Prevent erroneous socket closure

Tweak hit rate metric to get delta instead of aggregate

build.yaml

@@ -21,13 +21,25 @@ machamp:
         parent: make-build
         # https://github.sc-corp.net/Snapchat/img/tree/master/keydb/ubuntu-20-04
         builder_image: us.gcr.io/snapchat-build-artifacts/prod/snapchat/img/keydb/keydb-ubuntu-20-04@sha256:cf869a3f5d1de1e1d976bb906689c37b7031938eb68661b844a38c532f27248c
-        command: ./runtest --clients 4 --verbose 
+        command: ./runtest --clients 4 --verbose --tls --skipunit unit/memefficiency
+      tests-with-flash:
+        type: cmd
+        parent: make-build
+        # https://github.sc-corp.net/Snapchat/img/tree/master/keydb/ubuntu-20-04
+        builder_image: us.gcr.io/snapchat-build-artifacts/prod/snapchat/img/keydb/keydb-ubuntu-20-04@sha256:cf869a3f5d1de1e1d976bb906689c37b7031938eb68661b844a38c532f27248c
+        command: ./runtest --verbose --skipunit unit/memefficiency --flash
       cluster-test:
         type: cmd
         parent: make-build
         # https://github.sc-corp.net/Snapchat/img/tree/master/keydb/ubuntu-20-04
         builder_image: us.gcr.io/snapchat-build-artifacts/prod/snapchat/img/keydb/keydb-ubuntu-20-04@sha256:cf869a3f5d1de1e1d976bb906689c37b7031938eb68661b844a38c532f27248c
-        command: ./runtest-cluster 
+        command: ./runtest-cluster --tls
+      cluster-test-flash:
+        type: cmd
+        parent: make-build
+        # https://github.sc-corp.net/Snapchat/img/tree/master/keydb/ubuntu-20-04
+        builder_image: us.gcr.io/snapchat-build-artifacts/prod/snapchat/img/keydb/keydb-ubuntu-20-04@sha256:cf869a3f5d1de1e1d976bb906689c37b7031938eb68661b844a38c532f27248c
+        command: ./runtest-cluster --tls --flash
       sentinel-test:
         type: cmd
         parent: make-build
@@ -59,7 +71,7 @@ machamp:
       # to ensure a clearer docker build env
       code-checkout:
         type: cmd
-        command: echo checkout
+        command: echo checkout && git submodule init && git submodule update
         # default machamp builder image does not work for multi arch
         builder_image: us.gcr.io/snapchat-build-artifacts/prod/snapchat/img/ubuntu/ubuntu-23-04@sha256:bd43177a80e6ce1c3583e8ea959b88a9081c0f56b765ec9c5a157c27a637c23b
       docker:

ci.yaml

@@ -26,6 +26,16 @@ on:
           # references a build defined in build.yaml
           build_name: keydb-docker-build
           arch_types: ["amd64", "arm64"]
+  # Doc: go/cool-guide
+  cool:
+    workflows:
+      - workflow_type: backend_workflow
+        build_name: keydb-build
+        arch_types: ["amd64", "arm64"]
+      - workflow_type: backend_workflow
+        # references a build defined in build.yaml
+        build_name: keydb-docker-build
+        arch_types: ["amd64", "arm64"]
 
 # below defines which branch is release branch / release tag
 machamp:

machamp_scripts/Dockerfile

@@ -24,7 +24,7 @@ RUN set -eux; \
         gosu nobody true
 # build KeyDB
 ARG MAKE_JOBS=""
-ARG ENABLE_FLASH=""
+ARG ENABLE_FLASH="yes"
 COPY . /tmp/keydb-internal
 RUN set -eux; \
         cd /tmp/keydb-internal; \
@@ -73,7 +73,25 @@ RUN set -eux; \
                | xargs -r apt-mark manual \
         ; \
         apt-get purge -y --auto-remove -o APT::AutoRemove::RecommendsImportant=false; \
-        rm -rf /var/lib/apt/lists/*; \
+        rm -rf /var/lib/apt/lists/*
+# build RedisBloom
+RUN set -eux; \
+        savedAptMark="$(apt-mark showmanual)"; \
+        apt-get update; \
+        DEBIAN_FRONTEND=noninteractive apt-get -o Dpkg::Options::="--force-confnew" install -qqy --no-install-recommends \
+                ca-certificates \
+                wget \
+                make \
+                pkg-config \
+                build-essential \
+                git; \
+        git clone --recursive https://github.com/RedisBloom/RedisBloom.git; \
+        cd RedisBloom; \
+        ./sbin/setup; \
+        bash -l; \
+        make; \
+        redisBloomPath=`make run -n | awk '{print $NF}'`; \
+        cp $redisBloomPath /usr/local/lib/
 # create working directories and organize files
 RUN \
         mkdir /data && chown keydb:keydb /data; \

pkg/docker/Dockerfile

@@ -53,7 +53,7 @@ RUN set -eux; \
         grep -E '^ *createBoolConfig[(]"protected-mode",.*, *1 *,.*[)],$' ./src/config.cpp; \
         sed -ri 's!^( *createBoolConfig[(]"protected-mode",.*, *)1( *,.*[)],)$!\10\2!' ./src/config.cpp; \
         grep -E '^ *createBoolConfig[(]"protected-mode",.*, *0 *,.*[)],$' ./src/config.cpp; \
-        make -j$(nproc) BUILD_TLS=yes; \
+        make -j$(nproc) BUILD_TLS=yes ENABLE_FLASH=yes; \
         cd src; \
         strip keydb-cli keydb-benchmark keydb-check-rdb keydb-check-aof keydb-diagnostic-tool keydb-sentinel keydb-server; \
         mv keydb-server keydb-cli keydb-benchmark keydb-check-rdb keydb-check-aof keydb-diagnostic-tool keydb-sentinel /usr/local/bin/; \
@@ -83,6 +83,7 @@ RUN \
         sed -i 's/^\(logfile .*\)$/# \1/' /etc/keydb/keydb.conf; \
         sed -i 's/protected-mode yes/protected-mode no/g' /etc/keydb/keydb.conf; \
         sed -i 's/^\(bind .*\)$/# \1/' /etc/keydb/keydb.conf; \
+        cd /usr/local/bin; \
         ln -s keydb-cli redis-cli; \
         cd /etc/keydb; \
         ln -s keydb.conf redis.conf; \

pkg/docker/Dockerfile_Alpine

@@ -5,7 +5,7 @@ RUN mkdir -p /etc/keydb
 ARG BRANCH
 RUN set -eux; \
         \
-        apk add --no-cache su-exec; \
+        apk add --no-cache su-exec tini; \
         apk add --no-cache --virtual .build-deps \
                 coreutils \
                 gcc \
@@ -32,7 +32,7 @@ RUN set -eux; \
         grep -E '^ *createBoolConfig[(]"protected-mode",.*, *1 *,.*[)],$' ./src/config.cpp; \
         sed -ri 's!^( *createBoolConfig[(]"protected-mode",.*, *)1( *,.*[)],)$!\10\2!' ./src/config.cpp; \
         grep -E '^ *createBoolConfig[(]"protected-mode",.*, *0 *,.*[)],$' ./src/config.cpp; \
-        make -j$(nproc) BUILD_TLS=yes; \
+        make -j$(nproc) BUILD_TLS=yes ENABLE_FLASH=yes; \
         cd src; \
         strip keydb-cli keydb-benchmark keydb-check-rdb keydb-check-aof keydb-diagnostic-tool keydb-sentinel keydb-server; \
         mv keydb-server keydb-cli keydb-benchmark keydb-check-rdb keydb-check-aof keydb-diagnostic-tool keydb-sentinel /usr/local/bin/; \
@@ -54,6 +54,7 @@ RUN set -eux; \
         sed -i 's/^\(logfile .*\)$/# \1/' /etc/keydb/keydb.conf; \
         sed -i 's/protected-mode yes/protected-mode no/g' /etc/keydb/keydb.conf; \
         sed -i 's/^\(bind .*\)$/# \1/' /etc/keydb/keydb.conf; \
+        cd /usr/local/bin; \
         ln -s keydb-cli redis-cli; \
         cd /etc/keydb; \
         ln -s keydb.conf redis.conf; \
@@ -76,6 +77,6 @@ RUN set -eux; \
         chmod +x /usr/local/bin/docker-entrypoint.sh
 VOLUME /data
 WORKDIR /data
-ENTRYPOINT ["docker-entrypoint.sh"]
+ENTRYPOINT ["tini", "--", "docker-entrypoint.sh"]
 EXPOSE 6379
 CMD ["keydb-server", "/etc/keydb/keydb.conf"]

src/IStorage.h

@@ -1,10 +1,25 @@
 #pragma once
 #include <functional>
+#include <unordered_set>
 #include "sds.h"
+#include "ae.h"
 #include <string>
 
 #define METADATA_DB_IDENTIFIER "c299fde0-6d42-4ec4-b939-34f680ffe39f"
 
+struct StorageToken {
+    enum class TokenType {
+        SingleRead,
+        SingleWrite,
+        Delete,
+        BatchWrite,
+    };
+    TokenType type;
+    std::unordered_set<struct client *> setc;
+    struct redisDbPersistentData *db;
+    virtual ~StorageToken() {}
+};
+
 class IStorageFactory
 {
 public:
@@ -36,6 +51,12 @@ public:
     virtual bool enumerate_hashslot(callback fn, unsigned int hashslot) const = 0;
     virtual size_t count() const = 0;
 
+    virtual StorageToken *begin_retrieve(struct aeEventLoop *, aePostFunctionTokenProc, sds *, size_t) {return nullptr;};
+    virtual void complete_retrieve(StorageToken * /*tok*/, callbackSingle /*fn*/) {};
+
+    virtual StorageToken* begin_endWriteBatch(struct aeEventLoop *, aePostFunctionTokenProc*) { return nullptr; } // NOP
+    virtual void complete_endWriteBatch(StorageToken * /*tok*/) {};
+
     virtual void bulkInsert(char **rgkeys, size_t *rgcbkeys, char **rgvals, size_t *rgcbvals, size_t celem) {
         beginWriteBatch();
         for (size_t ielem = 0; ielem < celem; ++ielem) {
@@ -44,7 +65,7 @@ public:
         endWriteBatch();
     }
 
