr),
        prefix_extractor_(mem.prefix_extractor_),
-        valid_(false) {
+        valid_(false),
+        arena_mode_(arena != nullptr) {
    if (prefix_extractor_ != nullptr && !enforce_total_order) {
      bloom_ = mem.prefix_bloom_.get();
-      iter_.reset(mem.table_->GetDynamicPrefixIterator());
+      iter_ = mem.table_->GetDynamicPrefixIterator(arena);
    } else {
-      iter_.reset(mem.table_->GetIterator());
+      iter_ = mem.table_->GetIterator(arena);
+    }
+  }
+
+  ~MemTableIterator() {
+    if (arena_mode_) {
+      iter_->~Iterator();
+    } else {
+      delete iter_;
    }
  }

@@ -228,8 +239,9 @@ class MemTableIterator: public Iterator {
 private:
  DynamicBloom* bloom_;
  const SliceTransform* const prefix_extractor_;
-  std::unique_ptr<MemTableRep::Iterator> iter_;
+  MemTableRep::Iterator* iter_;
  bool valid_;
+  bool arena_mode_;

  // No copying allowed
  MemTableIterator(const MemTableIterator&);
@@ -237,8 +249,14 @@ class MemTableIterator: public Iterator {
 };

 Iterator* MemTable::NewIterator(const ReadOptions& options,
-    bool enforce_total_order) {
-  return new MemTableIterator(*this, options, enforce_total_order);
+                                bool enforce_total_order, Arena* arena) {
+  if (arena == nullptr) {
+    return new MemTableIterator(*this, options, enforce_total_order, nullptr);
+  } else {
+    auto mem = arena->AllocateAligned(sizeof(MemTableIterator));
+    return new (mem)
+        MemTableIterator(*this, options, enforce_total_order, arena);
+  }
 }

 port::RWMutex* MemTable::GetLock(const Slice& key) {
@@ -349,7 +367,7 @@ static bool SaveValue(void* arg, const char* entry) {
          s->value->assign(v.data(), v.size());
        }
        if (s->inplace_update_support) {
-          s->mem->GetLock(s->key->user_key())->Unlock();
+          s->mem->GetLock(s->key->user_key())->ReadUnlock();
        }
        *(s->found_final_value) = true;
        return false;
@@ -432,7 +450,7 @@ void MemTable::Update(SequenceNumber seq,
  Slice mem_key = lkey.memtable_key();

  std::unique_ptr<MemTableRep::Iterator> iter(
-    table_->GetIterator(lkey.user_key()));
+      table_->GetDynamicPrefixIterator());
  iter->Seek(lkey.internal_key(), mem_key.data());

  if (iter->Valid()) {
@@ -491,7 +509,7 @@ bool MemTable::UpdateCallback(SequenceNumber seq,
  Slice memkey = lkey.memtable_key();

  std::unique_ptr<MemTableRep::Iterator> iter(
-    table_->GetIterator(lkey.user_key()));
+      table_->GetDynamicPrefixIterator());
  iter->Seek(lkey.internal_key(), memkey.data());

  if (iter->Valid()) {
@@ -566,7 +584,7 @@ size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key) {
  // reps). By passing in the user key, we allow efficient iterator creation.
  // The iterator only needs to be ordered within the same user key.
  std::unique_ptr<MemTableRep::Iterator> iter(
-      table_->GetIterator(key.user_key()));
+      table_->GetDynamicPrefixIterator());
  iter->Seek(key.internal_key(), memkey.data());

  size_t num_successive_merges = 0;
@@ -593,7 +611,7 @@ size_t MemTable::CountSuccessiveMergeEntries(const LookupKey& key) {

 void MemTableRep::Get(const LookupKey& k, void* callback_args,
                      bool (*callback_func)(void* arg, const char* entry)) {
-  auto iter = GetIterator(k.user_key());
+  auto iter = GetDynamicPrefixIterator();
  for (iter->Seek(k.internal_key(), k.memtable_key().data());
       iter->Valid() && callback_func(callback_args, iter->key());
       iter->Next()) {
@@ -21,6 +21,7 @@

 namespace rocksdb {

+class Arena;
 class Mutex;
 class MemTableIterator;
 class MergeContext;
@@ -77,8 +78,12 @@ class MemTable {
  //
  // By default, it returns an iterator for prefix seek if prefix_extractor
  // is configured in Options.
+  // arena: If not null, the arena needs to be used to allocate the Iterator.
+  //        Calling ~Iterator of the iterator will destroy all the states but
+  //        those allocated in arena.
  Iterator* NewIterator(const ReadOptions& options,
-                        bool enforce_total_order = false);
+                        bool enforce_total_order = false,
+                        Arena* arena = nullptr);

  // Add an entry into memtable that maps key to value at the
  // specified sequence number and with the specified type.
@@ -11,6 +11,7 @@
 #include "db/version_set.h"
 #include "rocksdb/env.h"
 #include "rocksdb/iterator.h"
+#include "table/merger.h"
 #include "util/coding.h"
 #include "util/log_buffer.h"

@@ -78,6 +79,14 @@ void MemTableListVersion::AddIterators(const ReadOptions& options,
  }
 }

+void MemTableListVersion::AddIterators(
+    const ReadOptions& options, MergeIteratorBuilder* merge_iter_builder) {
+  for (auto& m : memlist_) {
+    merge_iter_builder->AddIterator(
+        m->NewIterator(options, merge_iter_builder->GetArena()));
+  }
+}
+
 uint64_t MemTableListVersion::GetTotalNumEntries() const {
  uint64_t total_num = 0;
  for (auto& m : memlist_) {
@@ -131,7 +140,7 @@ void MemTableList::PickMemtablesToFlush(autovector<MemTable*>* ret) {

 void MemTableList::RollbackMemtableFlush(const autovector<MemTable*>& mems,
                                         uint64_t file_number,
-                                         std::set<uint64_t>* pending_outputs) {
+                                         FileNumToPathIdMap* pending_outputs) {
  assert(!mems.empty());

  // If the flush was not successful, then just reset state.
@@ -153,7 +162,7 @@ void MemTableList::RollbackMemtableFlush(const autovector<MemTable*>& mems,
 Status MemTableList::InstallMemtableFlushResults(
    ColumnFamilyData* cfd, const autovector<MemTable*>& mems, VersionSet* vset,
    port::Mutex* mu, Logger* info_log, uint64_t file_number,
-    std::set<uint64_t>& pending_outputs, autovector<MemTable*>* to_delete,
+    FileNumToPathIdMap* pending_outputs, autovector<MemTable*>* to_delete,
    Directory* db_directory, LogBuffer* log_buffer) {
  mu->AssertHeld();

@@ -210,7 +219,7 @@ Status MemTableList::InstallMemtableFlushResults(
        // has been written to a committed version so that other concurrently
        // executing compaction threads do not mistakenly assume that this
        // file is not live.
-        pending_outputs.erase(m->file_number_);
+        pending_outputs->erase(m->file_number_);
        if (m->Unref() != nullptr) {
          to_delete->push_back(m);
        }
@@ -224,7 +233,7 @@ Status MemTableList::InstallMemtableFlushResults(
        m->flush_in_progress_ = false;
        m->edit_.Clear();
        num_flush_not_started_++;
-        pending_outputs.erase(m->file_number_);
+        pending_outputs->erase(m->file_number_);
        m->file_number_ = 0;
        imm_flush_needed.Release_Store((void *)1);
      }
@@ -15,6 +15,7 @@
 #include "rocksdb/iterator.h"

 #include "db/dbformat.h"
+#include "db/filename.h"
 #include "db/skiplist.h"
 #include "db/memtable.h"
 #include "rocksdb/db.h"
@@ -28,6 +29,7 @@ namespace rocksdb {
 class ColumnFamilyData;
 class InternalKeyComparator;
 class Mutex;
+class MergeIteratorBuilder;

 // keeps a list of immutable memtables in a vector. the list is immutable
 // if refcount is bigger than one. It is used as a state for Get() and
@@ -49,6 +51,9 @@ class MemTableListVersion {
  void AddIterators(const ReadOptions& options,
                    std::vector<Iterator*>* iterator_list);

+  void AddIterators(const ReadOptions& options,
+                    MergeIteratorBuilder* merge_iter_builder);
+
  uint64_t GetTotalNumEntries() const;

 private:
@@ -104,17 +109,14 @@ class MemTableList {
  // they can get picked up again on the next round of flush.
  void RollbackMemtableFlush(const autovector<MemTable*>& mems,
                             uint64_t file_number,
-                             std::set<uint64_t>* pending_outputs);
+                             FileNumToPathIdMap* pending_outputs);

  // Commit a successful flush in the manifest file
-  Status InstallMemtableFlushResults(ColumnFamilyData* cfd,
-                                     const autovector<MemTable*>& m,
-                                     VersionSet* vset, port::Mutex* mu,
-                                     Logger* info_log, uint64_t file_number,
-                                     std::set<uint64_t>& pending_outputs,
-                                     autovector<MemTable*>* to_delete,
-                                     Directory* db_directory,
-                                     LogBuffer* log_buffer);
+  Status InstallMemtableFlushResults(
+      ColumnFamilyData* cfd, const autovector<MemTable*>& m, VersionSet* vset,
+      port::Mutex* mu, Logger* info_log, uint64_t file_number,
+      FileNumToPathIdMap* pending_outputs, autovector<MemTable*>* to_delete,
+      Directory* db_directory, LogBuffer* log_buffer);

  // New memtables are inserted at the front of the list.
  // Takes ownership of the referenced held on *m by the caller of Add().
@@ -61,7 +61,8 @@ class PlainTableDBTest {
  // Return the current option configuration.
  Options CurrentOptions() {
    Options options;
-    options.table_factory.reset(NewPlainTableFactory(16, 2, 0.8, 3));
+    options.table_factory.reset(NewPlainTableFactory(0, 2, 0.8, 3, 0, kPrefix));
+    options.memtable_factory.reset(NewHashLinkListRepFactory(4, 0, 3, true, 3));
    options.prefix_extractor.reset(NewFixedPrefixTransform(8));
    options.allow_mmap_reads = true;
    return options;
@@ -178,16 +179,21 @@ class TestPlainTableReader : public PlainTableReader {
 public:
  TestPlainTableReader(const EnvOptions& storage_options,
                       const InternalKeyComparator& icomparator,
-                       uint64_t file_size, int bloom_bits_per_key,
-                       double hash_table_ratio, size_t index_sparseness,
+                       EncodingType encoding_type, uint64_t file_size,
+                       int bloom_bits_per_key, double hash_table_ratio,
+                       size_t index_sparseness,
                       const TableProperties* table_properties,
                       unique_ptr<RandomAccessFile>&& file,
                       const Options& options, bool* expect_bloom_not_match)
      : PlainTableReader(options, std::move(file), storage_options, icomparator,
-                         file_size, bloom_bits_per_key, hash_table_ratio,
-                         index_sparseness, table_properties, 2 * 1024 * 1024),
+                         encoding_type, file_size, table_properties),
        expect_bloom_not_match_(expect_bloom_not_match) {
-    Status s = PopulateIndex(const_cast<TableProperties*>(table_properties));
+    Status s = MmapDataFile();
+    ASSERT_TRUE(s.ok());
+
+    s = PopulateIndex(const_cast<TableProperties*>(table_properties),
+                      bloom_bits_per_key, hash_table_ratio, index_sparseness,
+                      2 * 1024 * 1024);
    ASSERT_TRUE(s.ok());
  }

@@ -209,9 +215,10 @@ class TestPlainTableFactory : public PlainTableFactory {
                                 uint32_t user_key_len, int bloom_bits_per_key,
                                 double hash_table_ratio,
                                 size_t index_sparseness,
-                                 size_t huge_page_tlb_size)
-      : PlainTableFactory(user_key_len, user_key_len, hash_table_ratio,
-                          index_sparseness, huge_page_tlb_size),
+                                 size_t huge_page_tlb_size,
+                                 EncodingType encoding_type)
+      : PlainTableFactory(user_key_len, bloom_bits_per_key, hash_table_ratio,
+                          index_sparseness, huge_page_tlb_size, encoding_type),
        bloom_bits_per_key_(bloom_bits_per_key),
        hash_table_ratio_(hash_table_ratio),
        index_sparseness_(index_sparseness),
@@ -226,10 +233,17 @@ class TestPlainTableFactory : public PlainTableFactory {
                                 options.env, options.info_log.get(), &props);
    ASSERT_TRUE(s.ok());

+    auto& user_props = props->user_collected_properties;
+    auto encoding_type_prop =
+        user_props.find(PlainTablePropertyNames::kEncodingType);
+    assert(encoding_type_prop != user_props.end());
+    EncodingType encoding_type = static_cast<EncodingType>(
+        DecodeFixed32(encoding_type_prop->second.c_str()));
+
    std::unique_ptr<PlainTableReader> new_reader(new TestPlainTableReader(
-        soptions, internal_comparator, file_size, bloom_bits_per_key_,
-        hash_table_ratio_, index_sparseness_, props, std::move(file), options,
-        expect_bloom_not_match_));
+        soptions, internal_comparator, encoding_type, file_size,
+        bloom_bits_per_key_, hash_table_ratio_, index_sparseness_, props,
+        std::move(file), options, expect_bloom_not_match_));

    *table = std::move(new_reader);
    return s;
@@ -245,6 +259,7 @@ class TestPlainTableFactory : public PlainTableFactory {
 TEST(PlainTableDBTest, Flush) {
  for (size_t huge_page_tlb_size = 0; huge_page_tlb_size <= 2 * 1024 * 1024;
       huge_page_tlb_size += 2 * 1024 * 1024) {
+    for (EncodingType encoding_type : {kPlain, kPrefix}) {
    for (int bloom_bits = 0; bloom_bits <= 117; bloom_bits += 117) {
      for (int total_order = 0; total_order <= 1; total_order++) {
        Options options = CurrentOptions();
@@ -252,11 +267,12 @@ TEST(PlainTableDBTest, Flush) {
        // Set only one bucket to force bucket conflict.
        // Test index interval for the same prefix to be 1, 2 and 4
        if (total_order) {
-          options.table_factory.reset(NewTotalOrderPlainTableFactory(
-              16, bloom_bits, 2, huge_page_tlb_size));
+          options.prefix_extractor.reset();
+          options.table_factory.reset(NewPlainTableFactory(
+              0, bloom_bits, 0, 2, huge_page_tlb_size, encoding_type));
        } else {
          options.table_factory.reset(NewPlainTableFactory(
-              16, bloom_bits, 0.75, 16, huge_page_tlb_size));
+              0, bloom_bits, 0.75, 16, huge_page_tlb_size, encoding_type));
        }
        DestroyAndReopen(&options);

@@ -272,21 +288,26 @@ TEST(PlainTableDBTest, Flush) {
        auto tp = row->second;
        ASSERT_EQ(total_order ? "4" : "12", (tp->user_collected_properties).at(
                                                "plain_table_hash_table_size"));
-        ASSERT_EQ(total_order ? "9" : "0", (tp->user_collected_properties).at(
-                                               "plain_table_sub_index_size"));
+        ASSERT_EQ("0", (tp->user_collected_properties)
+                           .at("plain_table_sub_index_size"));

        ASSERT_EQ("v3", Get("1000000000000foo"));
        ASSERT_EQ("v2", Get("0000000000000bar"));
      }
    }
+    }
  }
 }

 TEST(PlainTableDBTest, Flush2) {
  for (size_t huge_page_tlb_size = 0; huge_page_tlb_size <= 2 * 1024 * 1024;
       huge_page_tlb_size += 2 * 1024 * 1024) {
+    for (EncodingType encoding_type : {kPlain, kPrefix}) {
    for (int bloom_bits = 0; bloom_bits <= 117; bloom_bits += 117) {
      for (int total_order = 0; total_order <= 1; total_order++) {
+        if (encoding_type == kPrefix && total_order == 1) {
+          continue;
+        }
        bool expect_bloom_not_match = false;
        Options options = CurrentOptions();
        options.create_if_missing = true;
@@ -294,13 +315,13 @@ TEST(PlainTableDBTest, Flush2) {
        // Test index interval for the same prefix to be 1, 2 and 4
        if (total_order) {
          options.prefix_extractor = nullptr;
-          options.table_factory.reset(
-              new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits,
-                                        0, 2, huge_page_tlb_size));
+          options.table_factory.reset(new TestPlainTableFactory(
+              &expect_bloom_not_match, 0, bloom_bits, 0, 2, huge_page_tlb_size,
+              encoding_type));
        } else {
-          options.table_factory.reset(
-              new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits,
-                                        0.75, 16, huge_page_tlb_size));
+          options.table_factory.reset(new TestPlainTableFactory(
+              &expect_bloom_not_match, 0, bloom_bits, 0.75, 16,
+              huge_page_tlb_size, encoding_type));
        }
        DestroyAndReopen(&options);
        ASSERT_OK(Put("0000000000000bar", "b"));
@@ -339,14 +360,19 @@ TEST(PlainTableDBTest, Flush2) {
        }
      }
    }
+    }
  }
 }

 TEST(PlainTableDBTest, Iterator) {
  for (size_t huge_page_tlb_size = 0; huge_page_tlb_size <= 2 * 1024 * 1024;
       huge_page_tlb_size += 2 * 1024 * 1024) {
+    for (EncodingType encoding_type : {kPlain, kPrefix}) {
    for (int bloom_bits = 0; bloom_bits <= 117; bloom_bits += 117) {
      for (int total_order = 0; total_order <= 1; total_order++) {
+        if (encoding_type == kPrefix && total_order == 1) {
+          continue;
+        }
        bool expect_bloom_not_match = false;
        Options options = CurrentOptions();
        options.create_if_missing = true;
@@ -354,13 +380,13 @@ TEST(PlainTableDBTest, Iterator) {
        // Test index interval for the same prefix to be 1, 2 and 4
        if (total_order) {
          options.prefix_extractor = nullptr;
-          options.table_factory.reset(
-              new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits,
-                                        0, 2, huge_page_tlb_size));
+          options.table_factory.reset(new TestPlainTableFactory(
+              &expect_bloom_not_match, 16, bloom_bits, 0, 2, huge_page_tlb_size,
+              encoding_type));
        } else {
-          options.table_factory.reset(
-              new TestPlainTableFactory(&expect_bloom_not_match, 16, bloom_bits,
-                                        0.75, 16, huge_page_tlb_size));
+          options.table_factory.reset(new TestPlainTableFactory(
+              &expect_bloom_not_match, 16, bloom_bits, 0.75, 16,
+              huge_page_tlb_size, encoding_type));
        }
        DestroyAndReopen(&options);

@@ -447,6 +473,7 @@ TEST(PlainTableDBTest, Iterator) {
        delete iter;
      }
    }
+    }
  }
 }

@@ -458,7 +485,7 @@ std::string MakeLongKey(size_t length, char c) {

 TEST(PlainTableDBTest, IteratorLargeKeys) {
  Options options = CurrentOptions();
-  options.table_factory.reset(NewTotalOrderPlainTableFactory(0, 0, 16));
+  options.table_factory.reset(NewPlainTableFactory(0, 0, 0));
  options.create_if_missing = true;
  options.prefix_extractor.reset();
  DestroyAndReopen(&options);
@@ -494,6 +521,45 @@ TEST(PlainTableDBTest, IteratorLargeKeys) {
  delete iter;
 }

+namespace {
+std::string MakeLongKeyWithPrefix(size_t length, char c) {
+  return "00000000" + std::string(length - 8, c);
+}
+}  // namespace
+
+TEST(PlainTableDBTest, IteratorLargeKeysWithPrefix) {
+  Options options = CurrentOptions();
+  options.table_factory.reset(NewPlainTableFactory(16, 0, 0.8, 3, 0, kPrefix));
+  options.create_if_missing = true;
+  DestroyAndReopen(&options);
+
+  std::string key_list[] = {
+      MakeLongKeyWithPrefix(30, '0'), MakeLongKeyWithPrefix(16, '1'),
+      MakeLongKeyWithPrefix(32, '2'), MakeLongKeyWithPrefix(60, '3'),
+      MakeLongKeyWithPrefix(90, '4'), MakeLongKeyWithPrefix(50, '5'),
+      MakeLongKeyWithPrefix(26, '6')};
+
+  for (size_t i = 0; i < 7; i++) {
+    ASSERT_OK(Put(key_list[i], std::to_string(i)));
+  }
+
+  dbfull()->TEST_FlushMemTable();
+
+  Iterator* iter = dbfull()->NewIterator(ReadOptions());
+  iter->Seek(key_list[0]);
+
+  for (size_t i = 0; i < 7; i++) {
+    ASSERT_TRUE(iter->Valid());
+    ASSERT_EQ(key_list[i], iter->key().ToString());
+    ASSERT_EQ(std::to_string(i), iter->value().ToString());
+    iter->Next();
+  }
+
+  ASSERT_TRUE(!iter->Valid());
+
+  delete iter;
+}
+
 // A test comparator which compare two strings in this way:
 // (1) first compare prefix of 8 bytes in alphabet order,
 // (2) if two strings share the same prefix, sort the other part of the string
@@ -600,7 +666,7 @@ TEST(PlainTableDBTest, HashBucketConflict) {
      // Set only one bucket to force bucket conflict.
      // Test index interval for the same prefix to be 1, 2 and 4
      options.table_factory.reset(
-          NewTotalOrderPlainTableFactory(16, 0, 2 ^ i, huge_page_tlb_size));
+          NewPlainTableFactory(16, 0, 0, 2 ^ i, huge_page_tlb_size));
      DestroyAndReopen(&options);
      ASSERT_OK(Put("5000000000000fo0", "v1"));
      ASSERT_OK(Put("5000000000000fo1", "v2"));
@@ -687,7 +753,7 @@ TEST(PlainTableDBTest, HashBucketConflictReverseSuffixComparator) {
      // Set only one bucket to force bucket conflict.
      // Test index interval for the same prefix to be 1, 2 and 4
      options.table_factory.reset(
-          NewTotalOrderPlainTableFactory(16, 0, 2 ^ i, huge_page_tlb_size));
+          NewPlainTableFactory(16, 0, 0, 2 ^ i, huge_page_tlb_size));
      DestroyAndReopen(&options);
      ASSERT_OK(Put("5000000000000fo0", "v1"));
      ASSERT_OK(Put("5000000000000fo1", "v2"));
@@ -767,7 +833,7 @@ TEST(PlainTableDBTest, NonExistingKeyToNonEmptyBucket) {
  options.create_if_missing = true;
  // Set only one bucket to force bucket conflict.
  // Test index interval for the same prefix to be 1, 2 and 4
-  options.table_factory.reset(NewTotalOrderPlainTableFactory(16, 0, 5));
+  options.table_factory.reset(NewPlainTableFactory(16, 0, 0, 5));
  DestroyAndReopen(&options);
  ASSERT_OK(Put("5000000000000fo0", "v1"));
  ASSERT_OK(Put("5000000000000fo1", "v2"));
@@ -846,6 +912,49 @@ TEST(PlainTableDBTest, CompactionTrigger) {
  ASSERT_EQ(NumTableFilesAtLevel(1), 1);
 }

+TEST(PlainTableDBTest, AdaptiveTable) {
+  Options options = CurrentOptions();
+  options.create_if_missing = true;
+
+  options.table_factory.reset(NewPlainTableFactory());
+  DestroyAndReopen(&options);
+
+  ASSERT_OK(Put("1000000000000foo", "v1"));
+  ASSERT_OK(Put("0000000000000bar", "v2"));
+  ASSERT_OK(Put("1000000000000foo", "v3"));
+  dbfull()->TEST_FlushMemTable();
+
+  options.create_if_missing = false;
+  std::shared_ptr<TableFactory> dummy_factory;
+  std::shared_ptr<TableFactory> block_based_factory(
+      NewBlockBasedTableFactory());
+  options.table_factory.reset(NewAdaptiveTableFactory(
+      block_based_factory, dummy_factory, dummy_factory));
+  Reopen(&options);
+  ASSERT_EQ("v3", Get("1000000000000foo"));
+  ASSERT_EQ("v2", Get("0000000000000bar"));
+
+  ASSERT_OK(Put("2000000000000foo", "v4"));
+  ASSERT_OK(Put("3000000000000bar", "v5"));
+  dbfull()->TEST_FlushMemTable();
+  ASSERT_EQ("v4", Get("2000000000000foo"));
+  ASSERT_EQ("v5", Get("3000000000000bar"));
+
+  Reopen(&options);
+  ASSERT_EQ("v3", Get("1000000000000foo"));
+  ASSERT_EQ("v2", Get("0000000000000bar"));
+  ASSERT_EQ("v4", Get("2000000000000foo"));
+  ASSERT_EQ("v5", Get("3000000000000bar"));
+
+  options.table_factory.reset(NewBlockBasedTableFactory());
+  Reopen(&options);
+  ASSERT_NE("v3", Get("1000000000000foo"));
+
+  options.table_factory.reset(NewPlainTableFactory());
+  Reopen(&options);
+  ASSERT_NE("v5", Get("3000000000000bar"));
+}
+
 }  // namespace rocksdb

 int main(int argc, char** argv) {
@@ -189,6 +189,10 @@ class PrefixTest {
          options.memtable_factory.reset(
              NewHashLinkListRepFactory(bucket_count, 2 * 1024 * 1024));
          return true;
+        case kHashLinkListTriggerSkipList:
+          options.memtable_factory.reset(
+              NewHashLinkListRepFactory(bucket_count, 0, 3));
+          return true;
        default:
          return false;
      }
@@ -208,6 +212,7 @@ class PrefixTest {
    kHashSkipList,
    kHashLinkList,
    kHashLinkListHugePageTlb,
+    kHashLinkListTriggerSkipList,
    kEnd
  };
  int option_config_;
@@ -31,6 +31,8 @@

 #ifndef ROCKSDB_LITE

+#define __STDC_FORMAT_MACROS
+#include <inttypes.h>
 #include "db/builder.h"
 #include "db/db_impl.h"
 #include "db/dbformat.h"
@@ -63,8 +65,8 @@ class Repairer {
            NewLRUCache(10, options_.table_cache_numshardbits,
                        options_.table_cache_remove_scan_count_limit)),
        next_file_number_(1) {
-    table_cache_ = new TableCache(dbname_, &options_, storage_options_,
-                                  raw_table_cache_.get());
+    table_cache_ =
+        new TableCache(&options_, storage_options_, raw_table_cache_.get());
    edit_ = new VersionEdit();
  }

@@ -82,18 +84,17 @@ class Repairer {
      status = WriteDescriptor();
    }
    if (status.ok()) {
-      unsigned long long bytes = 0;
+      uint64_t bytes = 0;
      for (size_t i = 0; i < tables_.size(); i++) {
-        bytes += tables_[i].meta.file_size;
+        bytes += tables_[i].meta.fd.GetFileSize();
      }
      Log(options_.info_log,
          "**** Repaired rocksdb %s; "
-          "recovered %d files; %llu bytes. "
+          "recovered %zu files; %" PRIu64
+          "bytes. "
          "Some data may have been lost. "
          "****",
-          dbname_.c_str(),
-          static_cast<int>(tables_.size()),
-          bytes);
+          dbname_.c_str(), tables_.size(), bytes);
    }
    return status;
  }
@@ -115,7 +116,7 @@ class Repairer {
  VersionEdit* edit_;

  std::vector<std::string> manifests_;
-  std::vector<uint64_t> table_numbers_;
+  std::vector<FileDescriptor> table_fds_;
  std::vector<uint64_t> logs_;
  std::vector<TableInfo> tables_;
  uint64_t next_file_number_;
@@ -123,35 +124,43 @@ class Repairer {

  Status FindFiles() {
    std::vector<std::string> filenames;
-    Status status = env_->GetChildren(dbname_, &filenames);
-    if (!status.ok()) {
-      return status;
-    }
-    if (filenames.empty()) {
-      return Status::Corruption(dbname_, "repair found no files");
-    }
+    bool found_file = false;
+    for (uint32_t path_id = 0; path_id < options_.db_paths.size(); path_id++) {
+      Status status = env_->GetChildren(options_.db_paths[path_id], &filenames);
+      if (!status.ok()) {
+        return status;
+      }
+      if (!filenames.empty()) {
+        found_file = true;
+      }

-    uint64_t number;
-    FileType type;
-    for (size_t i = 0; i < filenames.size(); i++) {
-      if (ParseFileName(filenames[i], &number, &type)) {
-        if (type == kDescriptorFile) {
-          manifests_.push_back(filenames[i]);
-        } else {
-          if (number + 1 > next_file_number_) {
-            next_file_number_ = number + 1;
-          }
-          if (type == kLogFile) {
-            logs_.push_back(number);
-          } else if (type == kTableFile) {
-            table_numbers_.push_back(number);
+      uint64_t number;
+      FileType type;
+      for (size_t i = 0; i < filenames.size(); i++) {
+        if (ParseFileName(filenames[i], &number, &type)) {
+          if (type == kDescriptorFile) {
+            assert(path_id == 0);
+            manifests_.push_back(filenames[i]);
          } else {
-            // Ignore other files
+            if (number + 1 > next_file_number_) {
+              next_file_number_ = number + 1;
+            }
+            if (type == kLogFile) {
+              assert(path_id == 0);
+              logs_.push_back(number);
+            } else if (type == kTableFile) {
+              table_fds_.emplace_back(number, path_id, 0);
+            } else {
+              // Ignore other files
+            }
          }
        }
      }
    }
-    return status;
+    if (!found_file) {
+      return Status::Corruption(dbname_, "repair found no files");
+    }
+    return Status::OK();
  }

  void ConvertLogFilesToTables() {
@@ -159,8 +168,8 @@ class Repairer {
      std::string logname = LogFileName(dbname_, logs_[i]);
      Status status = ConvertLogToTable(logs_[i]);
      if (!status.ok()) {
-        Log(options_.info_log, "Log #%llu: ignoring conversion error: %s",
-            (unsigned long long) logs_[i],
+        Log(options_.info_log,
+            "Log #%" PRIu64 ": ignoring conversion error: %s", logs_[i],
            status.ToString().c_str());
      }
      ArchiveFile(logname);
@@ -174,10 +183,8 @@ class Repairer {
      uint64_t lognum;
      virtual void Corruption(size_t bytes, const Status& s) {
        // We print error messages for corruption, but continue repairing.
-        Log(info_log, "Log #%llu: dropping %d bytes; %s",
-            (unsigned long long) lognum,
-            static_cast<int>(bytes),
-            s.ToString().c_str());
+        Log(info_log, "Log #%" PRIu64 ": dropping %d bytes; %s", lognum,
+            static_cast<int>(bytes), s.ToString().c_str());
      }
    };

@@ -220,8 +227,7 @@ class Repairer {
      if (status.ok()) {
        counter += WriteBatchInternal::Count(&batch);
      } else {
-        Log(options_.info_log, "Log #%llu: ignoring %s",
-            (unsigned long long) log,
+        Log(options_.info_log, "Log #%" PRIu64 ": ignoring %s", log,
            status.ToString().c_str());
        status = Status::OK();  // Keep going with rest of file
      }
@@ -230,7 +236,7 @@ class Repairer {
    // Do not record a version edit for this conversion to a Table
    // since ExtractMetaData() will also generate edits.
    FileMetaData meta;
-    meta.number = next_file_number_++;
+    meta.fd = FileDescriptor(next_file_number_++, 0, 0);
    ReadOptions ro;
    Iterator* iter = mem->NewIterator(ro, true /* enforce_total_order */);
    status = BuildTable(dbname_, env_, options_, storage_options_, table_cache_,
@@ -240,27 +246,27 @@ class Repairer {
    delete cf_mems_default;
    mem = nullptr;
    if (status.ok()) {
-      if (meta.file_size > 0) {
-        table_numbers_.push_back(meta.number);
+      if (meta.fd.GetFileSize() > 0) {
+        table_fds_.push_back(meta.fd);
      }
    }
-    Log(options_.info_log, "Log #%llu: %d ops saved to Table #%llu %s",
-        (unsigned long long) log,
-        counter,
-        (unsigned long long) meta.number,
-        status.ToString().c_str());
+    Log(options_.info_log,
+        "Log #%" PRIu64 ": %d ops saved to Table #%" PRIu64 " %s", log, counter,
+        meta.fd.GetNumber(), status.ToString().c_str());
    return status;
  }

  void ExtractMetaData() {
-    for (size_t i = 0; i < table_numbers_.size(); i++) {
+    for (size_t i = 0; i < table_fds_.size(); i++) {
      TableInfo t;
-      t.meta.number = table_numbers_[i];
+      t.meta.fd = table_fds_[i];
      Status status = ScanTable(&t);
      if (!status.ok()) {
-        std::string fname = TableFileName(dbname_, table_numbers_[i]);
-        Log(options_.info_log, "Table #%llu: ignoring %s",
-            (unsigned long long) table_numbers_[i],
+        std::string fname = TableFileName(
+            options_.db_paths, t.meta.fd.GetNumber(), t.meta.fd.GetPathId());
+        Log(options_.info_log, "Table #%s: ignoring %s",
+            FormatFileNumber(t.meta.fd.GetNumber(), t.meta.fd.GetPathId())
+                .c_str(),
            status.ToString().c_str());
        ArchiveFile(fname);
      } else {
@@ -270,13 +276,16 @@ class Repairer {
  }

  Status ScanTable(TableInfo* t) {
-    std::string fname = TableFileName(dbname_, t->meta.number);
+    std::string fname = TableFileName(options_.db_paths, t->meta.fd.GetNumber(),
+                                      t->meta.fd.GetPathId());
    int counter = 0;
-    Status status = env_->GetFileSize(fname, &t->meta.file_size);
+    uint64_t file_size;
+    Status status = env_->GetFileSize(fname, &file_size);
+    t->meta.fd = FileDescriptor(t->meta.fd.GetNumber(), t->meta.fd.GetPathId(),
+                                file_size);
    if (status.ok()) {
-      FileMetaData dummy_meta(t->meta.number, t->meta.file_size);
      Iterator* iter = table_cache_->NewIterator(
-          ReadOptions(), storage_options_, icmp_, dummy_meta);
+          ReadOptions(), storage_options_, icmp_, t->meta.fd);
      bool empty = true;
      ParsedInternalKey parsed;
      t->min_sequence = 0;
@@ -284,9 +293,8 @@ class Repairer {
      for (iter->SeekToFirst(); iter->Valid(); iter->Next()) {
        Slice key = iter->key();
        if (!ParseInternalKey(key, &parsed)) {
-          Log(options_.info_log, "Table #%llu: unparsable key %s",
-              (unsigned long long) t->meta.number,
-              EscapeString(key).c_str());
+          Log(options_.info_log, "Table #%" PRIu64 ": unparsable key %s",
+              t->meta.fd.GetNumber(), EscapeString(key).c_str());
          continue;
        }

@@ -308,10 +316,8 @@ class Repairer {
      }
      delete iter;
    }
-    Log(options_.info_log, "Table #%llu: %d entries %s",
-        (unsigned long long) t->meta.number,
-        counter,
-        status.ToString().c_str());
+    Log(options_.info_log, "Table #%" PRIu64 ": %d entries %s",
+        t->meta.fd.GetNumber(), counter, status.ToString().c_str());
    return status;
  }

@@ -339,9 +345,9 @@ class Repairer {
    for (size_t i = 0; i < tables_.size(); i++) {
      // TODO(opt): separate out into multiple levels
      const TableInfo& t = tables_[i];
-      edit_->AddFile(0, t.meta.number, t.meta.file_size,
-                    t.meta.smallest, t.meta.largest,
-                    t.min_sequence, t.max_sequence);
+      edit_->AddFile(0, t.meta.fd.GetNumber(), t.meta.fd.GetPathId(),
+                     t.meta.fd.GetFileSize(), t.meta.smallest, t.meta.largest,
+                     t.min_sequence, t.max_sequence);
    }

    //fprintf(stderr, "NewDescriptor:\n%s\n", edit_.DebugString().c_str());
@@ -83,11 +83,11 @@ public:
                     unique_ptr<RandomAccessFile> && file, uint64_t file_size,
                     unique_ptr<TableReader>* table_reader);

-  Iterator* NewIterator(const ReadOptions&) override;
+  Iterator* NewIterator(const ReadOptions&, Arena* arena) override;

  Status Get(const ReadOptions&, const Slice& key, void* arg,
             bool (*handle_result)(void* arg, const ParsedInternalKey& k,
-                                   const Slice& v, bool),
+                                   const Slice& v),
             void (*mark_key_may_exist)(void*) = nullptr) override;

  uint64_t ApproximateOffsetOf(const Slice& key) override;
@@ -218,8 +218,14 @@ std::shared_ptr<const TableProperties> SimpleTableReader::GetTableProperties()
  return rep_->table_properties;
 }

-Iterator* SimpleTableReader::NewIterator(const ReadOptions& options) {
-  return new SimpleTableIterator(this);
+Iterator* SimpleTableReader::NewIterator(const ReadOptions& options,
+                                         Arena* arena) {
+  if (arena == nullptr) {
+    return new SimpleTableIterator(this);
+  } else {
+    auto mem = arena->AllocateAligned(sizeof(SimpleTableIterator));
+    return new (mem) SimpleTableIterator(this);
+  }
 }

 Status SimpleTableReader::GetOffset(const Slice& target, uint64_t* offset) {
@@ -279,7 +285,7 @@ Status SimpleTableReader::GetOffset(const Slice& target, uint64_t* offset) {
 Status SimpleTableReader::Get(const ReadOptions& options, const Slice& k,
                              void* arg,
                              bool (*saver)(void*, const ParsedInternalKey&,
-                                            const Slice&, bool),
+                                            const Slice&),
                              void (*mark_key_may_exist)(void*)) {
  Status s;
  SimpleTableIterator* iter = new SimpleTableIterator(this);
@@ -289,7 +295,7 @@ Status SimpleTableReader::Get(const ReadOptions& options, const Slice& k,
      return Status::Corruption(Slice());
    }

-    if (!(*saver)(arg, parsed_key, iter->value(), true)) {
+    if (!(*saver)(arg, parsed_key, iter->value())) {
      break;
    }
  }
@@ -13,6 +13,7 @@
 #include "db/version_edit.h"

 #include "rocksdb/statistics.h"
+#include "table/iterator_wrapper.h"
 #include "table/table_reader.h"
 #include "util/coding.h"
 #include "util/stop_watch.h"
@@ -30,15 +31,15 @@ static void UnrefEntry(void* arg1, void* arg2) {
  cache->Release(h);
 }

-static Slice GetSliceForFileNumber(uint64_t* file_number) {
+static Slice GetSliceForFileNumber(const uint64_t* file_number) {
  return Slice(reinterpret_cast<const char*>(file_number),
               sizeof(*file_number));
 }

-TableCache::TableCache(const std::string& dbname, const Options* options,
+TableCache::TableCache(const Options* options,
                       const EnvOptions& storage_options, Cache* const cache)
    : env_(options->env),
-      dbname_(dbname),
+      db_paths_(options->db_paths),
      options_(options),
      storage_options_(storage_options),
      cache_(cache) {}
@@ -56,20 +57,18 @@ void TableCache::ReleaseHandle(Cache::Handle* handle) {

 Status TableCache::FindTable(const EnvOptions& toptions,
                             const InternalKeyComparator& internal_comparator,
-                             uint64_t file_number, uint64_t file_size,
-                             Cache::Handle** handle, bool* table_io,
+                             const FileDescriptor& fd, Cache::Handle** handle,
                             const bool no_io) {
  Status s;
-  Slice key = GetSliceForFileNumber(&file_number);
+  uint64_t number = fd.GetNumber();
+  Slice key = GetSliceForFileNumber(&number);
  *handle = cache_->Lookup(key);
  if (*handle == nullptr) {
    if (no_io) { // Dont do IO and return a not-found status
      return Status::Incomplete("Table not found in table_cache, no_io is set");
    }
-    if (table_io != nullptr) {
-      *table_io = true;    // we had to do IO from storage
-    }
-    std::string fname = TableFileName(dbname_, file_number);
+    std::string fname =
+        TableFileName(db_paths_, fd.GetNumber(), fd.GetPathId());
    unique_ptr<RandomAccessFile> file;
    unique_ptr<TableReader> table_reader;
    s = env_->NewRandomAccessFile(fname, &file, toptions);
@@ -80,8 +79,8 @@ Status TableCache::FindTable(const EnvOptions& toptions,
      }
      StopWatch sw(env_, options_->statistics.get(), TABLE_OPEN_IO_MICROS);
      s = options_->table_factory->NewTableReader(
-          *options_, toptions, internal_comparator, std::move(file), file_size,
-          &table_reader);
+          *options_, toptions, internal_comparator, std::move(file),
+          fd.GetFileSize(), &table_reader);
    }

    if (!s.ok()) {
@@ -100,25 +99,25 @@ Status TableCache::FindTable(const EnvOptions& toptions,
 Iterator* TableCache::NewIterator(const ReadOptions& options,
                                  const EnvOptions& toptions,
                                  const InternalKeyComparator& icomparator,
-                                  const FileMetaData& file_meta,
+                                  const FileDescriptor& fd,
                                  TableReader** table_reader_ptr,
-                                  bool for_compaction) {
+                                  bool for_compaction, Arena* arena) {
  if (table_reader_ptr != nullptr) {
    *table_reader_ptr = nullptr;
  }
-  TableReader* table_reader = file_meta.table_reader;
+  TableReader* table_reader = fd.table_reader;
  Cache::Handle* handle = nullptr;
  Status s;
  if (table_reader == nullptr) {
-    s = FindTable(toptions, icomparator, file_meta.number, file_meta.file_size,
-                  &handle, nullptr, options.read_tier == kBlockCacheTier);
+    s = FindTable(toptions, icomparator, fd, &handle,
+                  options.read_tier == kBlockCacheTier);
    if (!s.ok()) {
-      return NewErrorIterator(s);
+      return NewErrorIterator(s, arena);
    }
    table_reader = GetTableReaderFromHandle(handle);
  }

-  Iterator* result = table_reader->NewIterator(options);
+  Iterator* result = table_reader->NewIterator(options, arena);
  if (handle != nullptr) {
    result->RegisterCleanup(&UnrefEntry, cache_, handle);
  }
@@ -135,16 +134,15 @@ Iterator* TableCache::NewIterator(const ReadOptions& options,

 Status TableCache::Get(const ReadOptions& options,
                       const InternalKeyComparator& internal_comparator,
-                       const FileMetaData& file_meta, const Slice& k, void* arg,
+                       const FileDescriptor& fd, const Slice& k, void* arg,
                       bool (*saver)(void*, const ParsedInternalKey&,
-                                     const Slice&, bool),
-                       bool* table_io, void (*mark_key_may_exist)(void*)) {
-  TableReader* t = file_meta.table_reader;
+                                     const Slice&),
+                       void (*mark_key_may_exist)(void*)) {
+  TableReader* t = fd.table_reader;
  Status s;
  Cache::Handle* handle = nullptr;
  if (!t) {
-    s = FindTable(storage_options_, internal_comparator, file_meta.number,
-                  file_meta.file_size, &handle, table_io,
+    s = FindTable(storage_options_, internal_comparator, fd, &handle,
                  options.read_tier == kBlockCacheTier);
    if (s.ok()) {
      t = GetTableReaderFromHandle(handle);
@@ -164,11 +162,10 @@ Status TableCache::Get(const ReadOptions& options,
 }
 Status TableCache::GetTableProperties(
    const EnvOptions& toptions,
-    const InternalKeyComparator& internal_comparator,
-    const FileMetaData& file_meta,
+    const InternalKeyComparator& internal_comparator, const FileDescriptor& fd,
    std::shared_ptr<const TableProperties>* properties, bool no_io) {
  Status s;
-  auto table_reader = file_meta.table_reader;
+  auto table_reader = fd.table_reader;
  // table already been pre-loaded?
  if (table_reader) {
    *properties = table_reader->GetTableProperties();
@@ -176,10 +173,8 @@ Status TableCache::GetTableProperties(
    return s;
  }