-    virtual std::vector<std::string> getExpirationCandidates(unsigned int count, long long time) = 0;
+    virtual std::vector<std::string> getExpirationCandidates(unsigned int count) = 0;
     virtual std::vector<std::string> getEvictionCandidates(unsigned int count) = 0;
     virtual void setExpire(const char *key, size_t cchKey, long long expire) = 0;
     virtual void removeExpire(const char *key, size_t cchKey, long long expire) = 0;

src/Makefile

@@ -73,10 +73,10 @@ ifneq ($(strip $(SANITIZE)),)
 endif
 
 ifeq ($(ENABLE_FLASH),yes)
+	FINAL_LIBS+= ../deps/rocksdb/librocksdb.a
 	FINAL_LIBS+= -lz -lcrypto -lbz2 -lzstd -llz4 -lsnappy
 	CXXFLAGS+= -I../deps/rocksdb/include/ -DENABLE_ROCKSDB
 	STORAGE_OBJ+= storage/rocksdb.o storage/rocksdbfactory.o
-	FINAL_LIBS+= ../deps/rocksdb/librocksdb.a 
 	DEPENDENCY_TARGETS+= rocksdb
 endif
 

src/SnapshotPayloadParseState.cpp

@@ -136,14 +136,33 @@ void SnapshotPayloadParseState::flushQueuedKeys() {
     int idb = current_database;
     serverAssert(vecqueuedKeys.size() == vecqueuedVals.size());
     auto sizePrev = vecqueuedKeys.size();
-    (*insertsInFlight)++;
-    std::weak_ptr<std::atomic<int>> insertsInFlightTmp = insertsInFlight; // C++ GRRRRRRRRRRRRRRRR, we don't want to capute "this" because that's dangerous
     if (current_database < cserver.dbnum) {
-        g_pserver->asyncworkqueue->AddWorkFunction([idb, vecqueuedKeys = std::move(this->vecqueuedKeys), vecqueuedKeysCb = std::move(this->vecqueuedKeysCb), vecqueuedVals = std::move(this->vecqueuedVals), vecqueuedValsCb = std::move(this->vecqueuedValsCb), insertsInFlightTmp, pallocator = m_spallocator.release()]() mutable {
-            g_pserver->db[idb]->bulkDirectStorageInsert(vecqueuedKeys.data(), vecqueuedKeysCb.data(), vecqueuedVals.data(), vecqueuedValsCb.data(), vecqueuedKeys.size());
-            (*(insertsInFlightTmp.lock()))--;
-            delete pallocator;
-        });
+        if (g_pserver->m_pstorageFactory) {
+            (*insertsInFlight)++;
+            std::weak_ptr<std::atomic<int>> insertsInFlightTmp = insertsInFlight; // C++ GRRRRRRRRRRRRRRRR, we don't want to capute "this" because that's dangerous
+            g_pserver->asyncworkqueue->AddWorkFunction([idb, vecqueuedKeys = std::move(this->vecqueuedKeys), vecqueuedKeysCb = std::move(this->vecqueuedKeysCb), vecqueuedVals = std::move(this->vecqueuedVals), vecqueuedValsCb = std::move(this->vecqueuedValsCb), insertsInFlightTmp, pallocator = m_spallocator.release()]() mutable {
+                g_pserver->db[idb]->bulkDirectStorageInsert(vecqueuedKeys.data(), vecqueuedKeysCb.data(), vecqueuedVals.data(), vecqueuedValsCb.data(), vecqueuedKeys.size());
+                (*(insertsInFlightTmp.lock()))--;
+                delete pallocator;
+            });
+        } else {
+            for (size_t ival = 0; ival < vecqueuedKeys.size(); ++ival) {
+                size_t offset = 0;
+                auto spexpire = deserializeExpire(vecqueuedVals[ival], vecqueuedValsCb[ival], &offset);    
+                auto o = deserializeStoredObject(vecqueuedVals[ival] + offset, vecqueuedValsCb[ival] - offset);
+                sds sdsKey = sdsnewlen(vecqueuedKeys[ival], -static_cast<ssize_t>(vecqueuedKeysCb[ival]));
+                if (dbMerge(g_pserver->db[idb], sdsKey, o, false)) {
+                    if (spexpire != nullptr)
+                        g_pserver->db[idb]->setExpire(sdsKey, std::move(*spexpire));
+                } else {
+                    sdsfree(sdsKey);
+                }
+            }
+            vecqueuedKeys.clear();
+            vecqueuedKeysCb.clear();
+            vecqueuedVals.clear();
+            vecqueuedValsCb.clear();
+        }
     } else {
         // else drop the data
         vecqueuedKeys.clear();
@@ -214,11 +233,6 @@ void SnapshotPayloadParseState::trimState() {
     
     if (stackParse.empty()) {
         flushQueuedKeys();
-        while (*insertsInFlight > 0) {
-            // TODO: ProcessEventsWhileBlocked
-            aeReleaseLock();
-            aeAcquireLock();
-        }
     }
 }
 

src/SnapshotPayloadParseState.h

@@ -63,4 +63,5 @@ public:
     void pushValue(const char *rgch, long long cch);
     void pushValue(long long value);
     bool shouldThrottle() const { return *insertsInFlight > (cserver.cthreads*4); }
+    bool hasIOInFlight() const { return *insertsInFlight > 0; }
 };

src/StorageCache.cpp

@@ -208,6 +208,30 @@ void StorageCache::retrieve(sds key, IStorage::callbackSingle fn) const
     m_spstorage->retrieve(key, sdslen(key), fn);
 }
 
+StorageToken *StorageCache::begin_retrieve(struct aeEventLoop *el, aePostFunctionTokenProc proc, sds *rgkey, size_t ckey) {
+    std::unique_lock<fastlock> ul(m_lock);
+    if (m_pdict != nullptr)
+    {
+        bool fAnyKey = false;
+        for (size_t ik = 0; ik < ckey; ++ik) {
+            uint64_t hash = dictSdsHash(rgkey[ik]);
+            dictEntry *de = dictFind(m_pdict, reinterpret_cast<void*>(hash));
+
+            if (de != nullptr)
+                fAnyKey = true;
+        }
+        if (!fAnyKey)
+            return nullptr; // All keys are missing - skip the io
+    }
+    ul.unlock();
+
+    return m_spstorage->begin_retrieve(el, proc, rgkey, ckey);
+}
+
+void StorageCache::complete_retrieve(StorageToken *tok, IStorage::callbackSingle fn) {
+    m_spstorage->complete_retrieve(tok, fn);
+}
+
 size_t StorageCache::count() const
 {
     std::unique_lock<fastlock> ul(m_lock, std::defer_lock);
@@ -233,4 +257,5 @@ void StorageCache::emergencyFreeCache() {
             dictRelease(d);
         });
     }
-}
+}
+

src/StorageCache.h

@@ -43,6 +43,10 @@ public:
     void insert(sds key, const void *data, size_t cbdata, bool fOverwrite);
     void bulkInsert(char **rgkeys, size_t *rgcbkeys, char **rgvals, size_t *rgcbvals, size_t celem);
     void retrieve(sds key, IStorage::callbackSingle fn) const;
+    StorageToken *begin_retrieve(struct aeEventLoop *el, aePostFunctionTokenProc proc, sds *rgkey, size_t ckey);
+    void complete_retrieve(StorageToken *tok, IStorage::callbackSingle fn);
+    StorageToken* begin_endWriteBatch(struct aeEventLoop *el, aePostFunctionTokenProc* proc) {return m_spstorage->begin_endWriteBatch(el,proc);} // NOP
+    void complete_endWriteBatch(StorageToken *tok) {m_spstorage->complete_endWriteBatch(tok);};
     bool erase(sds key);
     void emergencyFreeCache();
     bool keycacheIsEnabled() const { return m_pdict != nullptr; }
@@ -51,7 +55,7 @@ public:
     bool enumerate(IStorage::callback fn) const { return m_spstorage->enumerate(fn); }
     bool enumerate_hashslot(IStorage::callback fn, unsigned int hashslot) const { return m_spstorage->enumerate_hashslot(fn, hashslot); }
 
-    std::vector<std::string> getExpirationCandidates(unsigned int count, long long time) { return m_spstorage->getExpirationCandidates(count, time); }
+    std::vector<std::string> getExpirationCandidates(unsigned int count) { return m_spstorage->getExpirationCandidates(count); }
     std::vector<std::string> getEvictionCandidates(unsigned int count) { return m_spstorage->getEvictionCandidates(count); }
     void setExpire(const char *key, size_t cchKey, long long expire) { m_spstorage->setExpire(key, cchKey, expire); }
     void removeExpire(const char *key, size_t cchKey, long long expire) { m_spstorage->removeExpire(key, cchKey, expire); }

src/ae.cpp

@@ -113,6 +113,8 @@ enum class AE_ASYNC_OP
     PostCppFunction,
     DeleteFileEvent,
     CreateFileEvent,
+    PostAsynDBFunction,
+
 };
 
 struct aeCommand
@@ -125,6 +127,8 @@ struct aeCommand
         aePostFunctionProc *proc;
         aeFileProc *fproc;
         std::function<void()> *pfn;
+        aePostFunctionTokenProc* tproc;
+
     };
     void *clientData;
 };
@@ -175,6 +179,17 @@ void aeProcessCmd(aeEventLoop *eventLoop, int fd, void *, int )
             (*cmd.pfn)();
 
             delete cmd.pfn;
+            break;
+        }
+        case AE_ASYNC_OP::PostAsynDBFunction:
+        {   //added to support async api IStorage
+           if (cmd.fLock && !ulock.owns_lock()) {
+                g_forkLock.releaseRead();
+                ulock.lock();
+                g_forkLock.acquireRead();
+            }  
+            ((aePostFunctionTokenProc*)cmd.tproc)(eventLoop,(StorageToken*)cmd.clientData);
+            break; 
         }
             break;
         }
@@ -256,6 +271,20 @@ int aePostFunction(aeEventLoop *eventLoop, aePostFunctionProc *proc, void *arg)
     return AE_OK;
 }
 