-  bool table_io;
  Cache::Handle* table_handle = nullptr;
-  s = FindTable(toptions, internal_comparator, file_meta.number,
-                file_meta.file_size, &table_handle, &table_io, no_io);
+  s = FindTable(toptions, internal_comparator, fd, &table_handle, no_io);
  if (!s.ok()) {
    return s;
  }
@@ -11,6 +11,7 @@

 #pragma once
 #include <string>
+#include <vector>
 #include <stdint.h>

 #include "db/dbformat.h"
@@ -23,14 +24,13 @@
 namespace rocksdb {

 class Env;
-struct FileMetaData;
+class Arena;
+struct FileDescriptor;

-// TODO(sdong): try to come up with a better API to pass the file information
-//              other than simply passing FileMetaData.
 class TableCache {
 public:
-  TableCache(const std::string& dbname, const Options* options,
-             const EnvOptions& storage_options, Cache* cache);
+  TableCache(const Options* options, const EnvOptions& storage_options,
+             Cache* cache);
  ~TableCache();

  // Return an iterator for the specified file number (the corresponding
@@ -42,19 +42,19 @@ class TableCache {
  // returned iterator is live.
  Iterator* NewIterator(const ReadOptions& options, const EnvOptions& toptions,
                        const InternalKeyComparator& internal_comparator,
-                        const FileMetaData& file_meta,
+                        const FileDescriptor& file_fd,
                        TableReader** table_reader_ptr = nullptr,
-                        bool for_compaction = false);
+                        bool for_compaction = false, Arena* arena = nullptr);

  // If a seek to internal key "k" in specified file finds an entry,
  // call (*handle_result)(arg, found_key, found_value) repeatedly until
  // it returns false.
  Status Get(const ReadOptions& options,
             const InternalKeyComparator& internal_comparator,
-             const FileMetaData& file_meta, const Slice& k, void* arg,
+             const FileDescriptor& file_fd, const Slice& k, void* arg,
             bool (*handle_result)(void*, const ParsedInternalKey&,
-                                   const Slice&, bool),
-             bool* table_io, void (*mark_key_may_exist)(void*) = nullptr);
+                                   const Slice&),
+             void (*mark_key_may_exist)(void*) = nullptr);

  // Evict any entry for the specified file number
  static void Evict(Cache* cache, uint64_t file_number);
@@ -62,8 +62,8 @@ class TableCache {
  // Find table reader
  Status FindTable(const EnvOptions& toptions,
                   const InternalKeyComparator& internal_comparator,
-                   uint64_t file_number, uint64_t file_size, Cache::Handle**,
-                   bool* table_io = nullptr, const bool no_io = false);
+                   const FileDescriptor& file_fd, Cache::Handle**,
+                   const bool no_io = false);

  // Get TableReader from a cache handle.
  TableReader* GetTableReaderFromHandle(Cache::Handle* handle);
@@ -76,7 +76,7 @@ class TableCache {
  //            we set `no_io` to be true.
  Status GetTableProperties(const EnvOptions& toptions,
                            const InternalKeyComparator& internal_comparator,
-                            const FileMetaData& file_meta,
+                            const FileDescriptor& file_meta,
                            std::shared_ptr<const TableProperties>* properties,
                            bool no_io = false);

@@ -85,7 +85,7 @@ class TableCache {

 private:
  Env* const env_;
-  const std::string dbname_;
+  const std::vector<std::string> db_paths_;
  const Options* options_;
  const EnvOptions& storage_options_;
  Cache* const cache_;
@@ -18,25 +18,30 @@ namespace rocksdb {
 // Tag numbers for serialized VersionEdit.  These numbers are written to
 // disk and should not be changed.
 enum Tag {
-  kComparator           = 1,
-  kLogNumber            = 2,
-  kNextFileNumber       = 3,
-  kLastSequence         = 4,
-  kCompactPointer       = 5,
-  kDeletedFile          = 6,
-  kNewFile              = 7,
+  kComparator = 1,
+  kLogNumber = 2,
+  kNextFileNumber = 3,
+  kLastSequence = 4,
+  kCompactPointer = 5,
+  kDeletedFile = 6,
+  kNewFile = 7,
  // 8 was used for large value refs
-  kPrevLogNumber        = 9,
+  kPrevLogNumber = 9,

  // these are new formats divergent from open source leveldb
-  kNewFile2             = 100,  // store smallest & largest seqno
-
-  kColumnFamily         = 200,  // specify column family for version edit
-  kColumnFamilyAdd      = 201,
-  kColumnFamilyDrop     = 202,
-  kMaxColumnFamily      = 203,
+  kNewFile2 = 100,
+  kNewFile3 = 102,
+  kColumnFamily = 200,  // specify column family for version edit
+  kColumnFamilyAdd = 201,
+  kColumnFamilyDrop = 202,
+  kMaxColumnFamily = 203,
 };

+uint64_t PackFileNumberAndPathId(uint64_t number, uint64_t path_id) {
+  assert(number <= kFileNumberMask);
+  return number | (path_id * (kFileNumberMask + 1));
+}
+
 void VersionEdit::Clear() {
  comparator_.clear();
  max_level_ = 0;
@@ -93,10 +98,19 @@ void VersionEdit::EncodeTo(std::string* dst) const {

  for (size_t i = 0; i < new_files_.size(); i++) {
    const FileMetaData& f = new_files_[i].second;
-    PutVarint32(dst, kNewFile2);
+    if (f.fd.GetPathId() == 0) {
+      // Use older format to make sure user can roll back the build if they
+      // don't config multiple DB paths.
+      PutVarint32(dst, kNewFile2);
+    } else {
+      PutVarint32(dst, kNewFile3);
+    }
    PutVarint32(dst, new_files_[i].first);  // level
-    PutVarint64(dst, f.number);
-    PutVarint64(dst, f.file_size);
+    PutVarint64(dst, f.fd.GetNumber());
+    if (f.fd.GetPathId() != 0) {
+      PutVarint32(dst, f.fd.GetPathId());
+    }
+    PutVarint64(dst, f.fd.GetFileSize());
    PutLengthPrefixedSlice(dst, f.smallest.Encode());
    PutLengthPrefixedSlice(dst, f.largest.Encode());
    PutVarint64(dst, f.smallest_seqno);
@@ -230,12 +244,14 @@ Status VersionEdit::DecodeFrom(const Slice& src) {
        }
        break;

-      case kNewFile:
-        if (GetLevel(&input, &level, &msg) &&
-            GetVarint64(&input, &f.number) &&
-            GetVarint64(&input, &f.file_size) &&
+      case kNewFile: {
+        uint64_t number;
+        uint64_t file_size;
+        if (GetLevel(&input, &level, &msg) && GetVarint64(&input, &number) &&
+            GetVarint64(&input, &file_size) &&
            GetInternalKey(&input, &f.smallest) &&
            GetInternalKey(&input, &f.largest)) {
+          f.fd = FileDescriptor(number, 0, file_size);
          new_files_.push_back(std::make_pair(level, f));
        } else {
          if (!msg) {
@@ -243,15 +259,17 @@ Status VersionEdit::DecodeFrom(const Slice& src) {
          }
        }
        break;
-
-      case kNewFile2:
-        if (GetLevel(&input, &level, &msg) &&
-            GetVarint64(&input, &f.number) &&
-            GetVarint64(&input, &f.file_size) &&
+      }
+      case kNewFile2: {
+        uint64_t number;
+        uint64_t file_size;
+        if (GetLevel(&input, &level, &msg) && GetVarint64(&input, &number) &&
+            GetVarint64(&input, &file_size) &&
            GetInternalKey(&input, &f.smallest) &&
            GetInternalKey(&input, &f.largest) &&
            GetVarint64(&input, &f.smallest_seqno) &&
-            GetVarint64(&input, &f.largest_seqno) ) {
+            GetVarint64(&input, &f.largest_seqno)) {
+          f.fd = FileDescriptor(number, 0, file_size);
          new_files_.push_back(std::make_pair(level, f));
        } else {
          if (!msg) {
@@ -259,6 +277,27 @@ Status VersionEdit::DecodeFrom(const Slice& src) {
          }
        }
        break;
+      }
+
+      case kNewFile3: {
+        uint64_t number;
+        uint32_t path_id;
+        uint64_t file_size;
+        if (GetLevel(&input, &level, &msg) && GetVarint64(&input, &number) &&
+            GetVarint32(&input, &path_id) && GetVarint64(&input, &file_size) &&
+            GetInternalKey(&input, &f.smallest) &&
+            GetInternalKey(&input, &f.largest) &&
+            GetVarint64(&input, &f.smallest_seqno) &&
+            GetVarint64(&input, &f.largest_seqno)) {
+          f.fd = FileDescriptor(number, path_id, file_size);
+          new_files_.push_back(std::make_pair(level, f));
+        } else {
+          if (!msg) {
+            msg = "new-file2 entry";
+          }
+        }
+        break;
+      }

      case kColumnFamily:
        if (!GetVarint32(&input, &column_family_)) {
@@ -336,9 +375,9 @@ std::string VersionEdit::DebugString(bool hex_key) const {
    r.append("\n  AddFile: ");
    AppendNumberTo(&r, new_files_[i].first);
    r.append(" ");
-    AppendNumberTo(&r, f.number);
+    AppendNumberTo(&r, f.fd.GetNumber());
    r.append(" ");
-    AppendNumberTo(&r, f.file_size);
+    AppendNumberTo(&r, f.fd.GetFileSize());
    r.append(" ");
    r.append(f.smallest.DebugString(hex_key));
    r.append(" .. ");
@@ -19,31 +19,108 @@ namespace rocksdb {

 class VersionSet;

+const uint64_t kFileNumberMask = 0x3FFFFFFFFFFFFFFF;
+
+extern uint64_t PackFileNumberAndPathId(uint64_t number, uint64_t path_id);
+
+// A copyable structure contains information needed to read data from an SST
+// file. It can contains a pointer to a table reader opened for the file, or
+// file number and size, which can be used to create a new table reader for it.
+// The behavior is undefined when a copied of the structure is used when the
+// file is not in any live version any more.
+struct FileDescriptor {
+  // Table reader in table_reader_handle
+  TableReader* table_reader;
+  uint64_t packed_number_and_path_id;
+  uint64_t file_size;  // File size in bytes
+
+  FileDescriptor() : FileDescriptor(0, 0, 0) {}
+
+  FileDescriptor(uint64_t number, uint32_t path_id, uint64_t file_size)
+      : table_reader(nullptr),
+        packed_number_and_path_id(PackFileNumberAndPathId(number, path_id)),
+        file_size(file_size) {}
+
+  FileDescriptor& operator=(const FileDescriptor& fd) {
+    table_reader = fd.table_reader;
+    packed_number_and_path_id = fd.packed_number_and_path_id;
+    file_size = fd.file_size;
+    return *this;
+  }
+
+  uint64_t GetNumber() const {
+    return packed_number_and_path_id & kFileNumberMask;
+  }
+  uint32_t GetPathId() const {
+    return packed_number_and_path_id / (kFileNumberMask + 1);
+  }
+  uint64_t GetFileSize() const { return file_size; }
+};
+
 struct FileMetaData {
  int refs;
-  int allowed_seeks;          // Seeks allowed until compaction
-  uint64_t number;
-  uint64_t file_size;         // File size in bytes
-  InternalKey smallest;       // Smallest internal key served by table
-  InternalKey largest;        // Largest internal key served by table
-  bool being_compacted;       // Is this file undergoing compaction?
-  SequenceNumber smallest_seqno;// The smallest seqno in this file
-  SequenceNumber largest_seqno; // The largest seqno in this file
+  FileDescriptor fd;
+  InternalKey smallest;            // Smallest internal key served by table
+  InternalKey largest;             // Largest internal key served by table
+  bool being_compacted;            // Is this file undergoing compaction?
+  SequenceNumber smallest_seqno;   // The smallest seqno in this file
+  SequenceNumber largest_seqno;    // The largest seqno in this file

  // Needs to be disposed when refs becomes 0.
  Cache::Handle* table_reader_handle;
-  // Table reader in table_reader_handle
-  TableReader* table_reader;

-  FileMetaData(uint64_t number, uint64_t file_size)
+  // Stats for compensating deletion entries during compaction
+
+  // File size compensated by deletion entry.
+  // This is updated in Version::UpdateTemporaryStats() first time when the
+  // file is created or loaded.  After it is updated, it is immutable.
+  uint64_t compensated_file_size;
+  uint64_t num_entries;            // the number of entries.
+  uint64_t num_deletions;          // the number of deletion entries.
+  uint64_t raw_key_size;           // total uncompressed key size.
+  uint64_t raw_value_size;         // total uncompressed value size.
+
+  FileMetaData()
      : refs(0),
-        allowed_seeks(1 << 30),
-        number(number),
-        file_size(file_size),
        being_compacted(false),
        table_reader_handle(nullptr),
-        table_reader(nullptr) {}
-  FileMetaData() : FileMetaData(0, 0) {}
+        compensated_file_size(0),
+        num_entries(0),
+        num_deletions(0),
+        raw_key_size(0),
+        raw_value_size(0) {}
+};
+
+// A compressed copy of file meta data that just contain
+// smallest and largest key's slice
+struct FdWithKeyRange {
+  FileDescriptor fd;
+  Slice smallest_key;    // slice that contain smallest key
+  Slice largest_key;     // slice that contain largest key
+
+  FdWithKeyRange()
+      : fd(),
+        smallest_key(),
+        largest_key() {
+  }
+
+  FdWithKeyRange(FileDescriptor fd,
+      Slice smallest_key, Slice largest_key)
+      : fd(fd),
+        smallest_key(smallest_key),
+        largest_key(largest_key) {
+  }
+};
+
+// Data structure to store an array of FdWithKeyRange in one level
+// Actual data is guaranteed to be stored closely
+struct FileLevel {
+  size_t num_files;
+  FdWithKeyRange* files;
+  FileLevel() {
+    num_files = 0;
+    files = nullptr;
+  }
 };

 class VersionEdit {
@@ -81,16 +158,13 @@ class VersionEdit {
  // Add the specified file at the specified number.
  // REQUIRES: This version has not been saved (see VersionSet::SaveTo)
  // REQUIRES: "smallest" and "largest" are smallest and largest keys in file
-  void AddFile(int level, uint64_t file,
-               uint64_t file_size,
-               const InternalKey& smallest,
-               const InternalKey& largest,
-               const SequenceNumber& smallest_seqno,
+  void AddFile(int level, uint64_t file, uint64_t file_size,
+               uint64_t file_path_id, const InternalKey& smallest,
+               const InternalKey& largest, const SequenceNumber& smallest_seqno,
               const SequenceNumber& largest_seqno) {
    assert(smallest_seqno <= largest_seqno);
    FileMetaData f;
-    f.number = file;
-    f.file_size = file_size;
+    f.fd = FileDescriptor(file, file_size, file_path_id);
    f.smallest = smallest;
    f.largest = largest;
    f.smallest_seqno = smallest_seqno;
@@ -140,6 +214,7 @@ class VersionEdit {

 private:
  friend class VersionSet;
+  friend class Version;

  typedef std::set< std::pair<int, uint64_t>> DeletedFileSet;

@@ -30,11 +30,10 @@ TEST(VersionEditTest, EncodeDecode) {
  VersionEdit edit;
  for (int i = 0; i < 4; i++) {
    TestEncodeDecode(edit);
-    edit.AddFile(3, kBig + 300 + i, kBig + 400 + i,
+    edit.AddFile(3, kBig + 300 + i, kBig + 400 + i, 0,
                 InternalKey("foo", kBig + 500 + i, kTypeValue),
                 InternalKey("zoo", kBig + 600 + i, kTypeDeletion),
-                 kBig + 500 + i,
-                 kBig + 600 + i);
+                 kBig + 500 + i, kBig + 600 + i);
    edit.DeleteFile(4, kBig + 700 + i);
  }

@@ -51,6 +51,8 @@ class MergeContext;
 class ColumnFamilyData;
 class ColumnFamilySet;
 class TableCache;
+class MergeIteratorBuilder;
+

 // Return the smallest index i such that files[i]->largest >= key.
 // Return files.size() if there is no such file.
@@ -59,16 +61,24 @@ extern int FindFile(const InternalKeyComparator& icmp,
                    const std::vector<FileMetaData*>& files,
                    const Slice& key);

+// Return the smallest index i such that file_level.files[i]->largest >= key.
+// Return file_level.num_files if there is no such file.
+// REQUIRES: "file_level.files" contains a sorted list of
+// non-overlapping files.
+extern int FindFile(const InternalKeyComparator& icmp,
+                    const FileLevel& file_level,
+                    const Slice& key);
+
 // Returns true iff some file in "files" overlaps the user key range
 // [*smallest,*largest].
 // smallest==nullptr represents a key smaller than all keys in the DB.
 // largest==nullptr represents a key largest than all keys in the DB.
-// REQUIRES: If disjoint_sorted_files, files[] contains disjoint ranges
-//           in sorted order.
+// REQUIRES: If disjoint_sorted_files, file_level.files[]
+// contains disjoint ranges in sorted order.
 extern bool SomeFileOverlapsRange(
    const InternalKeyComparator& icmp,
    bool disjoint_sorted_files,
-    const std::vector<FileMetaData*>& files,
+    const FileLevel& file_level,
    const Slice* smallest_user_key,
    const Slice* largest_user_key);

@@ -80,29 +90,30 @@ class Version {
  void AddIterators(const ReadOptions&, const EnvOptions& soptions,
                    std::vector<Iterator*>* iters);

+  void AddIterators(const ReadOptions&, const EnvOptions& soptions,
+                    MergeIteratorBuilder* merger_iter_builder);
+
  // Lookup the value for key.  If found, store it in *val and
-  // return OK.  Else return a non-OK status.  Fills *stats.
+  // return OK.  Else return a non-OK status.
  // Uses *operands to store merge_operator operations to apply later
  // REQUIRES: lock is not held
-  struct GetStats {
-    FileMetaData* seek_file;
-    int seek_file_level;
-  };
  void Get(const ReadOptions&, const LookupKey& key, std::string* val,
-           Status* status, MergeContext* merge_context, GetStats* stats,
+           Status* status, MergeContext* merge_context,
           bool* value_found = nullptr);

-  // Adds "stats" into the current state.  Returns true if a new
-  // compaction may need to be triggered, false otherwise.
-  // REQUIRES: lock is held
-  bool UpdateStats(const GetStats& stats);
-
  // Updates internal structures that keep track of compaction scores
  // We use compaction scores to figure out which compaction to do next
  // REQUIRES: If Version is not yet saved to current_, it can be called without
  // a lock. Once a version is saved to current_, call only with mutex held
  void ComputeCompactionScore(std::vector<uint64_t>& size_being_compacted);

+  // Generate file_levels_ from files_
+  void GenerateFileLevels();
+
+  // Update scores, pre-calculated variables. It needs to be called before
+  // applying the version to the version set.
+  void PrepareApply(std::vector<uint64_t>& size_being_compacted);
+
  // Reference count management (so Versions do not disappear out from
  // under live iterators)
  void Ref();
@@ -189,7 +200,7 @@ class Version {
  int64_t MaxNextLevelOverlappingBytes();

  // Add all files listed in the current version to *live.
-  void AddLiveFiles(std::set<uint64_t>* live);
+  void AddLiveFiles(std::vector<FileDescriptor>* live);

  // Return a human readable string that describes this version's contents.
  std::string DebugString(bool hex = false) const;
@@ -197,6 +208,25 @@ class Version {
  // Returns the version nuber of this version
  uint64_t GetVersionNumber() const { return version_number_; }

+  uint64_t GetAverageValueSize() const {
+    if (num_non_deletions_ == 0) {
+      return 0;
+    }
+    assert(total_raw_key_size_ + total_raw_value_size_ > 0);
+    assert(total_file_size_ > 0);
+    return total_raw_value_size_ / num_non_deletions_ * total_file_size_ /
+           (total_raw_key_size_ + total_raw_value_size_);
+  }
+
+  // REQUIRES: lock is held
+  // On success, "tp" will contains the table properties of the file
+  // specified in "file_meta".  If the file name of "file_meta" is
+  // known ahread, passing it by a non-null "fname" can save a
+  // file-name conversion.
+  Status GetTableProperties(std::shared_ptr<const TableProperties>* tp,
+                            const FileMetaData* file_meta,
+                            const std::string* fname = nullptr);
+
  // REQUIRES: lock is held
  // On success, *props will be populated with all SSTables' table properties.
  // The keys of `props` are the sst file name, the values of `props` are the
@@ -218,6 +248,7 @@ class Version {
  friend class LevelCompactionPicker;
  friend class UniversalCompactionPicker;
  friend class FIFOCompactionPicker;
+  friend class ForwardIterator;

  class LevelFileNumIterator;
  class LevelFileIteratorState;
@@ -225,6 +256,18 @@ class Version {
  bool PrefixMayMatch(const ReadOptions& options, Iterator* level_iter,
                      const Slice& internal_prefix) const;

+  // Update num_non_empty_levels_.
+  void UpdateNumNonEmptyLevels();
+
+  // The helper function of UpdateTemporaryStats, which may fill the missing
+  // fields of file_mata from its associated TableProperties.
+  // Returns true if it does initialize FileMetaData.
+  bool MaybeInitializeFileMetaData(FileMetaData* file_meta);
+
+  // Update the temporary stats associated with the current version.
+  // This temporary stats will be used in compaction.
+  void UpdateTemporaryStats();
+
  // Sort all files for this version based on their file size and
  // record results in files_by_size_. The largest files are listed first.
  void UpdateFilesBySize();
@@ -234,18 +277,24 @@ class Version {
  const Comparator* user_comparator_;
  TableCache* table_cache_;
  const MergeOperator* merge_operator_;
+
+  autovector<FileLevel> file_levels_;   // A copy of list of files per level
  Logger* info_log_;
  Statistics* db_statistics_;
+  int num_levels_;              // Number of levels
+  int num_non_empty_levels_;    // Number of levels. Any level larger than it
+                                // is guaranteed to be empty.
  VersionSet* vset_;            // VersionSet to which this Version belongs
+  Arena arena_;                 // Used to allocate space for file_levels_
  Version* next_;               // Next version in linked list
  Version* prev_;               // Previous version in linked list
  int refs_;                    // Number of live refs to this version
-  int num_levels_;              // Number of levels