+int aePostFunction(aeEventLoop *eventLoop, aePostFunctionTokenProc *proc, StorageToken *token)
+{
+  //added to support async api IStorage
+    aeCommand cmd = {};
+    cmd.op = AE_ASYNC_OP::PostAsynDBFunction; 
+    cmd.tproc = proc;
+    cmd.clientData = (void*)token; 
+    cmd.fLock = false;
+    auto size = write(eventLoop->fdCmdWrite, &cmd, sizeof(cmd));
+    if (size != sizeof(cmd))
+        return AE_ERR;
+    return AE_OK;
+}
+
 int aePostFunction(aeEventLoop *eventLoop, std::function<void()> fn, bool fLock, bool fForceQueue)
 {
     if (eventLoop == g_eventLoopThisThread && !fForceQueue)
@@ -861,7 +890,7 @@ void aeReleaseForkLock()
 
 void aeForkLockInChild()
 {
-    g_forkLock.setNotify(false);
+    g_forkLock.setMulti(false);
 }
 
 int aeThreadOwnsLock()

src/ae.h

@@ -72,13 +72,15 @@ extern "C" {
 #define AE_NOTUSED(V) ((void) V)
 
 struct aeEventLoop;
-
 /* Types and data structures */
 typedef void aeFileProc(struct aeEventLoop *eventLoop, int fd, void *clientData, int mask);
 typedef int aeTimeProc(struct aeEventLoop *eventLoop, long long id, void *clientData);
 typedef void aeEventFinalizerProc(struct aeEventLoop *eventLoop, void *clientData);
 typedef void aeBeforeSleepProc(struct aeEventLoop *eventLoop);
 typedef void aePostFunctionProc(void *pvArgs);
+//added to support async api IStorage
+struct StorageToken;
+typedef void aePostFunctionTokenProc(struct aeEventLoop *el, struct StorageToken *token);
 
 /* File event structure */
 typedef struct aeFileEvent {
@@ -134,6 +136,8 @@ int aePostFunction(aeEventLoop *eventLoop, aePostFunctionProc *proc, void *arg);
 #ifdef __cplusplus
 }   // EXTERN C
 int aePostFunction(aeEventLoop *eventLoop, std::function<void()> fn, bool fLock = true, bool fForceQueue = false);
+//added to support async api IStorage
+int aePostFunction(aeEventLoop *eventLoop, aePostFunctionTokenProc *proc, StorageToken *token);
 extern "C" {
 #endif
 void aeDeleteEventLoop(aeEventLoop *eventLoop);

src/blocked.cpp

@@ -92,7 +92,8 @@ void blockClient(client *c, int btype) {
     /* Master client should never be blocked unless pause or module */
     serverAssert(!(c->flags & CLIENT_MASTER &&
                    btype != BLOCKED_MODULE &&
-                   btype != BLOCKED_PAUSE));
+                   btype != BLOCKED_PAUSE &&
+                   btype != BLOCKED_STORAGE));
 
     c->flags |= CLIENT_BLOCKED;
     c->btype = btype;
@@ -207,6 +208,8 @@ void unblockClient(client *c) {
     } else if (c->btype == BLOCKED_PAUSE) {
         listDelNode(g_pserver->paused_clients,c->paused_list_node);
         c->paused_list_node = NULL;
+    } else if (c->btype == BLOCKED_STORAGE) {
+        serverTL->vecclientsProcess.push_back(c);
     } else {
         serverPanic("Unknown btype in unblockClient().");
     }

src/cluster.cpp

@@ -5991,7 +5991,9 @@ clusterNode *getNodeByQuery(client *c, struct redisCommand *cmd, robj **argv, in
                 /* If it is not the first key, make sure it is exactly
                  * the same key as the first we saw. */
                 if (!equalStringObjects(firstkey,thiskey)) {
-                    if (slot != thisslot) {
+                    clusterNode* nThisKey = g_pserver->cluster->slots[thisslot];
+
+                    if ((slot != thisslot) && (nThisKey != n || migrating_slot || importing_slot || g_pserver->cluster->migrating_slots_to[slot] != nullptr || g_pserver->cluster->importing_slots_from[slot] != nullptr)) {
                         /* Error: multiple keys from different slots. */
                         getKeysFreeResult(&result);
                         if (error_code)

src/config.cpp

@@ -361,7 +361,7 @@ bool initializeStorageProvider(const char **err)
             // Create The Storage Factory (if necessary)
             serverLog(LL_NOTICE, "Initializing FLASH storage provider (this may take a long time)");
             adjustOpenFilesLimit();
-            g_pserver->m_pstorageFactory = CreateRocksDBStorageFactory(g_sdsArgs, cserver.dbnum, cserver.storage_conf, cserver.storage_conf ? strlen(cserver.storage_conf) : 0);
+            g_pserver->m_pstorageFactory = CreateRocksDBStorageFactory(g_sdsArgs, cserver.dbnum, cserver.storage_conf, cserver.storage_conf ? strlen(cserver.storage_conf) : 0, &g_pserver->asyncreadworkqueue, &g_pserver->asyncwriteworkqueue);
 #else
             serverLog(LL_WARNING, "To use the flash storage provider please compile KeyDB with ENABLE_FLASH=yes");
             serverLog(LL_WARNING, "Exiting due to the use of an unsupported storage provider");
@@ -474,6 +474,90 @@ void initConfigValues() {
     }
 }
 
+struct in_addr maskFromBitsIPV4(int bits) {
+    struct in_addr mask;
+    memset(&mask, 0, sizeof(mask));
+    mask.s_addr = htonl(ULONG_MAX << (IPV4_BITS - bits));
+    return mask;
+}
+
+bool validateAndAddIPV4(sds string) {
+    struct in_addr ip, mask;
+    int subnetParts;
+    sds* subnetSplit = sdssplitlen(string, sdslen(string), "/", 1, &subnetParts);
+    TCleanup splitCleanup([&](){
+        sdsfreesplitres(subnetSplit, subnetParts);
+    });
+    if (subnetParts > 2) {
+        return false;
+    } else if (subnetParts == 2) {
+        char* endptr;
+        long bits = strtol(subnetSplit[1], &endptr, 10);
+        if (*endptr != '\0' || bits < 0 || bits > IPV4_BITS) {
+            return false;
+        }
+        mask = maskFromBitsIPV4(bits);
+    } else {
+        mask = maskFromBitsIPV4(IPV4_BITS);
+    }
+    if (subnetParts > 0) {
+        if (inet_pton(AF_INET, subnetSplit[0], &ip) == 0) {
+            return false;
+        }
+        g_pserver->overload_ignorelist.emplace(ip, mask);
+        return true;
+    } else {
+        return false;
+    }
+}
+
+struct in6_addr maskFromBitsIPV6(int bits) {
+    struct in6_addr mask;
+    memset(&mask, 0, sizeof(mask));
+    for (unsigned int i = 0; i < sizeof(struct in6_addr); i++) {
+        if (bits >= CHAR_BIT) {
+            mask.s6_addr[i] = UCHAR_MAX;
+            bits -= CHAR_BIT;
+        } else if (bits > 0) {
+            mask.s6_addr[i] = UCHAR_MAX << (CHAR_BIT - bits);
+            bits = 0;
+        } else {
+            break;
+        }
+    }
+    return mask;
+}
+
+bool validateAndAddIPV6(sds string) {
+    struct in6_addr ip, mask;
+    int subnetParts;
+    sds* subnetSplit = sdssplitlen(string, sdslen(string), "/", 1, &subnetParts);
+    TCleanup splitCleanup([&](){
+        sdsfreesplitres(subnetSplit, subnetParts);
+    });
+    if (subnetParts > 2) {
+        return false;
+    } else if (subnetParts == 2) {
+        char* endptr;
+        long bits = strtol(subnetSplit[1], &endptr, 10);
+        if (*endptr != '\0' || bits < 0 || bits > IPV6_BITS) {
+            return false;
+        }
+        mask = maskFromBitsIPV6(bits);
+    } else {
+        mask = maskFromBitsIPV6(IPV6_BITS);
+    }
+    if (subnetParts > 0) {
+        if (inet_pton(AF_INET6, subnetSplit[0], &ip) == 0) {
+            return false;
+        }
+        g_pserver->overload_ignorelist_ipv6.emplace(ip, mask);
+        return true;
+    } else {
+        return false;
+    }
+}
+
 void loadServerConfigFromString(char *config) {
     const char *err = NULL;
     int linenum = 0, totlines, i;
@@ -779,6 +863,27 @@ void loadServerConfigFromString(char *config) {
             }
             for (int i = 1; i < argc; i++)
                 g_pserver->tls_auditlog_blocklist.emplace(argv[i], strlen(argv[i]));
+        } else if (!strcasecmp(argv[0], "tls-overload-ignorelist")) {
+            if (argc < 2) {
+                err = "must supply at least one element in the ignore list"; goto loaderr;
+            }
+            if (!g_pserver->tls_overload_ignorelist.empty()) {
+                err = "tls-overload-ignorelist may only be set once"; goto loaderr;
+            }
+            for (int i = 1; i < argc; i++)
+                g_pserver->tls_overload_ignorelist.emplace(argv[i], strlen(argv[i]));
+        } else if (!strcasecmp(argv[0], "overload-ignorelist")) {
+            if (argc < 2) {
+                err = "must supply at least one element in the ignore list"; goto loaderr;
+            }
+            if (!g_pserver->overload_ignorelist.empty() || !g_pserver->overload_ignorelist_ipv6.empty()) {
+                err = "overload-ignorelist may only be set once"; goto loaderr;
+            }
+            for (int i = 1; i < argc; i++) {
+                if (!validateAndAddIPV4(argv[i]) && !validateAndAddIPV6(argv[i])) {
+                    err = "overload-ignorelist must be a list of valid IP addresses or subnets"; goto loaderr;
+                }
+            }
         } else if (!strcasecmp(argv[0], "version-override") && argc == 2) {
             KEYDB_SET_VERSION = zstrdup(argv[1]);
             serverLog(LL_WARNING, "Warning version is overriden to: %s\n", KEYDB_SET_VERSION);
@@ -908,8 +1013,8 @@ void configSetCommand(client *c) {
 
     if (c->argc < 4 || c->argc > 4) {
         o = nullptr;
-        // Variadic set is only supported for tls-allowlist
-        if (strcasecmp(szFromObj(c->argv[2]), "tls-allowlist")) {
+        // Variadic set is only supported for tls-allowlist, tls-auditlog-blocklist, overload-ignorelist and tls-overload-ignorelist
+        if (strcasecmp(szFromObj(c->argv[2]), "tls-allowlist") && strcasecmp(szFromObj(c->argv[2]), "tls-auditlog-blocklist") && strcasecmp(szFromObj(c->argv[2]), "overload-ignorelist") && strcasecmp(szFromObj(c->argv[2]), "tls-overload-ignorelist") ) {
             addReplySubcommandSyntaxError(c);
             return;
         }
@@ -1084,6 +1189,28 @@ void configSetCommand(client *c) {
             robj *val = c->argv[i];
             g_pserver->tls_allowlist.emplace(szFromObj(val), sdslen(szFromObj(val)));
         }
+    } config_set_special_field("tls-auditlog-blocklist") {
+        g_pserver->tls_auditlog_blocklist.clear();
+        for (int i = 3; i < c->argc; ++i) {
+            robj *val = c->argv[i];
+            g_pserver->tls_auditlog_blocklist.emplace(szFromObj(val), sdslen(szFromObj(val)));
+        }
+    } config_set_special_field("tls-overload-ignorelist") {
+        g_pserver->tls_overload_ignorelist.clear();
+        for (int i = 3; i < c->argc; ++i) {
+            robj *val = c->argv[i];
+            g_pserver->tls_overload_ignorelist.emplace(szFromObj(val), sdslen(szFromObj(val)));
+        }
+    } config_set_special_field("overload-ignorelist") {
+        g_pserver->overload_ignorelist.clear();
+        g_pserver->overload_ignorelist_ipv6.clear();
+        for (int i = 3; i < c->argc; ++i) {
+            robj *val = c->argv[i];
+            if (!validateAndAddIPV4(szFromObj(val)) && !validateAndAddIPV6(szFromObj(val))) {
+                addReplyError(c, "overload-ignorelist must be a list of valid IP addresses or subnets");
+                return;
+            }
+        }
     /* Everything else is an error... */
     } config_set_else {
         addReplyErrorFormat(c,"Unsupported CONFIG parameter: %s",
@@ -1295,6 +1422,37 @@ void configGetCommand(client *c) {
         }
         matches++;
     }
+    if (stringmatch(pattern, "tls-auditlog-blocklist", 1)) {
+        addReplyBulkCString(c,"tls-auditlog-blocklist");
+        addReplyArrayLen(c, (long)g_pserver->tls_auditlog_blocklist.size());
+        for (auto &elem : g_pserver->tls_auditlog_blocklist) {
+            addReplyBulkCBuffer(c, elem.get(), elem.size()); // addReplyBulkSds will free which we absolutely don't want
+        }
+        matches++;
+    }
+    if (stringmatch(pattern, "tls-overload-ignorelist", 1)) {
+        addReplyBulkCString(c,"tls-overload-ignorelist");
+        addReplyArrayLen(c, (long)g_pserver->tls_overload_ignorelist.size());
+        for (auto &elem : g_pserver->tls_overload_ignorelist) {
+            addReplyBulkCBuffer(c, elem.get(), elem.size()); // addReplyBulkSds will free which we absolutely don't want
+        }
+        matches++;
+    }
+    if (stringmatch(pattern, "overload-ignorelist", 1)) {
+        addReplyBulkCString(c,"overload-ignorelist");
+        addReplyArrayLen(c, (long)g_pserver->overload_ignorelist.size() + (long)g_pserver->overload_ignorelist_ipv6.size());
+        for (auto &elem : g_pserver->overload_ignorelist) {
+            sds elem_sds = elem.getString();
+            addReplyBulkCBuffer(c, elem_sds, sdslen(elem_sds));
+            sdsfree(elem_sds);
+        }
+        for (auto &elem : g_pserver->overload_ignorelist_ipv6) {
+            sds elem_sds = elem.getString();
+            addReplyBulkCBuffer(c, elem_sds, sdslen(elem_sds));
+            sdsfree(elem_sds);
+        }
+        matches++;
+    }
 
     setDeferredMapLen(c,replylen,matches);
 }
@@ -1978,6 +2136,56 @@ int rewriteConfig(char *path, int force_all) {
         rewriteConfigMarkAsProcessed(state, "tls-allowlist"); // ensure the line is removed if it existed
     }
 
+    if (!g_pserver->tls_overload_ignorelist.empty()) {
+        sds conf = sdsnew("tls-overload-ignorelist ");
+        for (auto &elem : g_pserver->tls_overload_ignorelist) {
+            conf = sdscatsds(conf, (sds)elem.get());
+            conf = sdscat(conf, " ");
+        }
+        // trim the trailing space
+        sdsrange(conf, 0, -1);
+        rewriteConfigRewriteLine(state,"tls-overload-ignorelist",conf,1 /*force*/);
+        // note: conf is owned by rewriteConfigRewriteLine - no need to free
+    } else {
+        rewriteConfigMarkAsProcessed(state, "tls-overload-ignorelist"); // ensure the line is removed if it existed
+    }
+
+    if (!g_pserver->tls_auditlog_blocklist.empty()) {
+        sds conf = sdsnew("tls-auditlog-blocklist ");
+        for (auto &elem : g_pserver->tls_auditlog_blocklist) {
+            conf = sdscatsds(conf, (sds)elem.get());
+            conf = sdscat(conf, " ");
+        }
+        // trim the trailing space
+        sdsrange(conf, 0, -1);
+        rewriteConfigRewriteLine(state,"tls-auditlog-blocklist",conf,1 /*force*/);
+        // note: conf is owned by rewriteConfigRewriteLine - no need to free
+    } else {
+        rewriteConfigMarkAsProcessed(state, "tls-auditlog-blocklist"); // ensure the line is removed if it existed
+    }
+
+    if (!g_pserver->overload_ignorelist.empty() || !g_pserver->overload_ignorelist_ipv6.empty()) {
+        sds conf = sdsnew("overload-ignorelist ");
+        for (auto &elem : g_pserver->overload_ignorelist) {
+            sds elem_sds = elem.getString();
+            conf = sdscatsds(conf, elem_sds);
+            sdsfree(elem_sds);
+            conf = sdscat(conf, " ");
+        }
+        for (auto &elem : g_pserver->overload_ignorelist_ipv6) {
+            sds elem_sds = elem.getString();
+            conf = sdscatsds(conf, elem_sds);
+            sdsfree(elem_sds);
+            conf = sdscat(conf, " ");
+        }
+        // trim the trailing space
+        sdsrange(conf, 0, -1);
+        rewriteConfigRewriteLine(state,"overload-ignorelist",conf,1 /*force*/);
+        // note: conf is owned by rewriteConfigRewriteLine - no need to free
+    } else {
+        rewriteConfigMarkAsProcessed(state, "overload-ignorelist"); // ensure the line is removed if it existed
+    }
+
     /* Rewrite Sentinel config if in Sentinel mode. */
     if (g_pserver->sentinel_mode) rewriteConfigSentinelOption(state);
 
@@ -2516,6 +2724,14 @@ static int isValidProcTitleTemplate(char *val, const char **err) {
     return 1;
 }
 
+static int isValidDisklessEnum(int e, const char **err) {
+    if (e == REPL_DISKLESS_LOAD_SWAPDB && g_pserver->m_pstorageFactory != nullptr) {
+        *err = "cannot use swapdb with a storage provider";
+        return 0;
+    }
+    return 1;
+}
+
 static int updateProcTitleTemplate(char *val, char *prev, const char **err) {
     UNUSED(val);
     UNUSED(prev);
@@ -2883,7 +3099,7 @@ standardConfig configs[] = {
     /* Enum Configs */
     createEnumConfig("supervised", NULL, IMMUTABLE_CONFIG, supervised_mode_enum, cserver.supervised_mode, SUPERVISED_NONE, NULL, NULL),
     createEnumConfig("syslog-facility", NULL, IMMUTABLE_CONFIG, syslog_facility_enum, g_pserver->syslog_facility, LOG_LOCAL0, NULL, NULL),
-    createEnumConfig("repl-diskless-load", NULL, MODIFIABLE_CONFIG, repl_diskless_load_enum, g_pserver->repl_diskless_load, REPL_DISKLESS_LOAD_DISABLED, NULL, NULL),
+    createEnumConfig("repl-diskless-load", NULL, MODIFIABLE_CONFIG, repl_diskless_load_enum, g_pserver->repl_diskless_load, REPL_DISKLESS_LOAD_DISABLED, isValidDisklessEnum, NULL),
     createEnumConfig("loglevel", NULL, MODIFIABLE_CONFIG, loglevel_enum, cserver.verbosity, LL_NOTICE, NULL, NULL),
     createEnumConfig("maxmemory-policy", NULL, MODIFIABLE_CONFIG, maxmemory_policy_enum, g_pserver->maxmemory_policy, MAXMEMORY_NO_EVICTION, NULL, NULL),
     createEnumConfig("appendfsync", NULL, MODIFIABLE_CONFIG, aof_fsync_enum, g_pserver->aof_fsync, AOF_FSYNC_EVERYSEC, NULL, NULL),
@@ -2946,10 +3162,11 @@ standardConfig configs[] = {
     createLongLongConfig("latency-monitor-threshold", NULL, MODIFIABLE_CONFIG, 0, LLONG_MAX, g_pserver->latency_monitor_threshold, 0, INTEGER_CONFIG, NULL, NULL),
     createLongLongConfig("proto-max-bulk-len", NULL, MODIFIABLE_CONFIG, 1024*1024, LLONG_MAX, g_pserver->proto_max_bulk_len, 512ll*1024*1024, MEMORY_CONFIG, NULL, NULL), /* Bulk request max size */
     createLongLongConfig("stream-node-max-entries", NULL, MODIFIABLE_CONFIG, 0, LLONG_MAX, g_pserver->stream_node_max_entries, 100, INTEGER_CONFIG, NULL, NULL),
-    createLongLongConfig("repl-backlog-size", NULL, MODIFIABLE_CONFIG, 1, LLONG_MAX, g_pserver->repl_backlog_size, 1024*1024, MEMORY_CONFIG, NULL, updateReplBacklogSize), /* Default: 1mb */
+    createLongLongConfig("repl-backlog-size", NULL, MODIFIABLE_CONFIG, 1, LLONG_MAX, g_pserver->repl_backlog_config_size, 1024*1024, MEMORY_CONFIG, NULL, updateReplBacklogSize), /* Default: 1mb */
     createLongLongConfig("repl-backlog-disk-reserve", NULL, IMMUTABLE_CONFIG, 0, LLONG_MAX, cserver.repl_backlog_disk_size, 0, MEMORY_CONFIG, NULL, NULL),
     createLongLongConfig("max-snapshot-slip", NULL, MODIFIABLE_CONFIG, 0, 5000, g_pserver->snapshot_slip, 400, 0, NULL, NULL),
     createLongLongConfig("max-rand-count", NULL, MODIFIABLE_CONFIG, 0, LONG_MAX/2, g_pserver->rand_total_threshold, LONG_MAX/2, 0, NULL, NULL),
+    createLongLongConfig("repl-backlog-max-writes-per-event", NULL, MODIFIABLE_CONFIG, 0, LLONG_MAX, g_pserver->repl_backlog_max_writes_per_event, 8ll*1024*1024, MEMORY_CONFIG, NULL, NULL),
 