  // List of files per level, files in each level are arranged
  // in increasing order of keys
  std::vector<FileMetaData*>* files_;

+
  // A list for the same set of files that are stored in files_,
  // but files in each level are now sorted based on file
  // size. The file with the largest size is at the front.
@@ -261,11 +310,7 @@ class Version {
  // that on a running system, we need to look at only the first
  // few largest files because a new version is created every few
  // seconds/minutes (because of concurrent compactions).
-  static const int number_of_files_to_sort_ = 50;
-
-  // Next file to compact based on seek stats.
-  FileMetaData* file_to_compact_;
-  int file_to_compact_level_;
+  static const size_t number_of_files_to_sort_ = 50;

  // Level that should be compacted next and its compaction score.
  // Score < 1 means compaction is not strictly needed.  These fields
@@ -284,6 +329,16 @@ class Version {
  Version(ColumnFamilyData* cfd, VersionSet* vset, uint64_t version_number = 0);
  FileIndexer file_indexer_;

+  // total file size
+  uint64_t total_file_size_;
+  // the total size of all raw keys.
+  uint64_t total_raw_key_size_;
+  // the total size of all raw values.
+  uint64_t total_raw_value_size_;
+  // total number of non-deletion entries
+  uint64_t num_non_deletions_;
+
+
  ~Version();

  // re-initializes the index that is used to offset into files_by_size_
@@ -360,7 +415,7 @@ class VersionSet {
  // Arrange to reuse "file_number" unless a newer file number has
  // already been allocated.
  // REQUIRES: "file_number" was returned by a call to NewFileNumber().
-  void ReuseFileNumber(uint64_t file_number) {
+  void ReuseLogFileNumber(uint64_t file_number) {
    if (next_file_number_ == file_number + 1) {
      next_file_number_ = file_number;
    }
@@ -401,7 +456,7 @@ class VersionSet {
  Iterator* MakeInputIterator(Compaction* c);

  // Add all files listed in any live version to *live.
-  void AddLiveFiles(std::vector<uint64_t>* live_list);
+  void AddLiveFiles(std::vector<FileDescriptor>* live_list);

  // Return the approximate offset in the database of the data for
  // "key" as of version "v".
@@ -31,7 +31,7 @@ class FindFileTest {
           SequenceNumber smallest_seq = 100,
           SequenceNumber largest_seq = 100) {
    FileMetaData* f = new FileMetaData;
-    f->number = files_.size() + 1;
+    f->fd = FileDescriptor(files_.size() + 1, 0, 0);
    f->smallest = InternalKey(smallest, smallest_seq, kTypeValue);
    f->largest = InternalKey(largest, largest_seq, kTypeValue);
    files_.push_back(f);
@@ -42,18 +42,108 @@ class FindFileTest {
    InternalKeyComparator cmp(BytewiseComparator());
    return FindFile(cmp, files_, target.Encode());
  }
+};
+
+TEST(FindFileTest, Empty) {
+  ASSERT_EQ(0, Find("foo"));
+}
+
+TEST(FindFileTest, Single) {
+  Add("p", "q");
+  ASSERT_EQ(0, Find("a"));
+  ASSERT_EQ(0, Find("p"));
+  ASSERT_EQ(0, Find("p1"));
+  ASSERT_EQ(0, Find("q"));
+  ASSERT_EQ(1, Find("q1"));
+  ASSERT_EQ(1, Find("z"));
+}
+
+
+TEST(FindFileTest, Multiple) {
+  Add("150", "200");
+  Add("200", "250");
+  Add("300", "350");
+  Add("400", "450");
+  ASSERT_EQ(0, Find("100"));
+  ASSERT_EQ(0, Find("150"));
+  ASSERT_EQ(0, Find("151"));
+  ASSERT_EQ(0, Find("199"));
+  ASSERT_EQ(0, Find("200"));
+  ASSERT_EQ(1, Find("201"));
+  ASSERT_EQ(1, Find("249"));
+  ASSERT_EQ(1, Find("250"));
+  ASSERT_EQ(2, Find("251"));
+  ASSERT_EQ(2, Find("299"));
+  ASSERT_EQ(2, Find("300"));
+  ASSERT_EQ(2, Find("349"));
+  ASSERT_EQ(2, Find("350"));
+  ASSERT_EQ(3, Find("351"));
+  ASSERT_EQ(3, Find("400"));
+  ASSERT_EQ(3, Find("450"));
+  ASSERT_EQ(4, Find("451"));
+}
+
+class FindLevelFileTest {
+ public:
+  FileLevel level_files_;
+  bool disjoint_sorted_files_;
+  Arena arena_;
+
+  FindLevelFileTest() : disjoint_sorted_files_(true) { }
+
+  ~FindLevelFileTest() {
+  }
+
+  void LevelFileInit(size_t num = 0) {
+    char* mem = arena_.AllocateAligned(num * sizeof(FdWithKeyRange));
+    level_files_.files = new (mem)FdWithKeyRange[num];
+    level_files_.num_files = 0;
+  }
+
+  void Add(const char* smallest, const char* largest,
+           SequenceNumber smallest_seq = 100,
+           SequenceNumber largest_seq = 100) {
+    InternalKey smallest_key = InternalKey(smallest, smallest_seq, kTypeValue);
+    InternalKey largest_key = InternalKey(largest, largest_seq, kTypeValue);
+
+    Slice smallest_slice = smallest_key.Encode();
+    Slice largest_slice = largest_key.Encode();
+
+    char* mem = arena_.AllocateAligned(
+        smallest_slice.size() + largest_slice.size());
+    memcpy(mem, smallest_slice.data(), smallest_slice.size());
+    memcpy(mem + smallest_slice.size(), largest_slice.data(),
+        largest_slice.size());
+
+    // add compressd_level_
+    size_t num = level_files_.num_files;
+    auto& file = level_files_.files[num];
+    file.fd = FileDescriptor(num + 1, 0, 0);
+    file.smallest_key = Slice(mem, smallest_slice.size());
+    file.largest_key = Slice(mem + smallest_slice.size(),
+        largest_slice.size());
+    level_files_.num_files++;
+  }
+
+  int Find(const char* key) {
+    InternalKey target(key, 100, kTypeValue);
+    InternalKeyComparator cmp(BytewiseComparator());
+    return FindFile(cmp, level_files_, target.Encode());
+  }

  bool Overlaps(const char* smallest, const char* largest) {
    InternalKeyComparator cmp(BytewiseComparator());
    Slice s(smallest != nullptr ? smallest : "");
    Slice l(largest != nullptr ? largest : "");
-    return SomeFileOverlapsRange(cmp, disjoint_sorted_files_, files_,
+    return SomeFileOverlapsRange(cmp, disjoint_sorted_files_, level_files_,
                                 (smallest != nullptr ? &s : nullptr),
                                 (largest != nullptr ? &l : nullptr));
  }
 };

-TEST(FindFileTest, Empty) {
+TEST(FindLevelFileTest, LevelEmpty) {
+  LevelFileInit(0);
+
  ASSERT_EQ(0, Find("foo"));
  ASSERT_TRUE(! Overlaps("a", "z"));
  ASSERT_TRUE(! Overlaps(nullptr, "z"));
@@ -61,7 +151,9 @@ TEST(FindFileTest, Empty) {
  ASSERT_TRUE(! Overlaps(nullptr, nullptr));
 }

-TEST(FindFileTest, Single) {
+TEST(FindLevelFileTest, LevelSingle) {
+  LevelFileInit(1);
+
  Add("p", "q");
  ASSERT_EQ(0, Find("a"));
  ASSERT_EQ(0, Find("p"));
@@ -91,8 +183,9 @@ TEST(FindFileTest, Single) {
  ASSERT_TRUE(Overlaps(nullptr, nullptr));
 }

+TEST(FindLevelFileTest, LevelMultiple) {
+  LevelFileInit(4);

-TEST(FindFileTest, Multiple) {
  Add("150", "200");
  Add("200", "250");
  Add("300", "350");
@@ -130,7 +223,9 @@ TEST(FindFileTest, Multiple) {
  ASSERT_TRUE(Overlaps("450", "500"));
 }

-TEST(FindFileTest, MultipleNullBoundaries) {
+TEST(FindLevelFileTest, LevelMultipleNullBoundaries) {
+  LevelFileInit(4);
+
  Add("150", "200");
  Add("200", "250");
  Add("300", "350");
@@ -150,7 +245,9 @@ TEST(FindFileTest, MultipleNullBoundaries) {
  ASSERT_TRUE(Overlaps("450", nullptr));
 }

-TEST(FindFileTest, OverlapSequenceChecks) {
+TEST(FindLevelFileTest, LevelOverlapSequenceChecks) {
+  LevelFileInit(1);
+
  Add("200", "200", 5000, 3000);
  ASSERT_TRUE(! Overlaps("199", "199"));
  ASSERT_TRUE(! Overlaps("201", "300"));
@@ -159,7 +256,9 @@ TEST(FindFileTest, OverlapSequenceChecks) {
  ASSERT_TRUE(Overlaps("200", "210"));
 }

-TEST(FindFileTest, OverlappingFiles) {
+TEST(FindLevelFileTest, LevelOverlappingFiles) {
+  LevelFileInit(2);
+
  Add("150", "600");
  Add("400", "500");
  disjoint_sorted_files_ = false;
@@ -236,6 +236,23 @@ void WriteBatch::Delete(ColumnFamilyHandle* column_family, const Slice& key) {
  WriteBatchInternal::Delete(this, GetColumnFamilyID(column_family), key);
 }

+void WriteBatchInternal::Delete(WriteBatch* b, uint32_t column_family_id,
+                                const SliceParts& key) {
+  WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
+  if (column_family_id == 0) {
+    b->rep_.push_back(static_cast<char>(kTypeDeletion));
+  } else {
+    b->rep_.push_back(static_cast<char>(kTypeColumnFamilyDeletion));
+    PutVarint32(&b->rep_, column_family_id);
+  }
+  PutLengthPrefixedSliceParts(&b->rep_, key);
+}
+
+void WriteBatch::Delete(ColumnFamilyHandle* column_family,
+                        const SliceParts& key) {
+  WriteBatchInternal::Delete(this, GetColumnFamilyID(column_family), key);
+}
+
 void WriteBatchInternal::Merge(WriteBatch* b, uint32_t column_family_id,
                               const Slice& key, const Slice& value) {
  WriteBatchInternal::SetCount(b, WriteBatchInternal::Count(b) + 1);
@@ -71,6 +71,9 @@ class WriteBatchInternal {
  static void Put(WriteBatch* batch, uint32_t column_family_id,
                  const SliceParts& key, const SliceParts& value);

+  static void Delete(WriteBatch* batch, uint32_t column_family_id,
+                     const SliceParts& key);
+
  static void Delete(WriteBatch* batch, uint32_t column_family_id,
                     const Slice& key);

@@ -642,10 +642,6 @@ Default:1, i.e. pick maxfilesize amount of data as the source of a compaction.
 <li> <code>Options::max_grandparent_overlap_factor</code> -   Control maximum bytes of overlaps in grandparent (i.e., level+2) before we
 stop building a single file in a level->level+1 compaction.
 <p>
-<li> <code>Options::disable_seek_compaction</code> -  Disable compaction triggered by seek.
-With bloomfilter and fast storage, a miss on one level is very cheap if the file handle is cached in table cache
-(which is true if max_open_files is large).
-<p>
 <li> <code>Options::max_background_compactions</code> - Maximum number of concurrent background jobs, submitted to
 the default LOW priority thread pool
 </ul>
@@ -56,8 +56,14 @@ extern "C" {

 typedef struct rocksdb_t                 rocksdb_t;
 typedef struct rocksdb_cache_t           rocksdb_cache_t;
+typedef struct rocksdb_compactionfilter_t rocksdb_compactionfilter_t;
+typedef struct rocksdb_compactionfiltercontext_t
+    rocksdb_compactionfiltercontext_t;
+typedef struct rocksdb_compactionfilterfactory_t
+    rocksdb_compactionfilterfactory_t;
 typedef struct rocksdb_comparator_t      rocksdb_comparator_t;
 typedef struct rocksdb_env_t             rocksdb_env_t;
+typedef struct rocksdb_fifo_compaction_options_t rocksdb_fifo_compaction_options_t;
 typedef struct rocksdb_filelock_t        rocksdb_filelock_t;
 typedef struct rocksdb_filterpolicy_t    rocksdb_filterpolicy_t;
 typedef struct rocksdb_flushoptions_t    rocksdb_flushoptions_t;
@@ -75,6 +81,7 @@ typedef struct rocksdb_writebatch_t      rocksdb_writebatch_t;
 typedef struct rocksdb_writeoptions_t    rocksdb_writeoptions_t;
 typedef struct rocksdb_universal_compaction_options_t rocksdb_universal_compaction_options_t;
 typedef struct rocksdb_livefiles_t     rocksdb_livefiles_t;
+typedef struct rocksdb_column_family_handle_t rocksdb_column_family_handle_t;

 /* DB operations */

@@ -89,6 +96,45 @@ extern rocksdb_t* rocksdb_open_for_read_only(
    unsigned char error_if_log_file_exist,
    char** errptr);

+extern rocksdb_t* rocksdb_open_column_families(
+    const rocksdb_options_t* options,
+    const char* name,
+    int num_column_families,
+    const char** column_family_names,
+    const rocksdb_options_t** column_family_options,
+    rocksdb_column_family_handle_t** column_family_handles,
+    char** errptr);
+
+extern rocksdb_t* rocksdb_open_for_read_only_column_families(
+    const rocksdb_options_t* options,
+    const char* name,
+    int num_column_families,
+    const char** column_family_names,
+    const rocksdb_options_t** column_family_options,
+    rocksdb_column_family_handle_t** column_family_handles,
+    unsigned char error_if_log_file_exist,
+    char** errptr);
+
+char** rocksdb_list_column_families(
+    const rocksdb_options_t* options,
+    const char* name,
+    size_t* lencf,
+    char** errptr);
+void rocksdb_list_column_families_destroy(char** list, size_t len);
+
+extern rocksdb_column_family_handle_t* rocksdb_create_column_family(
+    rocksdb_t* db,
+    const rocksdb_options_t* column_family_options,
+    const char* column_family_name,
+    char** errptr);
+
+extern void rocksdb_drop_column_family(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* handle,
+    char** errptr);
+
+extern void rocksdb_column_family_handle_destroy(rocksdb_column_family_handle_t*);
+
 extern void rocksdb_close(rocksdb_t* db);

 extern void rocksdb_put(
@@ -98,12 +144,27 @@ extern void rocksdb_put(
    const char* val, size_t vallen,
    char** errptr);

+extern void rocksdb_put_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    const char* val, size_t vallen,
+    char** errptr);
+
 extern void rocksdb_delete(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
    const char* key, size_t keylen,
    char** errptr);

+void rocksdb_delete_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    char** errptr);
+
 extern void rocksdb_merge(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -111,6 +172,14 @@ extern void rocksdb_merge(
    const char* val, size_t vallen,
    char** errptr);

+extern void rocksdb_merge_cf(
+    rocksdb_t* db,
+    const rocksdb_writeoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    const char* val, size_t vallen,
+    char** errptr);
+
 extern void rocksdb_write(
    rocksdb_t* db,
    const rocksdb_writeoptions_t* options,
@@ -126,10 +195,23 @@ extern char* rocksdb_get(
    size_t* vallen,
    char** errptr);

+extern char* rocksdb_get_cf(
+    rocksdb_t* db,
+    const rocksdb_readoptions_t* options,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t keylen,
+    size_t* vallen,
+    char** errptr);
+
 extern rocksdb_iterator_t* rocksdb_create_iterator(
    rocksdb_t* db,
    const rocksdb_readoptions_t* options);

+extern rocksdb_iterator_t* rocksdb_create_iterator_cf(
+    rocksdb_t* db,
+    const rocksdb_readoptions_t* options,
+    rocksdb_column_family_handle_t* column_family);
+
 extern const rocksdb_snapshot_t* rocksdb_create_snapshot(
    rocksdb_t* db);

@@ -143,6 +225,11 @@ extern char* rocksdb_property_value(
    rocksdb_t* db,
    const char* propname);

+extern char* rocksdb_property_value_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    const char* propname);
+
 extern void rocksdb_approximate_sizes(
    rocksdb_t* db,
    int num_ranges,
@@ -150,11 +237,25 @@ extern void rocksdb_approximate_sizes(
    const char* const* range_limit_key, const size_t* range_limit_key_len,
    uint64_t* sizes);

+extern void rocksdb_approximate_sizes_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    int num_ranges,
+    const char* const* range_start_key, const size_t* range_start_key_len,
+    const char* const* range_limit_key, const size_t* range_limit_key_len,
+    uint64_t* sizes);
+
 extern void rocksdb_compact_range(
    rocksdb_t* db,
    const char* start_key, size_t start_key_len,
    const char* limit_key, size_t limit_key_len);

+extern void rocksdb_compact_range_cf(
+    rocksdb_t* db,
+    rocksdb_column_family_handle_t* column_family,
+    const char* start_key, size_t start_key_len,
+    const char* limit_key, size_t limit_key_len);
+
 extern void rocksdb_delete_file(
    rocksdb_t* db,
    const char* name);
@@ -204,6 +305,8 @@ extern void rocksdb_iter_get_error(const rocksdb_iterator_t*, char** errptr);
 /* Write batch */

 extern rocksdb_writebatch_t* rocksdb_writebatch_create();
+extern rocksdb_writebatch_t* rocksdb_writebatch_create_from(const char* rep,
+                                                            size_t size);
 extern void rocksdb_writebatch_destroy(rocksdb_writebatch_t*);
 extern void rocksdb_writebatch_clear(rocksdb_writebatch_t*);
 extern int rocksdb_writebatch_count(rocksdb_writebatch_t*);
@@ -211,13 +314,27 @@ extern void rocksdb_writebatch_put(
    rocksdb_writebatch_t*,
    const char* key, size_t klen,
    const char* val, size_t vlen);
+extern void rocksdb_writebatch_put_cf(
+    rocksdb_writebatch_t*,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen,
+    const char* val, size_t vlen);
 extern void rocksdb_writebatch_merge(
    rocksdb_writebatch_t*,
    const char* key, size_t klen,
    const char* val, size_t vlen);
+extern void rocksdb_writebatch_merge_cf(
+    rocksdb_writebatch_t*,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen,
+    const char* val, size_t vlen);
 extern void rocksdb_writebatch_delete(
    rocksdb_writebatch_t*,
    const char* key, size_t klen);
+extern void rocksdb_writebatch_delete_cf(
+    rocksdb_writebatch_t*,
+    rocksdb_column_family_handle_t* column_family,
+    const char* key, size_t klen);
 extern void rocksdb_writebatch_iterate(
    rocksdb_writebatch_t*,
    void* state,
@@ -229,6 +346,19 @@ extern const char* rocksdb_writebatch_data(rocksdb_writebatch_t*, size_t *size);

 extern rocksdb_options_t* rocksdb_options_create();
 extern void rocksdb_options_destroy(rocksdb_options_t*);
+extern void rocksdb_options_increase_parallelism(
+    rocksdb_options_t* opt, int total_threads);
+extern void rocksdb_options_optimize_for_point_lookup(
+    rocksdb_options_t* opt);
+extern void rocksdb_options_optimize_level_style_compaction(
+    rocksdb_options_t* opt, uint64_t memtable_memory_budget);
+extern void rocksdb_options_optimize_universal_style_compaction(
+    rocksdb_options_t* opt, uint64_t memtable_memory_budget);
+extern void rocksdb_options_set_compaction_filter(
+    rocksdb_options_t*,
+    rocksdb_compactionfilter_t*);
+extern void rocksdb_options_set_compaction_filter_factory(
+    rocksdb_options_t*, rocksdb_compactionfilterfactory_t*);
 extern void rocksdb_options_set_comparator(
    rocksdb_options_t*,
    rocksdb_comparator_t*);
@@ -243,6 +373,8 @@ extern void rocksdb_options_set_filter_policy(
    rocksdb_filterpolicy_t*);
 extern void rocksdb_options_set_create_if_missing(
    rocksdb_options_t*, unsigned char);
+extern void rocksdb_options_set_create_missing_column_families(
+    rocksdb_options_t*, unsigned char);
 extern void rocksdb_options_set_error_if_exists(
    rocksdb_options_t*, unsigned char);
 extern void rocksdb_options_set_paranoid_checks(
@@ -397,10 +529,48 @@ extern void rocksdb_options_set_compression(rocksdb_options_t*, int);

 enum {
  rocksdb_level_compaction = 0,
-  rocksdb_universal_compaction = 1
+  rocksdb_universal_compaction = 1,
+  rocksdb_fifo_compaction = 2
 };
 extern void rocksdb_options_set_compaction_style(rocksdb_options_t*, int);
 extern void rocksdb_options_set_universal_compaction_options(rocksdb_options_t*, rocksdb_universal_compaction_options_t*);
+extern void rocksdb_options_set_fifo_compaction_options(rocksdb_options_t* opt,
+    rocksdb_fifo_compaction_options_t* fifo);
+
+/* Compaction Filter */
+
+extern rocksdb_compactionfilter_t* rocksdb_compactionfilter_create(
+    void* state,
+    void (*destructor)(void*),
+    unsigned char (*filter)(
+        void*,
+        int level,
+        const char* key, size_t key_length,
+        const char* existing_value, size_t value_length,
+        char** new_value, size_t *new_value_length,
+        unsigned char* value_changed),
+    const char* (*name)(void*));
+extern void rocksdb_compactionfilter_destroy(rocksdb_compactionfilter_t*);
+
+/* Compaction Filter Context */
+
+extern unsigned char rocksdb_compactionfiltercontext_is_full_compaction(
+    rocksdb_compactionfiltercontext_t* context);
+
+extern unsigned char rocksdb_compactionfiltercontext_is_manual_compaction(
+    rocksdb_compactionfiltercontext_t* context);
+
+/* Compaction Filter Factory */
+
+extern rocksdb_compactionfilterfactory_t*
+    rocksdb_compactionfilterfactory_create(
+        void* state, void (*destructor)(void*),
+        rocksdb_compactionfilter_t* (*create_compaction_filter)(
+            void*, rocksdb_compactionfiltercontext_t* context),
+        const char* (*name)(void*));
+extern void rocksdb_compactionfilterfactory_destroy(
+    rocksdb_compactionfilterfactory_t*);
+
 /* Comparator */

 extern rocksdb_comparator_t* rocksdb_comparator_create(
@@ -546,6 +716,12 @@ extern void rocksdb_universal_compaction_options_set_stop_style(
 extern void rocksdb_universal_compaction_options_destroy(
  rocksdb_universal_compaction_options_t*);

+extern rocksdb_fifo_compaction_options_t* rocksdb_fifo_compaction_options_create();
+extern void rocksdb_fifo_compaction_options_set_max_table_files_size(
+    rocksdb_fifo_compaction_options_t* fifo_opts, uint64_t size);
+extern void rocksdb_fifo_compaction_options_destroy(
+    rocksdb_fifo_compaction_options_t* fifo_opts);
+
 extern int rocksdb_livefiles_count(
  const rocksdb_livefiles_t*);
 extern const char* rocksdb_livefiles_name(
@@ -86,7 +86,7 @@ class CompactionFilterV2 {
  //
  // Each entry in the return vector indicates if the corresponding kv should
  // be preserved in the output of this compaction run. The application can
-  // inspect the exisitng values of the keys and make decision based on it.
+  // inspect the existing values of the keys and make decision based on it.
  //
  // When a value is to be preserved, the application has the option
  // to modify the entry in existing_values and pass it back through an entry
@@ -108,7 +108,7 @@ class CompactionFilterV2 {
 };

 // Each compaction will create a new CompactionFilter allowing the
-// application to know about different campactions
+// application to know about different compactions
 class CompactionFilterFactory {
 public:
  virtual ~CompactionFilterFactory() { }
@@ -120,7 +120,7 @@ class CompactionFilterFactory {
  virtual const char* Name() const = 0;
 };

-// Default implementaion of CompactionFilterFactory which does not
+// Default implementation of CompactionFilterFactory which does not
 // return any filter
 class DefaultCompactionFilterFactory : public CompactionFilterFactory {
 public:
@@ -175,7 +175,7 @@ class CompactionFilterFactoryV2 {
  const SliceTransform* prefix_extractor_;
 };

-// Default implementaion of CompactionFilterFactoryV2 which does not
+// Default implementation of CompactionFilterFactoryV2 which does not
 // return any filter
 class DefaultCompactionFilterFactoryV2 : public CompactionFilterFactoryV2 {
 public:
@@ -55,6 +55,7 @@ class Env;
 // Metadata associated with each SST file.
 struct LiveFileMetaData {
  std::string column_family_name;  // Name of the column family
+  std::string db_path;
  std::string name;                // Name of the file
  int level;               // Level at which this file resides.
  size_t size;             // File size in bytes.
@@ -35,6 +35,7 @@ class WritableFile;
 class RandomRWFile;
 class Directory;
 struct DBOptions;
+class RateLimiter;

 using std::unique_ptr;
 using std::shared_ptr;
@@ -74,6 +75,9 @@ struct EnvOptions {
  // write. By default, we set it to true for MANIFEST writes and false for
  // WAL writes
  bool fallocate_with_keep_size = true;
+
+  // If not nullptr, write rate limiting is enabled for flush and compaction
+  RateLimiter* rate_limiter = nullptr;
 };

 class Env {
@@ -194,8 +198,16 @@ class Env {
  // REQUIRES: lock has not already been unlocked.
  virtual Status UnlockFile(FileLock* lock) = 0;

+  // Priority for scheduling job in thread pool
  enum Priority { LOW, HIGH, TOTAL };

+  // Priority for scheduling job in thread pool
+  enum IOPriority {
+    IO_LOW = 0,
+    IO_HIGH = 1,
+    IO_TOTAL = 2
+  };
+
  // Arrange to run "(*function)(arg)" once in a background thread, in
  // the thread pool specified by pri. By default, jobs go to the 'LOW'
  // priority thread pool.
@@ -371,7 +383,10 @@ class RandomAccessFile {
 // at a time to the file.
 class WritableFile {
 public:
-  WritableFile() : last_preallocated_block_(0), preallocation_block_size_ (0) {
+  WritableFile()
+    : last_preallocated_block_(0),
+      preallocation_block_size_(0),
+      io_priority_(Env::IO_TOTAL) {
  }
  virtual ~WritableFile();

@@ -390,6 +405,14 @@ class WritableFile {
    return Sync();
  }

+  /*
+   * Change the priority in rate limiter if rate limiting is enabled.
+   * If rate limiting is not enabled, this call has no effect.
+   */
+  virtual void SetIOPriority(Env::IOPriority pri) {
+    io_priority_ = pri;
+  }
+
  /*
   * Get the size of valid data in the file.
   */
@@ -474,6 +497,9 @@ class WritableFile {
  // No copying allowed
  WritableFile(const WritableFile&);
  void operator=(const WritableFile&);
+
+ protected:
+  Env::IOPriority io_priority_;
 };

 // A file abstraction for random reading and writing.
@@ -0,0 +1,34 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+#ifndef INCLUDE_ROCKSDB_IOSTATS_CONTEXT_H_
+#define INCLUDE_ROCKSDB_IOSTATS_CONTEXT_H_
+
+#include <stdint.h>
+#include <string>
+
+// A thread local context for gathering io-stats efficiently and transparently.
+namespace rocksdb {
+
+struct IOStatsContext {
+  // reset all io-stats counter to zero
+  void Reset();
+
+  std::string ToString() const;
+
+  // the thread pool id
+  uint64_t thread_pool_id;
+
+  // number of bytes that has been written.
+  uint64_t bytes_written;
+  // number of bytes that has been read.
+  uint64_t bytes_read;
+};
+
+extern __thread IOStatsContext iostats_context;
+
+}  // namespace rocksdb
+
+#endif  // INCLUDE_ROCKSDB_IOSTATS_CONTEXT_H_
@@ -4,13 +4,32 @@
 // of patent rights can be found in the PATENTS file in the same directory.
 #ifndef ROCKSDB_LITE
 #pragma once
+#include <string>
 #include "rocksdb/options.h"

 namespace rocksdb {

+// An interface for converting a slice to a readable string
+class SliceFormatter {
+ public:
+  virtual ~SliceFormatter() {}
+  virtual std::string Format(const Slice& s) const = 0;
+};
+
+// Options for customizing ldb tool (beyond the DB Options)
+struct LDBOptions {
+  // Create LDBOptions with default values for all fields
+  LDBOptions();
+
+  // Key formatter that converts a slice to a readable string.
+  // Default: Slice::ToString()
+  std::shared_ptr<SliceFormatter> key_formatter;
+};
+
 class LDBTool {
 public:
-  void Run(int argc, char** argv, Options = Options());
+  void Run(int argc, char** argv, Options db_options= Options(),
+           const LDBOptions& ldb_options = LDBOptions());
 };

 } // namespace rocksdb
@@ -142,16 +142,21 @@ class MemTableRep {
  };

  // Return an iterator over the keys in this representation.
-  virtual Iterator* GetIterator() = 0;
-
-  // Return an iterator over at least the keys with the specified user key. The
-  // iterator may also allow access to other keys, but doesn't have to. Default:
-  // GetIterator().
-  virtual Iterator* GetIterator(const Slice& user_key) { return GetIterator(); }
+  // arena: If not null, the arena needs to be used to allocate the Iterator.
+  //        When destroying the iterator, the caller will not call "delete"
+  //        but Iterator::~Iterator() directly. The destructor needs to destroy
+  //        all the states but those allocated in arena.
+  virtual Iterator* GetIterator(Arena* arena = nullptr) = 0;

  // Return an iterator that has a special Seek semantics. The result of
  // a Seek might only include keys with the same prefix as the target key.
-  virtual Iterator* GetDynamicPrefixIterator() { return GetIterator(); }
+  // arena: If not null, the arena needs to be used to allocate the Iterator.
+  //        When destroying the iterator, the caller will not call "delete"
+  //        but Iterator::~Iterator() directly. The destructor needs to destroy
+  //        all the states but those allocated in arena.
+  virtual Iterator* GetDynamicPrefixIterator(Arena* arena = nullptr) {
+    return GetIterator(arena);
+  }

  // Return true if the current MemTableRep supports merge operator.
  // Default: true
@@ -222,17 +227,27 @@ extern MemTableRepFactory* NewHashSkipListRepFactory(
    int32_t skiplist_branching_factor = 4
 );

-// The factory is to create memtables with a hashed linked list:
-// it contains a fixed array of buckets, each pointing to a sorted single
-// linked list (null if the bucket is empty).
+// The factory is to create memtables based on a hash table:
+// it contains a fixed array of buckets, each pointing to either a linked list
+// or a skip list if number of entries inside the bucket exceeds
+// threshold_use_skiplist.
 // @bucket_count: number of fixed array buckets
 // @huge_page_tlb_size: if <=0, allocate the hash table bytes from malloc.
 //                      Otherwise from huge page TLB. The user needs to reserve
 //                      huge pages for it to be allocated, like:
 //                          sysctl -w vm.nr_hugepages=20
 //                      See linux doc Documentation/vm/hugetlbpage.txt
+// @bucket_entries_logging_threshold: if number of entries in one bucket
+//                                    exceeds this number, log about it.
+// @if_log_bucket_dist_when_flash: if true, log distribution of number of
+//                                 entries when flushing.
+// @threshold_use_skiplist: a bucket switches to skip list if number of
+//                          entries exceed this parameter.
 extern MemTableRepFactory* NewHashLinkListRepFactory(
-    size_t bucket_count = 50000, size_t huge_page_tlb_size = 0);
+    size_t bucket_count = 50000, size_t huge_page_tlb_size = 0,
+    int bucket_entries_logging_threshold = 4096,
+    bool if_log_bucket_dist_when_flash = true,
+    uint32_t threshold_use_skiplist = 256);

 // This factory creates a cuckoo-hashing based mem-table representation.
 // Cuckoo-hash is a closed-hash strategy, in which all key/value pairs
@@ -34,12 +34,12 @@ class Snapshot;
 class TableFactory;
 class MemTableRepFactory;
 class TablePropertiesCollectorFactory;
+class RateLimiter;
 class Slice;
 class SliceTransform;
 class Statistics;
 class InternalKeyComparator;
-
-using std::shared_ptr;
+class RateLimiter;

 // DB contents are stored in a set of blocks, each of which holds a
 // sequence of key,value pairs.  Each block may be compressed before
@@ -132,7 +132,7 @@ struct ColumnFamilyOptions {
  // for the first time. It's necessary to specify a merge operator when
  // openning the DB in this case.
  // Default: nullptr
-  shared_ptr<MergeOperator> merge_operator;
+  std::shared_ptr<MergeOperator> merge_operator;

  // A single CompactionFilter instance to call into during compaction.
  // Allows an application to modify/delete a key-value during background
@@ -184,8 +184,9 @@ struct ColumnFamilyOptions {
  size_t write_buffer_size;

  // The maximum number of write buffers that are built up in memory.
-  // The default is 2, so that when 1 write buffer is being flushed to
-  // storage, new writes can continue to the other write buffer.
+  // The default and the minimum number is 2, so that when 1 write buffer
+  // is being flushed to storage, new writes can continue to the other
+  // write buffer.
  // Default: 2
  int max_write_buffer_number;

@@ -204,12 +205,12 @@ struct ColumnFamilyOptions {
  // If non-NULL use the specified cache for blocks.
  // If NULL, rocksdb will automatically create and use an 8MB internal cache.
  // Default: nullptr
-  shared_ptr<Cache> block_cache;
+  std::shared_ptr<Cache> block_cache;

  // If non-NULL use the specified cache for compressed blocks.
  // If NULL, rocksdb will not use a compressed block cache.
  // Default: nullptr
-  shared_ptr<Cache> block_cache_compressed;
+  std::shared_ptr<Cache> block_cache_compressed;

  // Approximate size of user data packed per block.  Note that the
  // block size specified here corresponds to uncompressed data.  The
@@ -367,10 +368,15 @@ struct ColumnFamilyOptions {
  // stop building a single file in a level->level+1 compaction.
  int max_grandparent_overlap_factor;

-  // Disable compaction triggered by seek.
-  // With bloomfilter and fast storage, a miss on one level
-  // is very cheap if the file handle is cached in table cache
-  // (which is true if max_open_files is large).
+  // We decided to remove seek compaction from RocksDB because:
+  // 1) It makes more sense for spinning disk workloads, while RocksDB is
+  // primarily designed for flash and memory,
+  // 2) It added some complexity to the important code-paths,
+  // 3) None of our internal customers were really using it.
+  //
+  // Since we removed seek compaction, this option is now obsolete.
+  // We left it here for backwards compatiblity (otherwise it would break the
+  // build), but we'll remove it at some point.
  // Default: true
  bool disable_seek_compaction;

@@ -547,12 +553,9 @@ struct ColumnFamilyOptions {

  // Control locality of bloom filter probes to improve cache miss rate.
  // This option only applies to memtable prefix bloom and plaintable
-  // prefix bloom. It essentially limits the max number of cache lines each
-  // bloom filter check can touch.
-  // This optimization is turned off when set to 0. The number should never
-  // be greater than number of probes. This option can boost performance
-  // for in-memory workload but should use with care since it can cause
-  // higher false positive rate.
+  // prefix bloom. It essentially limits every bloom checking to one cache line.
+  // This optimization is turned off when set to 0, and positive number to turn
+  // it on.
  // Default: 0
  uint32_t bloom_locality;

@@ -598,6 +601,10 @@ struct DBOptions {
  // Default: false
  bool create_if_missing;

+  // If true, missing column families will be automatically created.
+  // Default: false
+  bool create_missing_column_families;
+
  // If true, an error is raised if the database already exists.
  // Default: false
  bool error_if_exists;
@@ -618,11 +625,16 @@ struct DBOptions {
  // Default: Env::Default()
  Env* env;

+  // Use to control write rate of flush and compaction. Flush has higher
+  // priority than compaction. Rate limiting is disabled if nullptr.
+  // Default: nullptr
+  std::shared_ptr<RateLimiter> rate_limiter;
+
  // Any internal progress/error information generated by the db will
  // be written to info_log if it is non-nullptr, or to a file stored
  // in the same directory as the DB contents if info_log is nullptr.
  // Default: nullptr
-  shared_ptr<Logger> info_log;
+  std::shared_ptr<Logger> info_log;

  InfoLogLevel info_log_level;

@@ -645,7 +657,7 @@ struct DBOptions {
  // If non-null, then we should collect metrics about database operations
  // Statistics objects should not be shared between DB instances as
  // it does not use any locks to prevent concurrent updates.
-  shared_ptr<Statistics> statistics;
+  std::shared_ptr<Statistics> statistics;

  // If true, then the contents of data files are not synced
  // to stable storage. Their contents remain in the OS buffers till the
@@ -668,6 +680,13 @@ struct DBOptions {
  // Default value is 1800 (half an hour).
  int db_stats_log_interval;

+  // A list paths where SST files can be put into. A compaction style can
+  // determine which of those paths it will put the file to.
+  // If left empty, only one path will be used, which is db_name passed when
+  // opening the DB.
+  // Default: empty
+  std::vector<std::string> db_paths;
+
  // This specifies the info LOG dir.
  // If it is empty, the log files will be in the same dir as data.
  // If it is non empty, the log files will be in the specified dir,
@@ -961,7 +980,18 @@ struct WriteOptions {
  // and the write may got lost after a crash.
  bool disableWAL;

-  WriteOptions() : sync(false), disableWAL(false) {}
+  // If non-zero, then associated write waiting longer than the specified
+  // time MAY be aborted and returns Status::TimedOut. A write that takes
+  // less than the specified time is guaranteed to not fail with
+  // Status::TimedOut.
+  //
+  // The number of times a write call encounters a timeout is recorded in
+  // Statistics.WRITE_TIMEDOUT
+  //
+  // Default: 0
+  uint64_t timeout_hint_us;
+
+  WriteOptions() : sync(false), disableWAL(false), timeout_hint_us(0) {}
 };

 // Options that control flush operations
@@ -973,6 +1003,21 @@ struct FlushOptions {
  FlushOptions() : wait(true) {}
 };

+
+// Create a RateLimiter object, which can be shared among RocksDB instances to
+// control write rate of flush and compaction.
+// @rate_bytes_per_sec: desired total write rate in bytes per second.
+// @refill_period_us: token refill interval in micro-second.
+// @fairness: RateLimiter accepts high-pri requests and low-pri requests.
+// low-pri request is usually blocked in favor of hi-pri request. To prevent
+// low-pri request from being blocked for too long, it can get processed first
+// by 1/fairness chance.
+extern RateLimiter* NewRateLimiter(
+    int64_t rate_bytes_per_sec,
+    int64_t refill_period_us = 100 * 1000,
+    int32_t fairness = 10);
+
+
 }  // namespace rocksdb

 #endif  // STORAGE_ROCKSDB_INCLUDE_OPTIONS_H_
@@ -20,6 +20,9 @@ enum PerfLevel {
 // set the perf stats level
 void SetPerfLevel(PerfLevel level);

+// get current perf stats level
+PerfLevel GetPerfLevel();
+
 // A thread local context for gathering performance counter efficiently
 // and transparently.

@@ -107,6 +107,7 @@ class Slice {
 struct SliceParts {
  SliceParts(const Slice* _parts, int _num_parts) :
      parts(_parts), num_parts(_num_parts) { }
+  SliceParts() : parts(nullptr), num_parts(0) {}

  const Slice* parts;
  int num_parts;
@@ -115,9 +115,11 @@ enum Tickers {
  // head of the writers queue.
  WRITE_DONE_BY_SELF,
  WRITE_DONE_BY_OTHER,
+  WRITE_TIMEDOUT,        // Number of writes ending up with timed-out.
  WRITE_WITH_WAL,       // Number of Write calls that request WAL
  COMPACT_READ_BYTES,   // Bytes read during compaction
  COMPACT_WRITE_BYTES,  // Bytes written during compaction
+  FLUSH_WRITE_BYTES,    // Bytes written during flush

  // Number of table's properties loaded directly from file, without creating
  // table reader object.
@@ -125,6 +127,7 @@ enum Tickers {
  NUMBER_SUPERVERSION_ACQUIRES,
  NUMBER_SUPERVERSION_RELEASES,
  NUMBER_SUPERVERSION_CLEANUPS,
+  NUMBER_BLOCK_NOT_COMPRESSED,
  TICKER_ENUM_MAX
 };

@@ -175,7 +178,9 @@ const std::vector<std::pair<Tickers, std::string>> TickersNameMap = {
    {WAL_FILE_BYTES, "rocksdb.wal.bytes"},
    {WRITE_DONE_BY_SELF, "rocksdb.write.self"},
    {WRITE_DONE_BY_OTHER, "rocksdb.write.other"},
+    {WRITE_TIMEDOUT, "rocksdb.write.timedout"},
    {WRITE_WITH_WAL, "rocksdb.write.wal"},
+    {FLUSH_WRITE_BYTES, "rocksdb.flush.write.bytes"},
    {COMPACT_READ_BYTES, "rocksdb.compact.read.bytes"},
    {COMPACT_WRITE_BYTES, "rocksdb.compact.write.bytes"},
    {NUMBER_DIRECT_LOAD_TABLE_PROPERTIES,
@@ -183,6 +188,7 @@ const std::vector<std::pair<Tickers, std::string>> TickersNameMap = {
    {NUMBER_SUPERVERSION_ACQUIRES, "rocksdb.number.superversion_acquires"},
    {NUMBER_SUPERVERSION_RELEASES, "rocksdb.number.superversion_releases"},
    {NUMBER_SUPERVERSION_CLEANUPS, "rocksdb.number.superversion_cleanups"},
+    {NUMBER_BLOCK_NOT_COMPRESSED, "rocksdb.number.block.not_compressed"},
 };

 /**
@@ -65,6 +65,12 @@ class Status {
                                   const Slice& msg2 = Slice()) {
    return Status(kShutdownInProgress, msg, msg2);
  }
+  static Status TimedOut() {
+    return Status(kTimedOut);
+  }
+  static Status TimedOut(const Slice& msg, const Slice& msg2 = Slice()) {
+    return Status(kTimedOut, msg, msg2);
+  }

  // Returns true iff the status indicates success.
  bool ok() const { return code() == kOk; }
@@ -93,6 +99,8 @@ class Status {
  // Returns true iff the status indicates Incomplete
  bool IsShutdownInProgress() const { return code() == kShutdownInProgress; }

+  bool IsTimedOut() const { return code() == kTimedOut; }
+
  // Return a string representation of this status suitable for printing.
  // Returns the string "OK" for success.
  std::string ToString() const;
@@ -106,7 +114,8 @@ class Status {
    kIOError = 5,
    kMergeInProgress = 6,
    kIncomplete = 7,
-    kShutdownInProgress = 8
+    kShutdownInProgress = 8,
+    kTimedOut = 9
  };

  Code code() const {
@@ -74,6 +74,12 @@ struct BlockBasedTableOptions {

  IndexType index_type = kBinarySearch;

+  // Influence the behavior when kHashSearch is used.
+  // if false, stores a precise prefix to block range mapping
+  // if true, does not store prefix and allows prefix hash collision
+  // (less memory consumption)
+  bool hash_index_allow_collision = true;
+
  // Use the specified checksum type. Newly created table files will be
  // protected with this checksum type. Old table files will still be readable,
  // even though they have different checksum type.
@@ -91,6 +97,30 @@ extern TableFactory* NewBlockBasedTableFactory(
    const BlockBasedTableOptions& table_options = BlockBasedTableOptions());