     /* Unsigned Long Long configs */
     createULongLongConfig("maxmemory", NULL, MODIFIABLE_CONFIG, 0, LLONG_MAX, g_pserver->maxmemory, 0, MEMORY_CONFIG, NULL, updateMaxmemory),
@@ -2981,7 +3198,10 @@ standardConfig configs[] = {
     createSizeTConfig("semi-ordered-set-bucket-size", NULL, MODIFIABLE_CONFIG, 0, 1024, g_semiOrderedSetTargetBucketSize, 0, INTEGER_CONFIG, NULL, NULL),
     createSDSConfig("availability-zone", NULL, MODIFIABLE_CONFIG, 0, g_pserver->sdsAvailabilityZone, "", NULL, NULL),
     createIntConfig("overload-protect-percent", NULL, MODIFIABLE_CONFIG, 0, 200, g_pserver->overload_protect_threshold, 0, INTEGER_CONFIG, NULL, NULL),
+    createIntConfig("overload-protect-tenacity", NULL, MODIFIABLE_CONFIG, 0, 100, g_pserver->overload_protect_tenacity, 10, INTEGER_CONFIG, NULL, NULL),
     createIntConfig("force-eviction-percent", NULL, MODIFIABLE_CONFIG, 0, 100, g_pserver->force_eviction_percent, 0, INTEGER_CONFIG, NULL, NULL),
+    createBoolConfig("enable-async-rehash", NULL, MODIFIABLE_CONFIG, g_pserver->enable_async_rehash, 1, NULL, NULL),
+    createBoolConfig("enable-keydb-fastsync", NULL, MODIFIABLE_CONFIG, g_pserver->fEnableFastSync, 1, NULL, NULL),
 
 #ifdef USE_OPENSSL
     createIntConfig("tls-port", NULL, MODIFIABLE_CONFIG, 0, 65535, g_pserver->tls_port, 0, INTEGER_CONFIG, NULL, updateTLSPort), /* TCP port. */

src/connection.h

@@ -52,6 +52,7 @@ typedef enum {
 #define CONN_FLAG_READ_THREADSAFE        (1<<2)
 #define CONN_FLAG_WRITE_THREADSAFE       (1<<3)
 #define CONN_FLAG_AUDIT_LOGGING_REQUIRED (1<<4)
+#define CONN_FLAG_IGNORE_OVERLOAD (1<<5)
 
 #define CONN_TYPE_SOCKET            1
 #define CONN_TYPE_TLS               2

src/db.cpp

@@ -56,7 +56,6 @@ int expireIfNeeded(redisDb *db, robj *key, robj *o);
 void slotToKeyUpdateKeyCore(const char *key, size_t keylen, int add);
 
 std::unique_ptr<expireEntry> deserializeExpire(const char *str, size_t cch, size_t *poffset);
-sds serializeStoredObjectAndExpire(robj_roptr o);
 
 dictType dictChangeDescType {
     dictSdsHash,                /* hash function */
@@ -2529,6 +2528,12 @@ void slotToKeyFlush(int async) {
  * decrement the reference count to release the keys names. */
 unsigned int getKeysInSlot(unsigned int hashslot, robj **keys, unsigned int count) {
     if (g_pserver->m_pstorageFactory != nullptr) {
+        // We must commit so the storage engine agrees on the number of items in the hash slot
+        if (g_pserver->db[0]->FTrackingChanges()) {
+            if (g_pserver->db[0]->processChanges(false))
+                g_pserver->db[0]->commitChanges();
+            g_pserver->db[0]->trackChanges(false);
+        }
         int j = 0;
         g_pserver->db[0]->getStorageCache()->enumerate_hashslot([&](const char *key, size_t cchKey, const void *, size_t )->bool {
             keys[j++] = createStringObject(key, cchKey);
@@ -2558,6 +2563,12 @@ unsigned int getKeysInSlot(unsigned int hashslot, robj **keys, unsigned int coun
 unsigned int delKeysInSlot(unsigned int hashslot) {
     serverAssert(GlobalLocksAcquired());
     if (g_pserver->m_pstorageFactory != nullptr) {
+        // We must commit so the storage engine agrees on the number of items in the hash slot
+        if (g_pserver->db[0]->FTrackingChanges()) {
+            if (g_pserver->db[0]->processChanges(false))
+                g_pserver->db[0]->commitChanges();
+            g_pserver->db[0]->trackChanges(false);
+        }
         int j = 0;
         g_pserver->db[0]->getStorageCache()->enumerate_hashslot([&](const char *key, size_t cchKey, const void *, size_t )->bool {
             robj *keyobj = createStringObject(key, cchKey);
@@ -2703,11 +2714,6 @@ void redisDbPersistentData::prepOverwriteForSnapshot(char *key)
         auto itr = m_pdbSnapshot->find_cached_threadsafe(key);
         if (itr.key() != nullptr)
         {
-            if (itr.val()->FExpires()) {
-                // Note: I'm sure we could handle this, but its too risky at the moment.
-                //  There are known bugs doing this with expires
-                return;
-            }
             sds keyNew = sdsdupshared(itr.key());
             if (dictAdd(m_pdictTombstone, keyNew, (void*)dictHashKey(m_pdict, key)) != DICT_OK)
                 sdsfree(keyNew);
@@ -2745,6 +2751,13 @@ void redisDbPersistentData::clear(void(callback)(void*))
     if (m_spstorage != nullptr)
         m_spstorage->clear(callback);
     dictEmpty(m_pdictTombstone,callback);
+
+    // To avoid issues with async rehash we completly free the old dict and create a fresh one
+    dictRelease(m_pdict);
+    dictRelease(m_pdictTombstone);
+    m_pdict = dictCreate(&dbDictType, this);
+    m_pdictTombstone = dictCreate(&dbTombstoneDictType, this);
+
     m_pdbSnapshot = nullptr;
     m_numexpires = 0;
 }
@@ -2880,10 +2893,11 @@ LNotFound:
             {
                 dictAdd(m_pdict, sdsNewKey, o);
                 
+                o->SetFExpires(spexpire != nullptr);
                 if (spexpire != nullptr) {
                     o->expire = std::move(*spexpire);
+                    ++m_numexpires;
                 }
-                o->SetFExpires(spexpire != nullptr);
                 g_pserver->stat_storage_provider_read_hits++;
             } else {
                 sdsfree(sdsNewKey);
@@ -2987,6 +3001,7 @@ void redisDbPersistentData::processChangesAsync(std::atomic<int> &pendingJobs)
     dict *dictNew = dictCreate(&dbDictType, nullptr);
     std::swap(dictNew, m_pdict);
     m_cnewKeysPending = 0;
+    m_numexpires = 0;
     g_pserver->asyncworkqueue->AddWorkFunction([dictNew, this, &pendingJobs]{
         dictIterator *di = dictGetIterator(dictNew);
         dictEntry *de;
@@ -3042,7 +3057,18 @@ void redisDbPersistentData::commitChanges(const redisDbPersistentDataSnapshot **
         m_pdbSnapshotStorageFlush = nullptr;
     }
     if (m_spstorage != nullptr)
+    {
+#if 1
         m_spstorage->endWriteBatch();
+#else
+        auto *tok = m_spstorage->begin_endWriteBatch(serverTL->el, storageLoadCallback);
+        if (tok != nullptr)
+        {
+            tok->db = this;
+            tok->type = StorageToken::TokenType::BatchWrite;
+        }
+#endif
+    }
 }
 
 redisDbPersistentData::~redisDbPersistentData()
@@ -3114,7 +3140,14 @@ bool redisDbPersistentData::removeCachedValue(const char *key, dictEntry **ppde)
     // since we write ASAP the database already has a valid copy so safe to delete
     if (ppde != nullptr) {
         *ppde = dictUnlink(m_pdict, key);
+        robj *o = (robj*)dictGetVal(*ppde);
+        if (o->FExpires())
+            --m_numexpires;
     } else {
+        dictEntry *deT = dictFind(m_pdict, key);
+        robj *o = (robj*)dictGetVal(deT);
+        if (o->FExpires())
+            --m_numexpires;
         dictDelete(m_pdict, key);
     }
 
@@ -3134,8 +3167,11 @@ void redisDbPersistentData::trackChanges(bool fBulk, size_t sizeHint)
     if (fBulk)
         m_fAllChanged.fetch_add(1, std::memory_order_acq_rel);
 
-    if (sizeHint > 0)
+    if (sizeHint > 0) {
+        aeAcquireLock();
         dictExpand(m_dictChanged, sizeHint, false);
+        aeReleaseLock();
+    }
 }
 
 void redisDbPersistentData::removeAllCachedValues()
@@ -3158,6 +3194,7 @@ void redisDbPersistentData::removeAllCachedValues()
     } else {
         dictEmpty(m_pdict, nullptr);
     }
+    m_numexpires = 0;
 }
 
 void redisDbPersistentData::disableKeyCache()
@@ -3309,90 +3346,124 @@ void redisDbPersistentData::prefetchKeysAsync(client *c, parsed_command &command
                 }
             }
         }
+#else
+        UNUSED(c);
+        UNUSED(command);
 #endif
         return;
     }
+}
 
-    AeLocker lock;
+void redisDbPersistentData::prefetchKeysFlash(const std::unordered_set<client*> &setc)
+{
+    serverAssert(GlobalLocksAcquired());
+    std::vector<sds> veckeys;
+    std::unordered_set<client*> setcBlocked;
 
-    std::vector<robj*> veckeys;
-    lock.arm(c);
-    getKeysResult result = GETKEYS_RESULT_INIT;
-    auto cmd = lookupCommand(szFromObj(command.argv[0]));
-    if (cmd == nullptr)
-        return; // Bad command? It's not for us to judge, just bail
-    
-    if (command.argc < std::abs(cmd->arity))
-        return; // Invalid number of args
-    
-    int numkeys = getKeysFromCommand(cmd, command.argv, command.argc, &result);
-    for (int ikey = 0; ikey < numkeys; ++ikey)
-    {
-        robj *objKey = command.argv[result.keys[ikey]];
-        if (this->find_cached_threadsafe(szFromObj(objKey)) == nullptr)
-            veckeys.push_back(objKey);
-    }
-    lock.disarm();
+    for (client *c : setc) {
+        for (auto &command : c->vecqueuedcmd) {
+            getKeysResult result = GETKEYS_RESULT_INIT;
+            if (command.argc == 0)  // parse can do this it will be handled by processClient
+                break;
+            auto cmd = lookupCommand(szFromObj(command.argv[0]));
+            if (cmd == nullptr)
+                break; // Bad command? It's not for us to judge, just bail
+            
+            if (command.argc < std::abs(cmd->arity))
+                break; // Invalid number of args
+            
+            int numkeys = getKeysFromCommand(cmd, command.argv, command.argc, &result);
+            bool fQueued = false;
+            for (int ikey = 0; ikey < numkeys; ++ikey)
+            {
+                robj *objKey = command.argv[result.keys[ikey]];
+                if (this->find_cached_threadsafe(szFromObj(objKey)) == nullptr) {
+                    veckeys.push_back(szFromObj(objKey));
+                    fQueued = true;
+                }
+            }
 
-    getKeysFreeResult(&result);
-
-    std::vector<std::tuple<sds, robj*, std::unique_ptr<expireEntry>>> vecInserts;
-    for (robj *objKey : veckeys)
-    {
-        sds sharedKey = sdsdupshared((sds)szFromObj(objKey));
-        std::unique_ptr<expireEntry> spexpire;
-        robj *o = nullptr;
-        m_spstorage->retrieve((sds)szFromObj(objKey), [&](const char *, size_t, const void *data, size_t cb){
-                size_t offset = 0;
-                spexpire = deserializeExpire((const char*)data, cb, &offset);    
-                o = deserializeStoredObject(reinterpret_cast<const char*>(data) + offset, cb - offset);
-                serverAssert(o != nullptr);
-        });
-
-        if (o != nullptr) {
-            vecInserts.emplace_back(sharedKey, o, std::move(spexpire));
-        } else if (sharedKey != nullptr) {
-            sdsfree(sharedKey);
+            if (fQueued) {
+                setcBlocked.insert(c);
+            }
+            getKeysFreeResult(&result);
         }
     }
 
-    if (!vecInserts.empty()) {
-        lock.arm(c);
-        for (auto &tuple : vecInserts)
-        {
-            sds sharedKey = std::get<0>(tuple);
-            robj *o = std::get<1>(tuple);
-            std::unique_ptr<expireEntry> spexpire = std::move(std::get<2>(tuple));
+    if (veckeys.empty())
+        return;
 
-            if (o != nullptr)
-            {
-                if (this->find_cached_threadsafe(sharedKey) != nullptr)
-                {
-                    // While unlocked this was already ensured
-                    decrRefCount(o);
-                    sdsfree(sharedKey);
-                }
-                else
-                {
-                    if (spexpire != nullptr) {
-                        if (spexpire->when() < mstime()) {
-                            break;
-                        }
-                    }
-                    dictAdd(m_pdict, sharedKey, o);
-                    if (spexpire != nullptr)
-                        o->expire = std::move(*spexpire);
-                    o->SetFExpires(spexpire != nullptr);
-                }
-            }
-            else
-            {
-                if (sharedKey != nullptr)
-                    sdsfree(sharedKey); // BUG but don't bother crashing
+    StorageToken *tok = m_spstorage->begin_retrieve(serverTL->el, storageLoadCallback, veckeys.data(), veckeys.size());
+    if (tok != nullptr) {
+        for (client *c : setcBlocked) {
+            if (!(c->flags & CLIENT_BLOCKED)) {
+                blockClient(c, BLOCKED_STORAGE);
             }
         }
-        lock.disarm();
+        tok->setc = std::move(setcBlocked);
+        tok->type = StorageToken::TokenType::SingleRead;
+        tok->db = this;
     }
-
     return;
 }
+
+/*static*/ void redisDbPersistentData::storageLoadCallback(aeEventLoop *, StorageToken *tok) {
+    serverTL->setStorageTokensProcess.insert(tok);
+}
+
+void redisDbPersistentData::processStorageToken(StorageToken *tok) {
+    auto setc = std::move(tok->setc);
+    switch (tok->type)
+    {
+
+    case StorageToken::TokenType::SingleRead:
+    {
+        tok->db->m_spstorage->complete_retrieve(tok, [&](const char *szKey, size_t cbKey, const void *data, size_t cb) {
+            auto *db = tok->db;
+            size_t offset = 0;
+            sds key = sdsnewlen(szKey, -((ssize_t)cbKey));
+            auto spexpire = deserializeExpire((const char*)data, cb, &offset);    
+            robj *o = deserializeStoredObject(reinterpret_cast<const char*>(data) + offset, cb - offset);
+            serverAssert(o != nullptr);
+
+            if (db->find_cached_threadsafe(key) != nullptr) {
+            LUnneeded:
+                // While unlocked this was already ensured
+                decrRefCount(o);
+                sdsfree(key);
+            } else {
+                if (spexpire != nullptr) {
+                    if (spexpire->when() < mstime()) {
+                        goto LUnneeded;
+                    }
+                }
+                dictAdd(db->m_pdict, key, o);
+                o->SetFExpires(spexpire != nullptr);
+                if (spexpire != nullptr) {
+                    o->expire = std::move(*spexpire);
+                }
+            } 
+        });
+        break;
+    }
+    case StorageToken::TokenType::BatchWrite:
+    {
+        tok->db->m_spstorage->complete_endWriteBatch(tok);
+        break;
+    }
+    default:
+        serverAssert((tok->type == StorageToken::TokenType::SingleRead) || (tok->type == StorageToken::TokenType::BatchWrite));
+        break;
+    } //switch end
+
+    tok = nullptr; // Invalid past this point
+
+    for (client *c : setc)
+    {
+        std::unique_lock<fastlock> ul(c->lock);
+        if (c->flags & CLIENT_BLOCKED)
+            unblockClient(c);
+        
+        serverTL->vecclientsProcess.push_back(c);
+    }
+}

src/debug.cpp

@@ -79,6 +79,7 @@ void printCrashReport(void);
 void bugReportEnd(int killViaSignal, int sig);
 void logStackTrace(void *eip, int uplevel);
 void getTempFileName(char tmpfile[], int tmpfileNum);
+bool initializeStorageProvider(const char **err);
 
 /* ================================= Debugging ============================== */
 
@@ -308,7 +309,7 @@ void computeDatasetDigest(unsigned char *final) {
         mixDigest(final,&aux,sizeof(aux));
 
         /* Iterate this DB writing every entry */
-        db->iterate_threadsafe([final, db](const char *key, robj_roptr o)->bool {
+        db->iterate_threadsafe([final](const char *key, robj_roptr o)->bool {
             unsigned char digest[20];
             robj *keyobj;
 
@@ -535,30 +536,75 @@ NULL
                 return;
             }
         }
+        
+        if (g_pserver->m_pstorageFactory) {
+            protectClient(c);
 
-        /* The default behavior is to save the RDB file before loading
-         * it back. */
-        if (save) {
-            rdbSaveInfo rsi, *rsiptr;
-            rsiptr = rdbPopulateSaveInfo(&rsi);
-            if (rdbSave(nullptr, rsiptr) != C_OK) {
-                addReply(c,shared.err);
+            for (int idb = 0; idb < cserver.dbnum; ++idb) {
+                if (g_pserver->db[idb]->processChanges(false))
+                    g_pserver->db[idb]->commitChanges();
+                g_pserver->db[idb]->storageProviderDelete();
+            }
+            delete g_pserver->metadataDb;
+
+            if (flush) emptyDb(-1,EMPTYDB_NO_FLAGS,NULL);
+
+            delete g_pserver->m_pstorageFactory;
+
+            const char *err;
+            if (!initializeStorageProvider(&err))
+            {
+                serverLog(LL_WARNING, "Failed to initialize storage provider: %s",err);
+                exit(EXIT_FAILURE);
+            }
+
+            g_pserver->metadataDb = g_pserver->m_pstorageFactory->createMetadataDb();
+            for (int idb = 0; idb < cserver.dbnum; ++idb)
+            {
+                int dbid = idb;
+                std::string dbid_key = "db-" + std::to_string(idb);
+                g_pserver->metadataDb->retrieve(dbid_key.c_str(), dbid_key.length(), [&](const char *, size_t, const void *data, size_t){
+                    dbid = *(int*)data;
+                });
+                delete g_pserver->db[idb];
+                g_pserver->db[idb] = new (MALLOC_LOCAL) redisDb();
+                g_pserver->db[idb]->initialize(dbid);
+            }
+
+            moduleFireServerEvent(REDISMODULE_EVENT_LOADING, REDISMODULE_SUBEVENT_LOADING_FLASH_START, NULL);
+            for (int idb = 0; idb < cserver.dbnum; ++idb)
+            {
+                g_pserver->db[idb]->storageProviderInitialize();
+                g_pserver->db[idb]->trackChanges(false);
+            }
+            moduleFireServerEvent(REDISMODULE_EVENT_LOADING, REDISMODULE_SUBEVENT_LOADING_ENDED, NULL);
+            
+            unprotectClient(c);
+        } else {
+            /* The default behavior is to save the RDB file before loading
+            * it back. */
+            if (save) {
+                rdbSaveInfo rsi, *rsiptr;
+                rsiptr = rdbPopulateSaveInfo(&rsi);
+                if (rdbSave(nullptr, rsiptr) != C_OK) {
+                    addReply(c,shared.err);
+                    return;
+                }
+            }
+
+            /* The default behavior is to remove the current dataset from
+            * memory before loading the RDB file, however when MERGE is
+            * used together with NOFLUSH, we are able to merge two datasets. */
+            if (flush) emptyDb(-1,EMPTYDB_NO_FLAGS,NULL);
+
+            protectClient(c);
+            int ret = rdbLoadFile(g_pserver->rdb_filename,NULL,flags);
+            unprotectClient(c);
+            if (ret != C_OK) {
+                addReplyError(c,"Error trying to load the RDB dump");
                 return;
             }
         }
-
-        /* The default behavior is to remove the current dataset from
-         * memory before loading the RDB file, however when MERGE is
-         * used together with NOFLUSH, we are able to merge two datasets. */
-        if (flush) emptyDb(-1,EMPTYDB_NO_FLAGS,NULL);
-
-        protectClient(c);
-        int ret = rdbLoadFile(g_pserver->rdb_filename,NULL,flags);
-        unprotectClient(c);
-        if (ret != C_OK) {
-            addReplyError(c,"Error trying to load the RDB dump");
-            return;
-        }
         serverLog(LL_WARNING,"DB reloaded by DEBUG RELOAD");
         addReply(c,shared.ok);
     } else if (!strcasecmp(szFromObj(c->argv[1]),"loadaof")) {