 #ifndef ROCKSDB_LITE
+
+enum EncodingType : char {
+  // Always write full keys without any special encoding.
+  kPlain,
+  // Find opportunity to write the same prefix once for multiple rows.
+  // In some cases, when a key follows a previous key with the same prefix,
+  // instead of writing out the full key, it just writes out the size of the
+  // shared prefix, as well as other bytes, to save some bytes.
+  //
+  // When using this option, the user is required to use the same prefix
+  // extractor to make sure the same prefix will be extracted from the same key.
+  // The Name() value of the prefix extractor will be stored in the file. When
+  // reopening the file, the name of the options.prefix_extractor given will be
+  // bitwise compared to the prefix extractors stored in the file. An error
+  // will be returned if the two don't match.
+  kPrefix,
+};
+
+// Table Properties that are specific to plain table properties.
+struct PlainTablePropertyNames {
+  static const std::string kPrefixExtractorName;
+  static const std::string kEncodingType;
+};
+
 // -- Plain Table with prefix-only seek
 // For this factory, you need to set Options.prefix_extrator properly to make it
 // work. Look-up will starts with prefix hash lookup for key prefix. Inside the
@@ -107,41 +137,29 @@ extern TableFactory* NewBlockBasedTableFactory(
 //                    in the hash table
 // @index_sparseness: inside each prefix, need to build one index record for how
 //                    many keys for binary search inside each hash bucket.
+//                    For encoding type kPrefix, the value will be used when
+//                    writing to determine an interval to rewrite the full key.
+//                    It will also be used as a suggestion and satisfied when
+//                    possible.
 // @huge_page_tlb_size: if <=0, allocate hash indexes and blooms from malloc.
 //                      Otherwise from huge page TLB. The user needs to reserve
 //                      huge pages for it to be allocated, like:
 //                          sysctl -w vm.nr_hugepages=20
 //                      See linux doc Documentation/vm/hugetlbpage.txt
+// @encoding_type: how to encode the keys. See enum EncodingType above for
+//                 the choices. The value will determine how to encode keys
+//                 when writing to a new SST file. This value will be stored
+//                 inside the SST file which will be used when reading from the
+//                 file, which makes it possible for users to choose different
+//                 encoding type when reopening a DB. Files with different
+//                 encoding types can co-exist in the same DB and can be read.

 const uint32_t kPlainTableVariableLength = 0;
-extern TableFactory* NewPlainTableFactory(uint32_t user_key_len =
-                                              kPlainTableVariableLength,
-                                          int bloom_bits_per_prefix = 10,
-                                          double hash_table_ratio = 0.75,
-                                          size_t index_sparseness = 16,
-                                          size_t huge_page_tlb_size = 0);
-
-// -- Plain Table
-// This factory of plain table ignores Options.prefix_extractor and assumes no
-// hashable prefix available to the key structure. Lookup will be based on
-// binary search index only. Total order seek() can be issued.
-// @user_key_len: plain table has optimization for fix-sized keys, which can be
-//                specified via user_key_len.  Alternatively, you can pass
-//                `kPlainTableVariableLength` if your keys have variable
-//                lengths.
-// @bloom_bits_per_key: the number of bits used for bloom filer per key. You may
-//                  disable it by passing a zero.
-// @index_sparseness: need to build one index record for how many keys for
-//                    binary search.
-// @huge_page_tlb_size: if <=0, allocate hash indexes and blooms from malloc.
-//                      Otherwise from huge page TLB. The user needs to reserve
-//                      huge pages for it to be allocated, like:
-//                          sysctl -w vm.nr_hugepages=20
-//                      See linux doc Documentation/vm/hugetlbpage.txt
-extern TableFactory* NewTotalOrderPlainTableFactory(
+extern TableFactory* NewPlainTableFactory(
    uint32_t user_key_len = kPlainTableVariableLength,
-    int bloom_bits_per_key = 0, size_t index_sparseness = 16,
-    size_t huge_page_tlb_size = 0);
+    int bloom_bits_per_prefix = 10, double hash_table_ratio = 0.75,
+    size_t index_sparseness = 16, size_t huge_page_tlb_size = 0,
+    EncodingType encoding_type = kPlain, bool full_scan_mode = false);

 #endif  // ROCKSDB_LITE

@@ -203,4 +221,19 @@ class TableFactory {
      WritableFile* file, CompressionType compression_type) const = 0;
 };

+#ifndef ROCKSDB_LITE
+// Create a special table factory that can open both of block based table format
+// and plain table, based on setting inside the SST files. It should be used to
+// convert a DB from one table format to another.
+// @table_factory_to_write: the table factory used when writing to new files.
+// @block_based_table_factory:  block based table factory to use. If NULL, use
+//                              a default one.
+// @plain_table_factory: plain table factory to use. If NULL, use a default one.
+extern TableFactory* NewAdaptiveTableFactory(
+    std::shared_ptr<TableFactory> table_factory_to_write = nullptr,
+    std::shared_ptr<TableFactory> block_based_table_factory = nullptr,
+    std::shared_ptr<TableFactory> plain_table_factory = nullptr);
+
+#endif  // ROCKSDB_LITE
+
 }  // namespace rocksdb
@@ -8,6 +8,7 @@

 #include <stdint.h>
 #include <climits>
+#include <vector>

 namespace rocksdb {

@@ -61,6 +62,7 @@ class CompactionOptionsUniversal {
  // well as  the total size of C1...Ct as total_C, the compaction output file
  // will be compressed iff
  //   total_C / total_size < this percentage
+  // Default: -1
  int compression_size_percent;

  // The algorithm used to stop picking files into a single compaction run
@@ -68,14 +70,13 @@ class CompactionOptionsUniversal {
  CompactionStopStyle stop_style;

  // Default set of parameters
-  CompactionOptionsUniversal() :
-    size_ratio(1),
-    min_merge_width(2),
-    max_merge_width(UINT_MAX),
-    max_size_amplification_percent(200),
-    compression_size_percent(-1),
-    stop_style(kCompactionStopStyleTotalSize) {
-  }
+  CompactionOptionsUniversal()
+      : size_ratio(1),
+        min_merge_width(2),
+        max_merge_width(UINT_MAX),
+        max_size_amplification_percent(200),
+        compression_size_percent(-1),
+        stop_style(kCompactionStopStyleTotalSize) {}
 };

 }  // namespace rocksdb
@@ -6,7 +6,7 @@

 // Also update Makefile if you change these
 #define ROCKSDB_MAJOR 3
-#define ROCKSDB_MINOR 1
+#define ROCKSDB_MINOR 3
 #define ROCKSDB_PATCH 0

 // Do not use these. We made the mistake of declaring macros starting with
@@ -67,6 +67,10 @@ class WriteBatch {
  void Delete(ColumnFamilyHandle* column_family, const Slice& key);
  void Delete(const Slice& key) { Delete(nullptr, key); }

+  // variant that takes SliceParts
+  void Delete(ColumnFamilyHandle* column_family, const SliceParts& key);
+  void Delete(const SliceParts& key) { Delete(nullptr, key); }
+
  // Append a blob of arbitrary size to the records in this batch. The blob will
  // be stored in the transaction log but not in any other file. In particular,
  // it will not be persisted to the SST files. When iterating over this
@@ -0,0 +1,149 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <string>
+#include <vector>
+
+#include "utilities/stackable_db.h"
+#include "utilities/json_document.h"
+#include "rocksdb/db.h"
+
+namespace rocksdb {
+
+// IMPORTANT: DocumentDB is a work in progress. It is unstable and we might
+// change the API without warning. Talk to RocksDB team before using this in
+// production ;)
+
+// DocumentDB is a layer on top of RocksDB that provides a very simple JSON API.
+// When creating a DB, you specify a list of indexes you want to keep on your
+// data. You can insert a JSON document to the DB, which is automatically
+// indexed. Every document added to the DB needs to have "_id" field which is
+// automatically indexed and is an unique primary key. All other indexes are
+// non-unique.
+
+// NOTE: field names in the JSON are NOT allowed to start with '$' or
+// contain '.'. We don't currently enforce that rule, but will start behaving
+// badly.
+
+// Cursor is what you get as a result of executing query. To get all
+// results from a query, call Next() on a Cursor while  Valid() returns true
+class Cursor {
+ public:
+  Cursor() = default;
+  virtual ~Cursor() {}
+
+  virtual bool Valid() const = 0;
+  virtual void Next() = 0;
+  // Lifecycle of the returned JSONDocument is until the next Next() call
+  virtual const JSONDocument& document() const = 0;
+  virtual Status status() const = 0;
+
+ private:
+  // No copying allowed
+  Cursor(const Cursor&);
+  void operator=(const Cursor&);
+};
+
+struct DocumentDBOptions {
+  int background_threads = 4;
+  uint64_t memtable_size = 128 * 1024 * 1024;    // 128 MB
+  uint64_t cache_size = 1 * 1024 * 1024 * 1024;  // 1 GB
+};
+
+// TODO(icanadi) Add `JSONDocument* info` parameter to all calls that can be
+// used by the caller to get more information about the call execution (number
+// of dropped records, number of updated records, etc.)
+class DocumentDB : public StackableDB {
+ public:
+  struct IndexDescriptor {
+    // Currently, you can only define an index on a single field. To specify an
+    // index on a field X, set index description to JSON "{X: 1}"
+    // Currently the value needs to be 1, which means ascending.
+    // In the future, we plan to also support indexes on multiple keys, where
+    // you could mix ascending sorting (1) with descending sorting indexes (-1)
+    JSONDocument* description;
+    std::string name;
+  };
+
+  // Open DocumentDB with specified indexes. The list of indexes has to be
+  // complete, i.e. include all indexes present in the DB, except the primary
+  // key index.
+  // Otherwise, Open() will return an error
+  static Status Open(const DocumentDBOptions& options, const std::string& name,
+                     const std::vector<IndexDescriptor>& indexes,
+                     DocumentDB** db, bool read_only = false);
+
+  explicit DocumentDB(DB* db) : StackableDB(db) {}
+
+  // Create a new index. It will stop all writes for the duration of the call.
+  // All current documents in the DB are scanned and corresponding index entries
+  // are created
+  virtual Status CreateIndex(const WriteOptions& write_options,
+                             const IndexDescriptor& index) = 0;
+
+  // Drop an index. Client is responsible to make sure that index is not being
+  // used by currently executing queries
+  virtual Status DropIndex(const std::string& name) = 0;
+
+  // Insert a document to the DB. The document needs to have a primary key "_id"
+  // which can either be a string or an integer. Otherwise the write will fail
+  // with InvalidArgument.
+  virtual Status Insert(const WriteOptions& options,
+                        const JSONDocument& document) = 0;
+
+  // Deletes all documents matching a filter atomically
+  virtual Status Remove(const ReadOptions& read_options,
+                        const WriteOptions& write_options,
+                        const JSONDocument& query) = 0;
+
+  // Does this sequence of operations:
+  // 1. Find all documents matching a filter
+  // 2. For all documents, atomically:
+  // 2.1. apply the update operators
+  // 2.2. update the secondary indexes
+  //
+  // Currently only $set update operator is supported.
+  // Syntax is: {$set: {key1: value1, key2: value2, etc...}}
+  // This operator will change a document's key1 field to value1, key2 to
+  // value2, etc. New values will be set even if a document didn't have an entry
+  // for the specified key.
+  //
+  // You can not change a primary key of a document.
+  //
+  // Update example: Update({id: {$gt: 5}, $index: id}, {$set: {enabled: true}})
+  virtual Status Update(const ReadOptions& read_options,
+                        const WriteOptions& write_options,
+                        const JSONDocument& filter,
+                        const JSONDocument& updates) = 0;
+
+  // query has to be an array in which every element is an operator. Currently
+  // only $filter operator is supported. Syntax of $filter operator is:
+  // {$filter: {key1: condition1, key2: condition2, etc.}} where conditions can
+  // be either:
+  // 1) a single value in which case the condition is equality condition, or
+  // 2) a defined operators, like {$gt: 4}, which will match all documents that
+  // have key greater than 4.
+  //
+  // Supported operators are:
+  // 1) $gt -- greater than
+  // 2) $gte -- greater than or equal
+  // 3) $lt -- less than
+  // 4) $lte -- less than or equal
+  // If you want the filter to use an index, you need to specify it like this:
+  // {$filter: {...(conditions)..., $index: index_name}}
+  //
+  // Example query:
+  // * [{$filter: {name: John, age: {$gte: 18}, $index: age}}]
+  // will return all Johns whose age is greater or equal to 18 and it will use
+  // index "age" to satisfy the query.
+  virtual Cursor* Query(const ReadOptions& read_options,
+                        const JSONDocument& query) = 0;
+};
+
+}  // namespace rocksdb
+#endif  // ROCKSDB_LITE
@@ -0,0 +1,174 @@
+//  Copyright (c) 2013, Facebook, Inc.  All rights reserved.
+//  This source code is licensed under the BSD-style license found in the
+//  LICENSE file in the root directory of this source tree. An additional grant
+//  of patent rights can be found in the PATENTS file in the same directory.
+#pragma once
+#ifndef ROCKSDB_LITE
+
+#include <string>
+#include <map>
+#include <unordered_map>
+#include <vector>
+
+#include "rocksdb/slice.h"
+
+// We use JSONDocument for DocumentDB API
+// Implementation inspired by folly::dynamic and rapidjson
+
+namespace rocksdb {
+
+// NOTE: none of this is thread-safe
+class JSONDocument {
+ public:
+  // return nullptr on parse failure
+  static JSONDocument* ParseJSON(const char* json);
+
+  enum Type {
+    kNull,
+    kArray,
+    kBool,
+    kDouble,
+    kInt64,
+    kObject,
+    kString,
+  };
+
+  JSONDocument();  // null
+  /* implicit */ JSONDocument(bool b);
+  /* implicit */ JSONDocument(double d);
+  /* implicit */ JSONDocument(int64_t i);
+  /* implicit */ JSONDocument(const std::string& s);
+  /* implicit */ JSONDocument(const char* s);
+  // constructs JSONDocument of specific type with default value
+  explicit JSONDocument(Type type);
+
+  // copy constructor
+  JSONDocument(const JSONDocument& json_document);
+
+  ~JSONDocument();
+
+  Type type() const;
+
+  // REQUIRES: IsObject()
+  bool Contains(const std::string& key) const;
+  // Returns nullptr if !Contains()
+  // don't delete the returned pointer
+  // REQUIRES: IsObject()
+  const JSONDocument* Get(const std::string& key) const;
+  // REQUIRES: IsObject()
+  JSONDocument& operator[](const std::string& key);
+  // REQUIRES: IsObject()
+  const JSONDocument& operator[](const std::string& key) const;
+  // returns `this`, so you can chain operations.
+  // Copies value
+  // REQUIRES: IsObject()
+  JSONDocument* Set(const std::string& key, const JSONDocument& value);
+
+  // REQUIRES: IsArray() == true || IsObject() == true
+  size_t Count() const;
+
+  // REQUIRES: IsArray()
+  const JSONDocument* GetFromArray(size_t i) const;
+  // REQUIRES: IsArray()
+  JSONDocument& operator[](size_t i);
+  // REQUIRES: IsArray()
+  const JSONDocument& operator[](size_t i) const;
+  // returns `this`, so you can chain operations.
+  // Copies the value
+  // REQUIRES: IsArray() && i < Count()
+  JSONDocument* SetInArray(size_t i, const JSONDocument& value);
+  // REQUIRES: IsArray()
+  JSONDocument* PushBack(const JSONDocument& value);
+
+  bool IsNull() const;
+  bool IsArray() const;
+  bool IsBool() const;
+  bool IsDouble() const;
+  bool IsInt64() const;
+  bool IsObject() const;
+  bool IsString() const;
+
+  // REQUIRES: IsBool() == true
+  bool GetBool() const;
+  // REQUIRES: IsDouble() == true
+  double GetDouble() const;
+  // REQUIRES: IsInt64() == true
+  int64_t GetInt64() const;
+  // REQUIRES: IsString() == true
+  const std::string& GetString() const;
+
+  bool operator==(const JSONDocument& rhs) const;
+
+  std::string DebugString() const;
+
+ private:
+  class ItemsIteratorGenerator;
+
+ public:
+  // REQUIRES: IsObject()
+  ItemsIteratorGenerator Items() const;
+
+  // appends serialized object to dst
+  void Serialize(std::string* dst) const;
+  // returns nullptr if Slice doesn't represent valid serialized JSONDocument
+  static JSONDocument* Deserialize(const Slice& src);
+
+ private:
+  void SerializeInternal(std::string* dst, bool type_prefix) const;
+  // returns false if Slice doesn't represent valid serialized JSONDocument.
+  // Otherwise, true
+  bool DeserializeInternal(Slice* input);
+
+  typedef std::vector<JSONDocument*> Array;
+  typedef std::unordered_map<std::string, JSONDocument*> Object;
+
+  // iteration on objects
+  class const_item_iterator {
+   public:
+    typedef Object::const_iterator It;
+    typedef Object::value_type value_type;
+    /* implicit */ const_item_iterator(It it) : it_(it) {}
+    It& operator++() { return ++it_; }
+    bool operator!=(const const_item_iterator& other) {
+      return it_ != other.it_;
+    }
+    value_type operator*() { return *it_; }
+
+   private:
+    It it_;
+  };
+  class ItemsIteratorGenerator {
+   public:
+    /* implicit */ ItemsIteratorGenerator(const Object& object)
+        : object_(object) {}
+    const_item_iterator begin() { return object_.begin(); }
+    const_item_iterator end() { return object_.end(); }
+
+   private:
+    const Object& object_;
+  };
+
+  union Data {
+    Data() : n(nullptr) {}
+    ~Data() {}
+
+    void* n;
+    Array a;
+    bool b;
+    double d;
+    int64_t i;
+    std::string s;
+    Object o;
+  } data_;
+  const Type type_;
+
+  // Our serialization format's first byte specifies the encoding version. That
+  // way, we can easily change our format while providing backwards
+  // compatibility. This constant specifies the current version of the
+  // serialization format
+  static const char kSerializationFormatVersion;
+};
+
+}  // namespace rocksdb
+
+#endif  // ROCKSDB_LITE
@@ -171,6 +171,11 @@ class StackableDB : public DB {
    return db_->EnableFileDeletions(force);
  }

+  virtual void GetLiveFilesMetaData(
+      std::vector<LiveFileMetaData>* metadata) override {
+    db_->GetLiveFilesMetaData(metadata);
+  }
+
  virtual Status GetLiveFiles(std::vector<std::string>& vec, uint64_t* mfs,
                              bool flush_memtable = true) override {
      return db_->GetLiveFiles(vec, mfs, flush_memtable);
@@ -0,0 +1,54 @@
+# RocksJava Change Log
+
+## By 06/15/2014
+### New Features
+* Added basic Java binding for rocksdb::Env such that multiple RocksDB can share the same thread pool and environment.
+* Added RestoreBackupableDB
+
+## By 05/30/2014
+### Internal Framework Improvement
+* Added disOwnNativeHandle to RocksObject, which allows a RocksObject to give-up the ownership of its native handle.  This method is useful when sharing and transferring the ownership of RocksDB C++ resources.
+
+## By 05/15/2014
+### New Features
+* Added RocksObject --- the base class of all RocksDB classes which holds some RocksDB resources in the C++ side.
+* Use environmental variable JAVA_HOME in Makefile for RocksJava
+### Public API changes
+* Renamed org.rocksdb.Iterator to org.rocksdb.RocksIterator to avoid potential confliction with Java built-in Iterator.
+
+## By 04/30/2014
+### New Features
+* Added Java binding for MultiGet.
+* Added static method RocksDB.loadLibrary(), which loads necessary library files.
+* Added Java bindings for 60+ rocksdb::Options.
+* Added Java binding for BloomFilter.
+* Added Java binding for ReadOptions.
+* Added Java binding for memtables.
+* Added Java binding for sst formats.
+* Added Java binding for RocksDB Iterator which enables sequential scan operation.
+* Added Java binding for Statistics
+* Added Java binding for BackupableDB.
+
+### DB Benchmark
+* Added filluniquerandom, readseq benchmark.
+* 70+ command-line options.
+* Enabled BloomFilter configuration.
+
+## By 04/15/2014
+### New Features
+* Added Java binding for WriteOptions.
+* Added Java binding for WriteBatch, which enables batch-write.
+* Added Java binding for rocksdb::Options.
+* Added Java binding for block cache.
+* Added Java version DB Benchmark.
+
+### DB Benchmark
+* Added readwhilewriting benchmark.
+
+### Internal Framework Improvement
+* Avoid a potential byte-array-copy between c++ and Java in RocksDB.get.
+* Added SizeUnit in org.rocksdb.util to store consts like KB and GB.
+
+### 03/28/2014
+* RocksJava project started.
+* Added Java binding for RocksDB, which supports Open, Close, Get and Put.
@@ -1,7 +1,13 @@
-NATIVE_JAVA_CLASSES = org.rocksdb.RocksDB org.rocksdb.Options org.rocksdb.WriteBatch org.rocksdb.WriteBatchInternal org.rocksdb.WriteBatchTest org.rocksdb.WriteOptions org.rocksdb.BackupableDB org.rocksdb.BackupableDBOptions org.rocksdb.Statistics org.rocksdb.RocksIterator org.rocksdb.VectorMemTableConfig org.rocksdb.SkipListMemTableConfig org.rocksdb.HashLinkedListMemTableConfig org.rocksdb.HashSkipListMemTableConfig org.rocksdb.PlainTableConfig org.rocksdb.ReadOptions org.rocksdb.Filter org.rocksdb.BloomFilter
+NATIVE_JAVA_CLASSES = org.rocksdb.RocksDB org.rocksdb.Options org.rocksdb.WriteBatch org.rocksdb.WriteBatchInternal org.rocksdb.WriteBatchTest org.rocksdb.WriteOptions org.rocksdb.BackupableDB org.rocksdb.BackupableDBOptions org.rocksdb.Statistics org.rocksdb.RocksIterator org.rocksdb.VectorMemTableConfig org.rocksdb.SkipListMemTableConfig org.rocksdb.HashLinkedListMemTableConfig org.rocksdb.HashSkipListMemTableConfig org.rocksdb.PlainTableConfig org.rocksdb.ReadOptions org.rocksdb.Filter org.rocksdb.BloomFilter org.rocksdb.RestoreOptions org.rocksdb.RestoreBackupableDB org.rocksdb.RocksEnv
+
 NATIVE_INCLUDE = ./include
 ROCKSDB_JAR = rocksdbjni.jar