src/expire.cpp

@@ -518,11 +518,13 @@ void pexpireMemberAtCommand(client *c)
             } while (sampled == 0 ||
                     (expired*100/sampled) > config_cycle_acceptable_stale);
         } else {
-            long long now;
+            long prev_expired;
+            long long now = mstime();
             size_t tried = 0;
             std::vector<std::string> keys;
             do {
-                keys = db->getStorageCache()->getExpirationCandidates(ACTIVE_EXPIRE_CYCLE_LOOKUPS_PER_LOOP, now);
+                prev_expired = total_expired;
+                keys = db->getStorageCache()->getExpirationCandidates(ACTIVE_EXPIRE_CYCLE_LOOKUPS_PER_LOOP);
                 for (std::string key : keys) {
                     robj* keyobj = createStringObject(key.c_str(), key.size());
                     db->find(szFromObj(keyobj));
@@ -533,7 +535,7 @@ void pexpireMemberAtCommand(client *c)
                 }
                 total_sampled += keys.size();
                 elapsed = ustime()-start;
-            } while (keys.size() > 0 && (elapsed < timelimit));
+            } while (keys.size() > 0 && (elapsed < timelimit) && (total_expired - prev_expired) > 0);
 
             if (ustime()-start > timelimit) {
                 timelimit_exit = 1;

src/expire.h

@@ -160,6 +160,8 @@ public:
 
     expireEntry &operator=(expireEntry &&e)
     {
+        if (FFat())
+            delete pfatentry();
         s = e.s;
         e.s.m_whenAndPtrUnion = 0;
         e.s.fFat = false;
@@ -167,6 +169,8 @@ public:
     }
 
     expireEntry &operator=(expireEntry &e) {
+        if (FFat())
+            delete pfatentry();
         if (e.FFat()) {
             s.m_whenAndPtrUnion = reinterpret_cast<long long>(new expireEntryFat(*e.pfatentry()));
             s.fFat = true;

src/module.cpp

@@ -690,8 +690,9 @@ void moduleHandlePropagationAfterCommandCallback(RedisModuleCtx *ctx) {
 }
 
 /* Free the context after the user function was called. */
-void moduleFreeContext(RedisModuleCtx *ctx) {
-    moduleHandlePropagationAfterCommandCallback(ctx);
+void moduleFreeContext(RedisModuleCtx *ctx, bool propogate) {
+    if (propogate)
+        moduleHandlePropagationAfterCommandCallback(ctx);
     autoMemoryCollect(ctx);
     poolAllocRelease(ctx);
     if (ctx->postponed_arrays) {
@@ -7057,7 +7058,7 @@ RedisModuleString *RM_DictPrev(RedisModuleCtx *ctx, RedisModuleDictIter *di, voi
 /* Compare the element currently pointed by the iterator to the specified
  * element given by key/keylen, according to the operator 'op' (the set of
  * valid operators are the same valid for RedisModule_DictIteratorStart).
- * If the comparision is successful the command returns REDISMODULE_OK
+ * If the comparison is successful the command returns REDISMODULE_OK
  * otherwise REDISMODULE_ERR is returned.
  *
  * This is useful when we want to just emit a lexicographical range, so

src/modules/keydb_modstatsd/Makefile

@@ -4,7 +4,7 @@ OBJECT_FILES := modmain.o
 MODSNAP_CXX_FLAGS := -std=gnu++14
 
 %.o: %.cpp
-	$(CXX) -o $@ -c $< $(MODULE_FLAGS) -I../../../deps/cpp-statsd-client/include $(MODSNAP_CXX_FLAGS)
+	$(CXX) -o $@ -c $< $(MODULE_FLAGS) -I../../../deps/cpp-statsd-client/include $(MODSNAP_CXX_FLAGS) -g
 
 modstatsd.so: $(OBJECT_FILES)
 	$(CXX) -shared $(OBJECT_FILES) -o modstatsd.so

src/modules/keydb_modstatsd/modmain.cpp

@@ -67,8 +67,6 @@ public:
 
 /* constants */
 static time_t c_infoUpdateSeconds = 10;
-// the current Redis Cluster setup we configure replication factor as 2, each non-empty master node should have 2 replicas, given that there are 3 zones in each regions
-static const int EXPECTED_NUMBER_OF_REPLICAS = 2;
 
 StatsdClientWrapper *g_stats = nullptr;
 std::string m_strPrefix { "keydb" };
@@ -76,7 +74,7 @@ std::string m_strPrefix { "keydb" };
 const std::regex g_replica_or_db_info_regex { "^(slave|db)(\\d+)" };
 const char *g_string_counter_separator = "__";
 const uint64_t g_stats_buffer_size_bytes = 1600;
-utsname sysName;
+std::string nodeName;
 int unameResult;
 
 enum class StatsD_Type {
@@ -94,6 +92,7 @@ struct StatsRecord {
 
     /* Dynamic Values */
     long long prevVal = 0;
+    long long currVal = 0;
 };
 
 std::unordered_map<std::string, StatsRecord> g_mapInfoFields = {
@@ -101,6 +100,8 @@ std::unordered_map<std::string, StatsRecord> g_mapInfoFields = {
     { "used_memory", { StatsD_Type::STATSD_GAUGE_BYTES, false /* prefixOnly */}},
     { "used_memory_rss", { StatsD_Type::STATSD_GAUGE_BYTES }},
     { "maxmemory", { StatsD_Type::STATSD_GAUGE_BYTES, false /* prefixOnly */}},
+    { "maxstorage", { StatsD_Type::STATSD_GAUGE_BYTES, false /* prefixOnly */}},
+    { "storage_used", { StatsD_Type::STATSD_GAUGE_BYTES, false /* prefixOnly */}},
     { "used_memory_dataset_perc", { StatsD_Type::STATSD_GAUGE_FLOAT }},
     { "avg_lock_contention", { StatsD_Type::STATSD_GAUGE_LONGLONG }},
     { "repl_backlog_size", { StatsD_Type::STATSD_GAUGE_BYTES }},
@@ -253,6 +254,7 @@ std::unordered_map<std::string, StatsRecord> g_mapInfoFields = {
     { "cluster_size", { StatsD_Type::STATSD_GAUGE_LONGLONG }},
     { "storage_flash_available_bytes", { StatsD_Type::STATSD_GAUGE_BYTES }},
     { "storage_flash_total_bytes", { StatsD_Type::STATSD_GAUGE_BYTES }},
+    { "last_overload_cpu_reading", { StatsD_Type::STATSD_GAUGE_FLOAT }},
 };
 
 /* Globals */
@@ -297,6 +299,8 @@ void handleStatItem(struct RedisModuleCtx *ctx, std::string name, StatsRecord &r
     switch (record.type) {
         case StatsD_Type::STATSD_GAUGE_LONGLONG: {
             long long val = strtoll(pchValue, nullptr, 10);
+            record.prevVal = record.currVal;
+            record.currVal = val;
             g_stats->gauge(name, val, record.prefixOnly);
             RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"%s\": %lld", name.c_str(), val);
             break;
@@ -318,9 +322,10 @@ void handleStatItem(struct RedisModuleCtx *ctx, std::string name, StatsRecord &r
 
         case StatsD_Type::STATSD_DELTA: {
             long long val = strtoll(pchValue, nullptr, 10);
+            record.prevVal = record.currVal;
+            record.currVal = val;
             g_stats->count(name, val - record.prevVal, record.prefixOnly);
             RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric count for \"%s\": %lld", name.c_str() , val - record.prevVal);
-            record.prevVal = val;
             break;
         }
 
@@ -431,6 +436,20 @@ void handle_info_response(struct RedisModuleCtx *ctx, const char *szReply, size_
     }
 