+ROCKSDBJNILIB = ./librocksdbjni.so
+ifeq ($(PLATFORM), OS_MACOSX)
+  ROCKSDBJNILIB = ./librocksdbjni.jnilib
+endif
+
 clean:
 	-find . -name "*.class" -exec rm {} \;
 	-find . -name "hs*.log" -exec rm {} \;
@@ -9,7 +15,9 @@ clean:

 java:
 	javac org/rocksdb/util/*.java org/rocksdb/*.java
-	jar -cf $(ROCKSDB_JAR) org/rocksdb/*.class org/rocksdb/util/*.class
+	@cp ../HISTORY.md ./HISTORY-CPP.md
+	jar -cf $(ROCKSDB_JAR) org/rocksdb/*.class org/rocksdb/util/*.class HISTORY*.md $(ROCKSDBJNILIB)
+	@rm -f ./HISTORY-CPP.md
 	javah -d $(NATIVE_INCLUDE) -jni $(NATIVE_JAVA_CLASSES)

 sample: java
@@ -26,6 +34,7 @@ test: java
 	java -ea -Djava.library.path=.:../ -cp "$(ROCKSDB_JAR):.:./*" org.rocksdb.test.BackupableDBTest
 	java -ea -Djava.library.path=.:../ -cp "$(ROCKSDB_JAR):.:./*" org.rocksdb.test.OptionsTest
 	java -ea -Djava.library.path=.:../ -cp "$(ROCKSDB_JAR):.:./*" org.rocksdb.test.ReadOptionsTest
+	java -ea -Djava.library.path=.:../ -cp "$(ROCKSDB_JAR):.:./*" org.rocksdb.test.StatisticsCollectorTest

 db_bench: java
 	javac org/rocksdb/benchmark/*.java
@@ -43,7 +43,8 @@ public class RocksDBSample {
        .setDisableSeekCompaction(true)
        .setBlockSize(64 * SizeUnit.KB)
        .setMaxBackgroundCompactions(10)
-        .setFilter(filter);
+        .setFilter(filter)
+        .setCompressionType(CompressionType.SNAPPY_COMPRESSION);
    Statistics stats = options.statisticsPtr();

    assert(options.createIfMissing() == true);
@@ -52,6 +53,7 @@ public class RocksDBSample {
    assert(options.disableSeekCompaction() == true);
    assert(options.blockSize() == 64 * SizeUnit.KB);
    assert(options.maxBackgroundCompactions() == 10);
+    assert(options.compressionType() == CompressionType.SNAPPY_COMPRESSION);

    assert(options.memTableFactoryName().equals("SkipListFactory"));
    options.setMemTableConfig(
@@ -25,11 +25,14 @@ public class BackupableDB extends RocksDB {
  public static BackupableDB open(
      Options opt, BackupableDBOptions bopt, String db_path)
      throws RocksDBException {
-    // since BackupableDB c++ will handle the life cycle of
-    // the returned RocksDB of RocksDB.open(), here we store
-    // it as a BackupableDB member variable to avoid GC.
-    BackupableDB bdb = new BackupableDB(RocksDB.open(opt, db_path));
-    bdb.open(bdb.db_.nativeHandle_, bopt.nativeHandle_);
+
+    RocksDB db = RocksDB.open(opt, db_path);
+    BackupableDB bdb = new BackupableDB();
+    bdb.open(db.nativeHandle_, bopt.nativeHandle_);
+
+    // Prevent the RocksDB object from attempting to delete
+    // the underly C++ DB object.
+    db.disOwnNativeHandle();

    return bdb;
  }
@@ -64,9 +67,8 @@ public class BackupableDB extends RocksDB {
   * A protected construction that will be used in the static factory
   * method BackupableDB.open().
   */
-  protected BackupableDB(RocksDB db) {
+  protected BackupableDB() {
    super();
-    db_ = db;
  }

  @Override protected void finalize() {
@@ -75,6 +77,4 @@ public class BackupableDB extends RocksDB {

  protected native void open(long rocksDBHandle, long backupDBOptionsHandle);
  protected native void createNewBackup(long handle, boolean flag);
-
-  private final RocksDB db_;
 }
@@ -11,11 +11,39 @@ package org.rocksdb;
 *
 * Note that dispose() must be called before an Options instance
 * become out-of-scope to release the allocated memory in c++.
+ *
+ * @param path Where to keep the backup files. Has to be different than dbname.
+       Best to set this to dbname_ + "/backups"
+ * @param shareTableFiles If share_table_files == true, backup will assume that
+ *     table files with same name have the same contents. This enables
+ *     incremental backups and avoids unnecessary data copies. If
+ *     share_table_files == false, each backup will be on its own and will not
+ *     share any data with other backups. default: true
+ * @param sync If sync == true, we can guarantee you'll get consistent backup
+ *     even on a machine crash/reboot. Backup process is slower with sync
+ *     enabled. If sync == false, we don't guarantee anything on machine reboot.
+ *     However, chances are some of the backups are consistent. Default: true
+ * @param destroyOldData If true, it will delete whatever backups there are
+ *     already. Default: false
+ * @param backupLogFiles If false, we won't backup log files. This option can be
+ *     useful for backing up in-memory databases where log file are persisted,
+ *     but table files are in memory. Default: true
+ * @param backupRateLimit Max bytes that can be transferred in a second during
+ *     backup. If 0 or negative, then go as fast as you can. Default: 0
+ * @param restoreRateLimit Max bytes that can be transferred in a second during
+ *     restore. If 0 or negative, then go as fast as you can. Default: 0
 */
 public class BackupableDBOptions extends RocksObject {
-  public BackupableDBOptions(String path) {
+  public BackupableDBOptions(String path, boolean shareTableFiles, boolean sync,
+      boolean destroyOldData, boolean backupLogFiles, long backupRateLimit,
+      long restoreRateLimit) {
    super();
-    newBackupableDBOptions(path);
+
+    backupRateLimit = (backupRateLimit <= 0) ? 0 : backupRateLimit;
+    restoreRateLimit = (restoreRateLimit <= 0) ? 0 : restoreRateLimit;
+
+    newBackupableDBOptions(path, shareTableFiles, sync, destroyOldData,
+        backupLogFiles, backupRateLimit, restoreRateLimit);
  }

  /**
@@ -32,13 +60,14 @@ public class BackupableDBOptions extends RocksObject {
   * Release the memory allocated for the current instance
   * in the c++ side.
   */
-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose(nativeHandle_);
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

-  private native void newBackupableDBOptions(String path);
+  private native void newBackupableDBOptions(String path,
+      boolean shareTableFiles, boolean sync, boolean destroyOldData,
+      boolean backupLogFiles, long backupRateLimit, long restoreRateLimit);
  private native String backupDir(long handle);
-  private native void dispose(long handle);
+  private native void disposeInternal(long handle);
 }
@@ -0,0 +1,25 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+public enum CompressionType {
+  NO_COMPRESSION((byte) 0),
+  SNAPPY_COMPRESSION((byte) 1),
+  ZLIB_COMPRESSION((byte) 2),
+  BZLIB2_COMPRESSION((byte) 3),
+  LZ4_COMPRESSION((byte) 4),
+  LZ4HC_COMPRESSION((byte) 5);
+  
+  private final byte value_;
+
+  private CompressionType(byte value) {
+    value_ = value;
+  }
+
+  public byte getValue() {
+    return value_;
+  }
+}
@@ -22,11 +22,10 @@ public abstract class Filter extends RocksObject {
   * RocksDB instances referencing the filter are closed.
   * Otherwise an undefined behavior will occur.
   */
-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose0(nativeHandle_);
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

-  private native void dispose0(long handle);
+  private native void disposeInternal(long handle);
 }
@@ -24,6 +24,7 @@ public class Options extends RocksObject {
    super();
    cacheSize_ = DEFAULT_CACHE_SIZE;
    newOptions();
+    env_ = RocksEnv.getDefault();
  }

  /**
@@ -42,6 +43,24 @@ public class Options extends RocksObject {
    return this;
  }

+  /**
+   * Use the specified object to interact with the environment,
+   * e.g. to read/write files, schedule background work, etc.
+   * Default: RocksEnv.getDefault()
+   */
+  public Options setEnv(RocksEnv env) {
+    assert(isInitialized());
+    setEnv(nativeHandle_, env.nativeHandle_);
+    env_ = env;
+    return this;
+  }
+  private native void setEnv(long optHandle, long envHandle);
+
+  public RocksEnv getEnv() {
+    return env_;
+  }
+  private native long getEnvHandle(long handle);
+
  /**
   * Return true if the create_if_missing flag is set to true.
   * If true, the database will be created if it is missing.
@@ -1288,6 +1307,34 @@ public class Options extends RocksObject {
  }
  private native void setBlockRestartInterval(
      long handle, int blockRestartInterval);
+      
+  /**
+   * Compress blocks using the specified compression algorithm.  This
+     parameter can be changed dynamically.
+   * 
+   * Default: SNAPPY_COMPRESSION, which gives lightweight but fast compression.
+   * 
+   * @return Compression type.
+   */ 
+  public CompressionType compressionType() {
+    return CompressionType.values()[compressionType(nativeHandle_)];
+  }
+  private native byte compressionType(long handle);
+      
+  /**
+   * Compress blocks using the specified compression algorithm.  This
+     parameter can be changed dynamically.
+   * 
+   * Default: SNAPPY_COMPRESSION, which gives lightweight but fast compression.
+   * 
+   * @param compressionType Compression Type.
+   * @return the reference to the current option.
+   */
+  public Options setCompressionType(CompressionType compressionType) {
+    setCompressionType(nativeHandle_, compressionType.getValue());
+    return this;
+  }
+  private native void setCompressionType(long handle, byte compressionType);

  /**
   * If true, place whole keys in the filter (not just prefixes).
@@ -2311,10 +2358,9 @@ public class Options extends RocksObject {
   * Release the memory allocated for the current instance
   * in the c++ side.
   */
-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose0();
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

  static final int DEFAULT_PLAIN_TABLE_BLOOM_BITS_PER_KEY = 10;
@@ -2322,7 +2368,7 @@ public class Options extends RocksObject {
  static final int DEFAULT_PLAIN_TABLE_INDEX_SPARSENESS = 16;

  private native void newOptions();
-  private native void dispose0();
+  private native void disposeInternal(long handle);
  private native void setCreateIfMissing(long handle, boolean flag);
  private native boolean createIfMissing(long handle);
  private native void setWriteBufferSize(long handle, long writeBufferSize);
@@ -2352,4 +2398,5 @@ public class Options extends RocksObject {

  long cacheSize_;
  Filter filter_;
+  RocksEnv env_;
 }
@@ -18,19 +18,6 @@ public class ReadOptions extends RocksObject {
  }
  private native void newReadOptions();

-  /**
-   * Release the memory allocated for the current instance
-   * in the c++ side.
-   *
-   * Calling other methods after dispose() leads to undefined behavior.
-   */
-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose(nativeHandle_);
-    }
-  }
-  private native void dispose(long handle);
-
  /**
   * If true, all data read from underlying storage will be
   * verified against corresponding checksums.
@@ -127,4 +114,12 @@ public class ReadOptions extends RocksObject {
  }
  private native void setTailing(
      long handle, boolean tailing);
+
+
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
+  }
+  private native void disposeInternal(long handle);
+
 }
@@ -0,0 +1,84 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+/**
+ * This class is used to access information about backups and restore from them.
+ *
+ * Note that dispose() must be called before this instance become out-of-scope
+ * to release the allocated memory in c++.
+ *
+ * @param options Instance of BackupableDBOptions.
+ */
+public class RestoreBackupableDB extends RocksObject {
+  public RestoreBackupableDB(BackupableDBOptions options) {
+    super();
+    nativeHandle_ = newRestoreBackupableDB(options.nativeHandle_);
+  }
+
+  /**
+   * Restore from backup with backup_id
+   * IMPORTANT -- if options_.share_table_files == true and you restore DB
+   * from some backup that is not the latest, and you start creating new
+   * backups from the new DB, they will probably fail.
+   *
+   * Example: Let's say you have backups 1, 2, 3, 4, 5 and you restore 3.
+   * If you add new data to the DB and try creating a new backup now, the
+   * database will diverge from backups 4 and 5 and the new backup will fail.
+   * If you want to create new backup, you will first have to delete backups 4
+   * and 5.
+   */
+  public void restoreDBFromBackup(long backupId, String dbDir, String walDir,
+      RestoreOptions restoreOptions) throws RocksDBException {
+    restoreDBFromBackup0(nativeHandle_, backupId, dbDir, walDir,
+        restoreOptions.nativeHandle_);
+  }
+
+  /**
+   * Restore from the latest backup.
+   */
+  public void restoreDBFromLatestBackup(String dbDir, String walDir,
+      RestoreOptions restoreOptions) throws RocksDBException {
+    restoreDBFromLatestBackup0(nativeHandle_, dbDir, walDir,
+        restoreOptions.nativeHandle_);
+  }
+
+  /**
+   * Deletes old backups, keeping latest numBackupsToKeep alive.
+   *
+   * @param Number of latest backups to keep
+   */
+  public void purgeOldBackups(int numBackupsToKeep) throws RocksDBException {
+    purgeOldBackups0(nativeHandle_, numBackupsToKeep);
+  }
+
+  /**
+   * Deletes a specific backup.
+   *
+   * @param ID of backup to delete.
+   */
+  public void deleteBackup(long backupId) throws RocksDBException {
+    deleteBackup0(nativeHandle_, backupId);
+  }
+
+  /**
+   * Release the memory allocated for the current instance
+   * in the c++ side.
+   */
+  @Override public synchronized void disposeInternal() {
+    assert(isInitialized());
+    dispose(nativeHandle_);
+  }
+
+  private native long newRestoreBackupableDB(long options);
+  private native void restoreDBFromBackup0(long nativeHandle, long backupId,
+      String dbDir, String walDir, long restoreOptions) throws RocksDBException;
+  private native void restoreDBFromLatestBackup0(long nativeHandle,
+      String dbDir, String walDir, long restoreOptions) throws RocksDBException;
+  private native void purgeOldBackups0(long nativeHandle, int numBackupsToKeep);
+  private native void deleteBackup0(long nativeHandle, long backupId);
+  private native void dispose(long nativeHandle);
+}
@@ -0,0 +1,37 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+/**
+ * RestoreOptions to control the behavior of restore.
+ *
+ * Note that dispose() must be called before this instance become out-of-scope
+ * to release the allocated memory in c++.
+ *
+ * @param If true, restore won't overwrite the existing log files in wal_dir. It
+ *     will also move all log files from archive directory to wal_dir. Use this
+ *     option in combination with BackupableDBOptions::backup_log_files = false
+ *     for persisting in-memory databases.
+ *     Default: false
+ */
+public class RestoreOptions extends RocksObject {
+  public RestoreOptions(boolean keepLogFiles) {
+    super();
+    nativeHandle_ = newRestoreOptions(keepLogFiles);
+  }
+
+  /**
+   * Release the memory allocated for the current instance
+   * in the c++ side.
+   */
+  @Override public synchronized void disposeInternal() {
+    assert(isInitialized());
+    dispose(nativeHandle_);
+  }
+
+  private native long newRestoreOptions(boolean keepLogFiles);
+  private native void dispose(long handle);
+}
@@ -21,7 +21,7 @@ import org.rocksdb.util.Environment;
 public class RocksDB extends RocksObject {
  public static final int NOT_FOUND = -1;
  private static final String[] compressionLibs_ = {
-      "snappy", "zlib", "bzip2", "lz4", "lz4hc"};
+      "snappy", "z", "bzip2", "lz4", "lz4hc"};

  /**
   * Loads the necessary library files.
@@ -93,10 +93,7 @@ public class RocksDB extends RocksObject {
    // This allows to use the rocksjni default Options instead of
    // the c++ one.
    Options options = new Options();
-    db.open(options.nativeHandle_, options.cacheSize_, path);
-    db.transferCppRawPointersOwnershipFrom(options);
-    options.dispose();
-    return db;
+    return open(options, path);
  }

  /**
@@ -114,11 +111,9 @@ public class RocksDB extends RocksObject {
    return db;
  }

-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose(nativeHandle_);
-      nativeHandle_ = 0;
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

  /**
@@ -315,10 +310,6 @@ public class RocksDB extends RocksObject {
    return new RocksIterator(iterator0(nativeHandle_));
  }

-  @Override protected void finalize() {
-    close();
-  }
-
  /**
   * Private constructor.
   */
@@ -370,7 +361,7 @@ public class RocksDB extends RocksObject {
      long handle, long writeOptHandle,
      byte[] key, int keyLen) throws RocksDBException;
  protected native long iterator0(long optHandle);
-  protected native void dispose(long handle);
+  private native void disposeInternal(long handle);

  protected Filter filter_;
 }
@@ -0,0 +1,102 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+/**
+ * A RocksEnv is an interface used by the rocksdb implementation to access
+ * operating system functionality like the filesystem etc.
+ *
+ * All Env implementations are safe for concurrent access from
+ * multiple threads without any external synchronization.
+ */
+public class RocksEnv extends RocksObject {
+  public static final int FLUSH_POOL = 0;
+  public static final int COMPACTION_POOL = 1;
+
+  static {
+    default_env_ = new RocksEnv(getDefaultEnvInternal());
+  }
+  private static native long getDefaultEnvInternal();
+
+  /**
+   * Returns the default environment suitable for the current operating
+   * system.
+   *
+   * The result of getDefault() is a singleton whose ownership belongs
+   * to rocksdb c++.  As a result, the returned RocksEnv will not
+   * have the ownership of its c++ resource, and calling its dispose()
+   * will be no-op.
+   */
+  public static RocksEnv getDefault() {
+    return default_env_;
+  }
+
+  /**
+   * Sets the number of background worker threads of the flush pool
+   * for this environment.
+   * default number: 1
+   */
+  public RocksEnv setBackgroundThreads(int num) {
+    return setBackgroundThreads(num, FLUSH_POOL);
+  }
+
+  /**
+   * Sets the number of background worker threads of the specified thread
+   * pool for this environment.
+   *
+   * @param num the number of threads
+   * @param poolID the id to specified a thread pool.  Should be either
+   *     FLUSH_POOL or COMPACTION_POOL.
+   * Default number: 1
+   */
+  public RocksEnv setBackgroundThreads(int num, int poolID) {
+    setBackgroundThreads(nativeHandle_, num, poolID);
+    return this;
+  }
+  private native void setBackgroundThreads(
+      long handle, int num, int priority);
+
+  /**
+   * Returns the length of the queue associated with the specified
+   * thread pool.
+   *
+   * @param poolID the id to specified a thread pool.  Should be either
+   *     FLUSH_POOL or COMPACTION_POOL.
+   */
+  public int getThreadPoolQueueLen(int poolID) {
+    return getThreadPoolQueueLen(nativeHandle_, poolID);
+  }
+  private native int getThreadPoolQueueLen(long handle, int poolID);
+
+  /**
+   * Package-private constructor that uses the specified native handle
+   * to construct a RocksEnv.  Note that the ownership of the input handle
+   * belongs to the caller, and the newly created RocksEnv will not take
+   * the ownership of the input handle.  As a result, calling dispose()
+   * of the created RocksEnv will be no-op.
+   */
+  RocksEnv(long handle) {
+    super();
+    nativeHandle_ = handle;
+    disOwnNativeHandle();
+  }
+
+  /**
+   * The helper function of dispose() which all subclasses of RocksObject
+   * must implement to release their associated C++ resource.
+   */
+  protected void disposeInternal() {
+    disposeInternal(nativeHandle_);
+  }
+  private native void disposeInternal(long handle);
+
+  /**
+   * The static default RocksEnv.  The ownership of its native handle
+   * belongs to rocksdb c++ and is not able to be released on the Java
+   * side.
+   */
+  static RocksEnv default_env_;
+}
@@ -118,15 +118,13 @@ public class RocksIterator extends RocksObject {
  /**
   * Deletes underlying C++ iterator pointer.
   */
-  @Override public synchronized void dispose() {
-    if(isInitialized()) {
-      dispose(nativeHandle_);
-      nativeHandle_ = 0;
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

  private native boolean isValid0(long handle);
-  private native void dispose(long handle);
+  private native void disposeInternal(long handle);
  private native void seekToFirst0(long handle);
  private native void seekToLast0(long handle);
  private native void next0(long handle);
@@ -6,30 +6,112 @@
 package org.rocksdb;

 /**
- * RocksObject is the base-class of all RocksDB related class that has
- * a pointer to some c++ rocksdb object.  Although RocksObject
- * will release its c++ resource on its finalize() once it has been
- * garbage-collected, it is suggested to call dispose() manually to
- * release its c++ resource once an instance of RocksObject is no
- * longer used.
+ * RocksObject is the base-class of all RocksDB classes that has a pointer to
+ * some c++ rocksdb object.
+ *
+ * RocksObject has dispose() function, which releases its associated c++ resource.
+ * This function can be either called manually, or being called automatically
+ * during the regular Java GC process.  However, since Java may wrongly assume a
+ * RocksObject only contains a long member variable and think it is small in size,
+ * Java may give RocksObject low priority in the GC process.  For this, it is
+ * suggested to call dispose() manually.  However, it is safe to let RocksObject go
+ * out-of-scope without manually calling dispose() as dispose() will be called
+ * in the finalizer during the regular GC process.
 */
 public abstract class RocksObject {
  protected RocksObject() {
    nativeHandle_ = 0;
+    owningHandle_ = true;
  }

  /**
-   * Release the c++ object pointed by the native handle.
+   * Release the c++ object manually pointed by the native handle.
+   *
+   * Note that dispose() will also be called during the GC process
+   * if it was not called before its RocksObject went out-of-scope.
+   * However, since Java may wrongly wrongly assume those objects are
+   * small in that they seems to only hold a long variable. As a result,
+   * they might have low priority in the GC process.  To prevent this,
+   * it is suggested to call dispose() manually.
+   *
+   * Note that once an instance of RocksObject has been disposed,
+   * calling its function will lead undefined behavior.
   */
-  public abstract void dispose();
+  public final synchronized void dispose() {
+    if (isOwningNativeHandle() && isInitialized()) {
+      disposeInternal();
+    }
+    nativeHandle_ = 0;
+    disOwnNativeHandle();
+  }

+  /**
+   * The helper function of dispose() which all subclasses of RocksObject
+   * must implement to release their associated C++ resource.
+   */
+  protected abstract void disposeInternal();
+
+  /**
+   * Revoke ownership of the native object.
+   *
+   * This will prevent the object from attempting to delete the underlying
+   * native object in its finalizer. This must be used when another object
+   * takes over ownership of the native object or both will attempt to delete
+   * the underlying object when garbage collected.
+   *
+   * When disOwnNativeHandle() is called, dispose() will simply set nativeHandle_
+   * to 0 without releasing its associated C++ resource.  As a result,
+   * incorrectly use this function may cause memory leak, and this function call
+   * will not affect the return value of isInitialized().
+   *
+   * @see dispose()
+   * @see isInitialized()
+   */
+  protected void disOwnNativeHandle() {
+    owningHandle_ = false;
+  }
+
+  /**
+   * Returns true if the current RocksObject is responsable to release its
+   * native handle.
+   *
+   * @return true if the current RocksObject is responsible to release its
+   *   native handle.
+   *
+   * @see disOwnNativeHandle()
+   * @see dispose()
+   */
+  protected boolean isOwningNativeHandle() {
+    return owningHandle_;
+  }
+
+  /**
+   * Returns true if the associated native handle has been initialized.
+   *
+   * @return true if the associated native handle has been initialized.
+   *
+   * @see dispose()
+   */
  protected boolean isInitialized() {
    return (nativeHandle_ != 0);
  }

+  /**
+   * Simply calls dispose() and release its c++ resource if it has not
+   * yet released.
+   */
  @Override protected void finalize() {
    dispose();
  }

+  /**
+   * A long variable holding c++ pointer pointing to some RocksDB C++ object.
+   */
  protected long nativeHandle_;
+
+  /**
+   * A flag indicating whether the current RocksObject is responsible to
+   * release the c++ object stored in its nativeHandle_.
+   */
+  private boolean owningHandle_;
 }
@@ -0,0 +1,93 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+import java.util.concurrent.ArrayBlockingQueue;
+import java.util.concurrent.ThreadPoolExecutor;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicBoolean;
+
+/**
+ * Helper class to collect DB statistics periodically at a period specified in
+ * constructor. Callback function (provided in constructor) is called with
+ * every statistics collection.
+ *
+ * Caller should call start() to start statistics collection. Shutdown() should
+ * be called to stop stats collection and should be called before statistics (
+ * provided in constructor) reference has been disposed.
+ */
+public class StatisticsCollector {
+  private final Statistics _statistics;
+  private final ThreadPoolExecutor _threadPoolExecutor;
+  private final int _statsCollectionInterval;
+  private final StatisticsCollectorCallback _statsCallback;
+  private volatile boolean _isRunning = true;
+
+  /**
+   * Constructor for statistics collector.
+   * @param statistics Reference of DB statistics.
+   * @param statsCollectionIntervalInMilliSeconds Statistics collection time 
+   *        period (specified in milliseconds) 
+   * @param statsCallback Reference of statistics callback interface.
+   */
+  public StatisticsCollector(Statistics statistics,
+      int statsCollectionIntervalInMilliSeconds,
+      StatisticsCollectorCallback statsCallback) {
+    _statistics = statistics;
+    _statsCollectionInterval = statsCollectionIntervalInMilliSeconds;
+    _statsCallback = statsCallback;
+
+    _threadPoolExecutor = new ThreadPoolExecutor(1, 1, 0, TimeUnit.SECONDS,
+        new ArrayBlockingQueue<Runnable>(1));
+  }
+
+  public void start() {
+    _threadPoolExecutor.submit(collectStatistics());
+  }
+
+  public void shutDown() throws InterruptedException {
+    _isRunning = false;
+
+    _threadPoolExecutor.shutdown();
+    // Wait for collectStatistics runnable to finish so that disposal of
+    // statistics does not cause any exceptions to be thrown.
+    _threadPoolExecutor.awaitTermination(Integer.MAX_VALUE, TimeUnit.SECONDS);
+  }
+
+  private Runnable collectStatistics() {
+    return new Runnable() {
+
+      @Override
+      public void run() {
+        while (_isRunning) {
+          try {
+            // Collect ticker data
+            for(TickerType ticker : TickerType.values()) {
+              long tickerValue = _statistics.getTickerCount(ticker);
+              _statsCallback.tickerCallback(ticker, tickerValue);
+            }
+
+            // Collect histogram data
+            for(HistogramType histogramType : HistogramType.values()) {
+              HistogramData histogramData =
+                  _statistics.geHistogramData(histogramType);
+              _statsCallback.histogramCallback(histogramType, histogramData);
+            }
+
+            Thread.sleep(_statsCollectionInterval);
+          }
+          catch (InterruptedException e) {
+            Thread.currentThread().interrupt();
+            throw new RuntimeException("Thread got interrupted!", e);
+          }
+          catch (Exception e) {
+            throw new RuntimeException("Error while calculating statistics", e);
+          }
+        }
+      }
+    };
+  }
+}
@@ -0,0 +1,34 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb;
+
+/**
+ * Callback interface provided to StatisticsCollector.
+ * 
+ * Thread safety:
+ * StatisticsCollector doesn't make any guarantees about thread safety. 
+ * If the same reference of StatisticsCollectorCallback is passed to multiple
+ * StatisticsCollector references, then its the responsibility of the 
+ * user to make StatisticsCollectorCallback' implementation thread-safe.
+ * 
+ * @param tickerType
+ * @param tickerCount
+*/
+public interface StatisticsCollectorCallback {
+  /**
+   * Callback function to get ticker values.
+   * @param tickerType Ticker type.
+   * @param tickerCount Value of ticker type.
+  */
+  void tickerCallback(TickerType tickerType, long tickerCount);
+
+  /**
+   * Callback function to get histogram values.
+   * @param histType Histogram type.
+   * @param histData Histogram data.
+  */
+  void histogramCallback(HistogramType histType, HistogramData histData);
+}
@@ -86,10 +86,9 @@ public class WriteBatch extends RocksObject {
  /**
   * Delete the c++ side pointer.
   */
-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose0();
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

  private native void newWriteBatch(int reserved_bytes);
@@ -99,7 +98,7 @@ public class WriteBatch extends RocksObject {
                            byte[] value, int valueLen);
  private native void remove(byte[] key, int keyLen);
  private native void putLogData(byte[] blob, int blobLen);
-  private native void dispose0();
+  private native void disposeInternal(long handle);
 }

 /**
@@ -17,10 +17,9 @@ public class WriteOptions extends RocksObject {
    newWriteOptions();
  }

-  @Override public synchronized void dispose() {
-    if (isInitialized()) {
-      dispose0(nativeHandle_);
-    }
+  @Override protected void disposeInternal() {
+    assert(isInitialized());
+    disposeInternal(nativeHandle_);
  }

  /**
@@ -96,5 +95,5 @@ public class WriteOptions extends RocksObject {
  private native boolean sync(long handle);
  private native void setDisableWAL(long handle, boolean flag);
  private native boolean disableWAL(long handle);
-  private native void dispose0(long handle);
+  private native void disposeInternal(long handle);
 }
@@ -22,6 +22,7 @@
 package org.rocksdb.benchmark;

 import java.lang.Runnable;
+import java.lang.Math;
 import java.io.File;
 import java.nio.ByteBuffer;
 import java.util.Collection;
@@ -240,7 +241,8 @@ public class DbBenchmark {
        if (entriesPerBatch_ == 1) {
          for (long i = 0; i < numEntries_; ++i) {
            getKey(key, i, keyRange_);
-            db_.put(writeOpt_, key, DbBenchmark.this.gen_.generate(valueSize_));
+            DbBenchmark.this.gen_.generate(value);
+            db_.put(writeOpt_, key, value);
            stats_.finishedSingleOp(keySize_ + valueSize_);
            writeRateControl(i);
            if (isFinished()) {
@@ -252,7 +254,8 @@ public class DbBenchmark {
            WriteBatch batch = new WriteBatch();
            for (long j = 0; j < entriesPerBatch_; j++) {
              getKey(key, i + j, keyRange_);
-              batch.put(key, DbBenchmark.this.gen_.generate(valueSize_));
+              DbBenchmark.this.gen_.generate(value);
+              db_.put(writeOpt_, key, value);
              stats_.finishedSingleOp(keySize_ + valueSize_);
            }
            db_.write(writeOpt_, batch);
@@ -460,7 +463,7 @@ public class DbBenchmark {
      if (compressionType_.equals("snappy")) {
        System.loadLibrary("snappy");
      } else if (compressionType_.equals("zlib")) {
-        System.loadLibrary("zlib");
+        System.loadLibrary("z");
      } else if (compressionType_.equals("bzip2")) {
        System.loadLibrary("bzip2");
      } else if (compressionType_.equals("lz4")) {
@@ -473,7 +476,6 @@ public class DbBenchmark {
                        "No compression is used.%n",
          compressionType_, e.toString());
      compressionType_ = "none";
-      compressionRatio_ = 1.0;
    }
    gen_ = new RandomGenerator(randSeed_, compressionRatio_);
  }
@@ -526,6 +528,8 @@ public class DbBenchmark {
        (Integer)flags_.get(Flag.max_write_buffer_number));
    options.setMaxBackgroundCompactions(
        (Integer)flags_.get(Flag.max_background_compactions));
+    options.getEnv().setBackgroundThreads(
+        (Integer)flags_.get(Flag.max_background_compactions));
    options.setMaxBackgroundFlushes(
        (Integer)flags_.get(Flag.max_background_flushes));
    options.setCacheSize(
@@ -534,8 +538,6 @@ public class DbBenchmark {
        (Long)flags_.get(Flag.block_size));
    options.setMaxOpenFiles(
        (Integer)flags_.get(Flag.open_files));
-    options.setCreateIfMissing(
-        !(Boolean)flags_.get(Flag.use_existing_db));
    options.setTableCacheRemoveScanCountLimit(
        (Integer)flags_.get(Flag.cache_remove_scan_count_limit));
    options.setDisableDataSync(
@@ -939,7 +941,7 @@ public class DbBenchmark {
        "\tflag and also specify a benchmark that wants a fresh database,\n" +
        "\tthat benchmark will fail.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    num(1000000,
@@ -1028,7 +1030,7 @@ public class DbBenchmark {
    use_plain_table(false,
        "Use plain-table sst format.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    cache_size(-1L,
@@ -1085,7 +1087,7 @@ public class DbBenchmark {
    },
    histogram(false,"Print histogram of operation timings.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    min_write_buffer_number_to_merge(
@@ -1203,12 +1205,12 @@ public class DbBenchmark {
    verify_checksum(false,"Verify checksum for every block read\n" +
        "\tfrom storage.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    statistics(false,"Database statistics.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    writes(-1,"Number of write operations to do. If negative, do\n" +
@@ -1219,23 +1221,23 @@ public class DbBenchmark {
    },
    sync(false,"Sync all writes to disk.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    disable_data_sync(false,"If true, do not wait until data is\n" +
        "\tsynced to disk.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    use_fsync(false,"If true, issue fsync instead of fdatasync.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    disable_wal(false,"If true, do not write WAL for write.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    wal_dir("", "If not empty, use the given dir for WAL.") {
@@ -1312,7 +1314,7 @@ public class DbBenchmark {
    disable_seek_compaction(false,"Option to disable compaction\n" +
        "\ttriggered by read.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    delete_obsolete_files_period_micros(0,"Option to delete\n" +
@@ -1393,12 +1395,12 @@ public class DbBenchmark {
    },
    readonly(false,"Run read only benchmarks.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    disable_auto_compactions(false,"Do not auto trigger compactions.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    source_compaction_factor(1,"Cap the size of data in level-K for\n" +
@@ -1423,26 +1425,26 @@ public class DbBenchmark {
    bufferedio(rocksdb::EnvOptions().use_os_buffer,
        "Allow buffered io using OS buffers.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    */
    mmap_read(false,
        "Allow reads to occur via mmap-ing files.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    mmap_write(false,
        "Allow writes to occur via mmap-ing files.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    advise_random_on_open(defaultOptions_.adviseRandomOnOpen(),
        "Advise random access on table file open.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    compaction_fadvice("NORMAL",
@@ -1454,13 +1456,13 @@ public class DbBenchmark {
    use_tailing_iterator(false,
        "Use tailing iterator to access a series of keys instead of get.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    use_adaptive_mutex(defaultOptions_.useAdaptiveMutex(),
        "Use adaptive mutex.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    bytes_per_sync(defaultOptions_.bytesPerSync(),
@@ -1474,7 +1476,7 @@ public class DbBenchmark {
    filter_deletes(false," On true, deletes use bloom-filter and drop\n" +
        "\tthe delete if key not present.") {
      @Override public Object parseValue(String value) {
-        return Boolean.parseBoolean(value);
+        return parseBoolean(value);
      }
    },
    max_successive_merges(0,"Maximum number of successive merge\n" +
@@ -1495,8 +1497,6 @@ public class DbBenchmark {
      desc_ = desc;
    }

-    protected abstract Object parseValue(String value);
-
    public Object getDefaultValue() {
      return defaultValue_;
    }
@@ -1505,6 +1505,17 @@ public class DbBenchmark {
      return desc_;
    }

+    public boolean parseBoolean(String value) {
+      if (value.equals("1")) {
+        return true;
+      } else if (value.equals("0")) {
+        return false;
+      }
+      return Boolean.parseBoolean(value);
+    }
+
+    protected abstract Object parseValue(String value);
+
    private final Object defaultValue_;
    private final String desc_;
  }
@@ -1513,24 +1524,54 @@ public class DbBenchmark {
    private final byte[] data_;
    private int dataLength_;
    private int position_;
+    private double compressionRatio_;
    Random rand_;

    private RandomGenerator(long seed, double compressionRatio) {
      // We use a limited amount of data over and over again and ensure
      // that it is larger than the compression window (32KB), and also
+      byte[] value = new byte[100];
      // large enough to serve all typical value sizes we want to write.
      rand_ = new Random(seed);
-      dataLength_ = 1048576 + 100;
+      dataLength_ = value.length * 10000;
      data_ = new byte[dataLength_];
-      // TODO(yhchiang): mimic test::CompressibleString?
-      for (int i = 0; i < dataLength_; ++i) {
-        data_[i] = (byte) (' ' + rand_.nextInt(95));
+      compressionRatio_ = compressionRatio;
+      int pos = 0;
+      while (pos < dataLength_) {
+        compressibleBytes(value);
+        System.arraycopy(value, 0, data_, pos,
+                         Math.min(value.length, dataLength_ - pos));
+        pos += value.length;
      }
    }

-    private byte[] generate(int length) {
-      position_ = rand_.nextInt(data_.length - length);
-      return Arrays.copyOfRange(data_, position_, position_ + length);
+    private void compressibleBytes(byte[] value) {
+      int baseLength = value.length;
+      if (compressionRatio_ < 1.0d) {
+        baseLength = (int) (compressionRatio_ * value.length + 0.5);
+      }
+      if (baseLength <= 0) {
+        baseLength = 1;
+      }
+      int pos;
+      for (pos = 0; pos < baseLength; ++pos) {
+        value[pos] = (byte) (' ' + rand_.nextInt(95));  // ' ' .. '~'
+      }
+      while (pos < value.length) {
+        System.arraycopy(value, 0, value, pos,
+                         Math.min(baseLength, value.length - pos));
+        pos += baseLength;
+      }
+    }
+
+    private void generate(byte[] value) {
+      if (position_ + value.length > data_.length) {
+        position_ = 0;
+        assert(value.length <= data_.length);
+      }
+      position_ += value.length;
+      System.arraycopy(data_, position_ - value.length,
+                       value, 0, value.length);
    }
  }

@@ -18,15 +18,37 @@ public class BackupableDBTest {
    Options opt = new Options();
    opt.setCreateIfMissing(true);

-    BackupableDBOptions bopt = new BackupableDBOptions(backup_path);
+    BackupableDBOptions bopt = new BackupableDBOptions(backup_path, false,
+        true, false, true, 0, 0);
    BackupableDB bdb = null;

    try {
      bdb = BackupableDB.open(opt, bopt, db_path);
-      bdb.put("hello".getBytes(), "BackupableDB".getBytes());
+
+      bdb.put("abc".getBytes(), "def".getBytes());
+      bdb.put("ghi".getBytes(), "jkl".getBytes());
      bdb.createNewBackup(true);
-      byte[] value = bdb.get("hello".getBytes());
-      assert(new String(value).equals("BackupableDB"));
+
+      // delete record after backup
+      bdb.remove("abc".getBytes());
+      byte[] value = bdb.get("abc".getBytes());
+      assert(value == null);
+      bdb.close();
+
+      // restore from backup
+      RestoreOptions ropt = new RestoreOptions(false);
+      RestoreBackupableDB rdb = new RestoreBackupableDB(bopt);
+      rdb.restoreDBFromLatestBackup(db_path, db_path,
+          ropt);
+      rdb.dispose();
+      ropt.dispose();
+
+      // verify that backed up data contains deleted record
+      bdb = BackupableDB.open(opt, bopt, db_path);
+      value = bdb.get("abc".getBytes());
+      assert(new String(value).equals("def"));
+
+      System.out.println("Backup and restore test passed");
    } catch (RocksDBException e) {
      System.err.format("[ERROR]: %s%n", e);
      e.printStackTrace();
@@ -0,0 +1,40 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb.test;
+
+import org.rocksdb.*;
+
+public class StatisticsCollectorTest {
+  static final String db_path = "/tmp/backupablejni_db";
+  static {
+    RocksDB.loadLibrary();
+  }
+
+  public static void main(String[] args) 
+      throws InterruptedException, RocksDBException {
+    Options opt = new Options().createStatistics().setCreateIfMissing(true);
+    Statistics stats = opt.statisticsPtr();
+
+    RocksDB db = RocksDB.open(db_path);
+
+    StatsCallbackMock callback = new StatsCallbackMock();
+    StatisticsCollector statsCollector = new StatisticsCollector(stats, 100,
+        callback);
+    statsCollector.start();
+
+    Thread.sleep(1000);
+
+    assert(callback.tickerCallbackCount > 0);
+    assert(callback.histCallbackCount > 0);
+
+    statsCollector.shutDown();
+
+    db.close();
+    opt.dispose();
+
+    System.out.println("Stats collector test passed.!");
+  }
+}
@@ -0,0 +1,22 @@
+// Copyright (c) 2014, Facebook, Inc.  All rights reserved.
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree. An additional grant
+// of patent rights can be found in the PATENTS file in the same directory.
+
+package org.rocksdb.test;
+
+import org.rocksdb.*;
+
+public class StatsCallbackMock implements StatisticsCollectorCallback {
+  public int tickerCallbackCount = 0;
+  public int histCallbackCount = 0;
+
+  public void tickerCallback(TickerType tickerType, long tickerCount) {
+    tickerCallbackCount++;
+  }
+
+  public void histogramCallback(HistogramType histType,
+      HistogramData histData) {
+    histCallbackCount++;
+  }
+}
@@ -4,7 +4,8 @@
 // of patent rights can be found in the PATENTS file in the same directory.
 //
 // This file implements the "bridge" between Java and C++ and enables
-// calling c++ rocksdb::DB methods from Java side.
+// calling c++ rocksdb::BackupableDB and rocksdb::BackupableDBOptions methods
+// from Java side.

 #include <stdio.h>
 #include <stdlib.h>
@@ -51,9 +52,18 @@ void Java_org_rocksdb_BackupableDB_createNewBackup(
 * Signature: (Ljava/lang/String;)V
 */
 void Java_org_rocksdb_BackupableDBOptions_newBackupableDBOptions(
-    JNIEnv* env, jobject jobj, jstring jpath) {
+    JNIEnv* env, jobject jobj, jstring jpath, jboolean jshare_table_files,
+    jboolean jsync, jboolean jdestroy_old_data, jboolean jbackup_log_files,
+    jlong jbackup_rate_limit, jlong jrestore_rate_limit) {
+  jbackup_rate_limit = (jbackup_rate_limit <= 0) ? 0 : jbackup_rate_limit;
+  jrestore_rate_limit = (jrestore_rate_limit <= 0) ? 0 : jrestore_rate_limit;
+
  const char* cpath = env->GetStringUTFChars(jpath, 0);
-  auto bopt = new rocksdb::BackupableDBOptions(cpath);
+
+  auto bopt = new rocksdb::BackupableDBOptions(cpath, nullptr,
+      jshare_table_files, nullptr, jsync, jdestroy_old_data, jbackup_log_files,
+      jbackup_rate_limit, jrestore_rate_limit);
+
  env->ReleaseStringUTFChars(jpath, cpath);

  rocksdb::BackupableDBOptionsJni::setHandle(env, jobj, bopt);
@@ -72,10 +82,10 @@ jstring Java_org_rocksdb_BackupableDBOptions_backupDir(

 /*
 * Class:     org_rocksdb_BackupableDBOptions
- * Method:    dispose
+ * Method:    disposeInternal
 * Signature: (J)V
 */
-void Java_org_rocksdb_BackupableDBOptions_dispose(
+void Java_org_rocksdb_BackupableDBOptions_disposeInternal(
    JNIEnv* env, jobject jopt, jlong jhandle) {
  auto bopt = reinterpret_cast<rocksdb::BackupableDBOptions*>(jhandle);
  assert(bopt);