     #undef SAFETY_CHECK_POINTER
+
+    // Emit keyspace hit rate metric which is computed from info values
+    auto hit = g_mapInfoFields.find("keyspace_hits");
+    auto miss = g_mapInfoFields.find("keyspace_misses");
+    if (hit != g_mapInfoFields.end() && miss != g_mapInfoFields.end()) {
+        long long hitValDelta = hit->second.currVal - hit->second.prevVal;
+        long long missValDelta = miss->second.currVal - miss->second.prevVal;
+        long long total = hitValDelta + missValDelta;
+        if (total > 0) {
+            double hitRate = (double)hitValDelta / total * 1000000;
+            g_stats->gauge("keyspace_hit_rate_per_million", hitRate, false /* prefixOnly */);
+            RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"keyspace_hit_rate_per_million\": %f", hitRate);
+        }
+    }
 }
 
 void handle_cluster_nodes_response(struct RedisModuleCtx *ctx, const char *szReply, size_t len) {
@@ -438,6 +457,8 @@ void handle_cluster_nodes_response(struct RedisModuleCtx *ctx, const char *szRep
     const char *pchLineStart = szReply;
     long long primaries = 0;
     long long replicas = 0;
+    long long handshakes = 0;
+    long long errors = 0;
     while (SAFETY_CHECK_POINTER(pchLineStart) && *pchLineStart != '\0') {
         // Loop Each Line
         const char *pchLineEnd = pchLineStart;
@@ -450,6 +471,10 @@ void handle_cluster_nodes_response(struct RedisModuleCtx *ctx, const char *szRep
             ++primaries;
         } else if (std::string::npos != line.find("slave")) {
             ++replicas;
+        } else if (std::string::npos != line.find("handshake")) {
+            ++handshakes;
+        } else if (std::string::npos != line.find("fail?") || std::string::npos != line.find("fail") || std::string::npos != line.find("noaddr") || std::string::npos != line.find("nofailover") || std::string::npos != line.find("noflags") ) {
+            ++errors;
         } else {
             RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_WARNING, "Unexpected NODE format returned by \"CLUSTER NODES\" command: \"%s\"", line.c_str());
         }
@@ -481,6 +506,10 @@ void handle_cluster_nodes_response(struct RedisModuleCtx *ctx, const char *szRep
     RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"primaries\": %llu", primaries);
     g_stats->gauge("replicas", replicas);
     RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"replicas\": %llu", replicas);
+    g_stats->gauge("handshakes", handshakes);
+    RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"handshakes\": %llu", handshakes);
+    g_stats->gauge("cluster_nodes_error_nodes", errors);
+    RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"cluster_nodes_error_nodes\": %llu", errors);
     #undef SAFETY_CHECK_POINTER
 }
 
@@ -566,10 +595,8 @@ void emit_metrics_for_insufficient_replicas(struct RedisModuleCtx *ctx, long lon
     // check if the current node is a primary
     if (strncmp(role, "master", len) == 0) {
         RedisModuleCallReply *replicasReply = RedisModule_CallReplyArrayElement(reply, 2);
-        // check if there are less than 2 connected replicas
-        if (RedisModule_CallReplyLength(replicasReply) < EXPECTED_NUMBER_OF_REPLICAS) {
-            g_stats->increment("lessThanExpectedReplicas_error", 1);
-        }
+        size_t numberOfActiveReplicas = RedisModule_CallReplyLength(replicasReply);
+        g_stats->gauge("numberOfActiveReplicas", numberOfActiveReplicas);
     }
     RedisModule_FreeCallReply(reply);
 }
@@ -582,7 +609,7 @@ void event_cron_handler(struct RedisModuleCtx *ctx, RedisModuleEvent eid, uint64
         size_t startTime = ustime();
 
 #ifdef __linux__
-        /* Log CPU Usage */
+	/* Log CPU Usage */
 	static long long s_mscpuLast = 0;
 	struct rusage self_ru;
 	getrusage(RUSAGE_SELF, &self_ru);
@@ -591,7 +618,7 @@ void event_cron_handler(struct RedisModuleCtx *ctx, RedisModuleEvent eid, uint64
 			+ (self_ru.ru_stime.tv_sec * 1000) + (self_ru.ru_stime.tv_usec / 1000);
 
 
-        g_stats->gauge("cpu_load_perc", (double)(mscpuCur - s_mscpuLast) / ((curTime - lastTime)*1000), true /* prefixOnly */);   
+	g_stats->gauge("cpu_load_perc", ((double)(mscpuCur - s_mscpuLast) / ((curTime - lastTime)*1000))*100, false /* prefixOnly */);   
 	s_mscpuLast = mscpuCur;
 #endif
 
@@ -600,7 +627,7 @@ void event_cron_handler(struct RedisModuleCtx *ctx, RedisModuleEvent eid, uint64
 
         /* node name */
         if (unameResult == 0) {
-            g_stats->increment("node_name" + std::string(g_string_counter_separator) + sysName.nodename);
+            g_stats->increment("node_name" + std::string(g_string_counter_separator) + nodeName);
         }
 
         /* Log INFO Fields */
@@ -653,7 +680,7 @@ void event_cron_handler(struct RedisModuleCtx *ctx, RedisModuleEvent eid, uint64
         reply = RedisModule_Call(ctx, "dbsize", "");
         long long keys = RedisModule_CallReplyInteger(reply);
         RedisModule_FreeCallReply(reply);
-        g_stats->gauge("keys", keys);
+        g_stats->gauge("keys", keys, false /* prefixOnly */);
         RedisModule_Log(ctx, REDISMODULE_LOGLEVEL_DEBUG, "Emitting metric \"keys\": %llu", keys);
         g_stats->timing("emit_keys_metric_time_taken_us", ustime() - commandStartTime);
 
@@ -672,17 +699,20 @@ extern "C" int RedisModule_OnLoad(RedisModuleCtx *ctx, RedisModuleString **argv,
     RedisModule_AutoMemory(ctx);
     /* Use pod name if available*/
     const char *podName = getenv("POD_NAME");
+    utsname sysName;
     unameResult = uname(&sysName);
+    if (unameResult == 0) {
+        nodeName = std::string(sysName.nodename);
+        std::replace(nodeName.begin(), nodeName.end(), '.', '-');
+    }
     if (podName != nullptr) {
         m_strPrefix = podName;
+        std::replace(m_strPrefix.begin(), m_strPrefix.end(), '.', '-');
     }
-    else {
-        if (unameResult == 0) {
-            m_strPrefix = std::string(sysName.nodename);
-            unameResult = 1;
-        }
+    else if (unameResult == 0) {
+        m_strPrefix = nodeName;
+        unameResult = 1;
     }
-    std::replace(m_strPrefix.begin(), m_strPrefix.end(), '.', '-');
 
     for (int iarg = 0; iarg < argc; ++iarg) {
         size_t len = 0;

src/networking.cpp

@@ -38,6 +38,7 @@
 #include <vector>
 #include <mutex>
 #include "aelocker.h"
+#include <sys/socket.h>
 
 static void setProtocolError(const char *errstr, client *c);
 __thread int ProcessingEventsWhileBlocked = 0; /* See processEventsWhileBlocked(). */
@@ -163,7 +164,6 @@ client *createClient(connection *conn, int iel) {
     c->multibulklen = 0;
     c->bulklen = -1;
     c->sentlen = 0;
-    c->sentlenAsync = 0;
     c->flags = 0;
     c->fPendingAsyncWrite = FALSE;
     c->fPendingAsyncWriteHandler = FALSE;
@@ -1174,6 +1174,32 @@ int chooseBestThreadForAccept()
     return ielMinLoad;
 }
 
+bool checkOverloadIgnorelist(connection *conn) {
+    if (conn->flags & CONN_FLAG_IGNORE_OVERLOAD) {
+        return true;
+    }
+    struct sockaddr_storage sa;
+    socklen_t salen = sizeof(sa);
+    if (getpeername(conn->fd, (struct sockaddr *)&sa, &salen) == -1) {
+        return false;
+    }
+    if (sa.ss_family == AF_INET) {
+        struct sockaddr_in *s = (struct sockaddr_in *)&sa;
+        for (auto ignore : g_pserver->overload_ignorelist) {
+            if (ignore.match(s->sin_addr))
+                return true;
+        }
+    }
+    if (sa.ss_family == AF_INET6) {
+        struct sockaddr_in6 *s = (struct sockaddr_in6 *)&sa;
+        for (auto ignore : g_pserver->overload_ignorelist_ipv6) {
+            if (ignore.match(s->sin6_addr))
+                return true;
+        }
+    }
+    return false;
+}
+
 void clientAcceptHandler(connection *conn) {
     client *c = (client*)connGetPrivateData(conn);
 
@@ -1205,7 +1231,6 @@ void clientAcceptHandler(connection *conn) {
     {
         char cip[NET_IP_STR_LEN+1] = { 0 };
         connPeerToString(conn, cip, sizeof(cip)-1, NULL);
-
         if (strcmp(cip,"127.0.0.1") && strcmp(cip,"::1")) {
             const char *err =
                 "-DENIED KeyDB is running in protected mode because protected "
@@ -1237,6 +1262,14 @@ void clientAcceptHandler(connection *conn) {
         }
     }
 
+    if (checkOverloadIgnorelist(conn))
+    {
+        c->flags |= CLIENT_IGNORE_OVERLOAD;
+        char cip[NET_IP_STR_LEN+1] = { 0 };
+        connPeerToString(conn, cip, sizeof(cip)-1, NULL);
+        serverLog(LL_NOTICE, "Ignoring client from %s for loadshedding", cip);
+    }
+
     g_pserver->stat_numconnections++;
     moduleFireServerEvent(REDISMODULE_EVENT_CLIENT_CHANGE,
                           REDISMODULE_SUBEVENT_CLIENT_CHANGE_CONNECTED,
@@ -1594,6 +1627,7 @@ void unlinkClient(client *c) {
     }
 
     serverTL->vecclientsProcess.erase(std::remove(serverTL->vecclientsProcess.begin(), serverTL->vecclientsProcess.end(), c), serverTL->vecclientsProcess.end());
+    serverTL->setclientsPrefetch.erase(c);
 
     /* Clear the tracking status. */
     if (c->flags & CLIENT_TRACKING) disableTracking(c);
@@ -1607,7 +1641,7 @@ bool freeClient(client *c) {
 
     /* If a client is protected, yet we need to free it right now, make sure
      * to at least use asynchronous freeing. */
-    if (c->flags & CLIENT_PROTECTED || c->casyncOpsPending || c->replstate == SLAVE_STATE_FASTSYNC_TX) {
+    if (c->flags & CLIENT_PROTECTED || c->casyncOpsPending || c->replstate == SLAVE_STATE_FASTSYNC_TX || c->btype == BLOCKED_STORAGE) {
         freeClientAsync(c);
         return false;
     }
@@ -1838,21 +1872,23 @@ int writeToClient(client *c, int handler_installed) {
 
         while (clientHasPendingReplies(c)) {
             long long repl_end_idx = getReplIndexFromOffset(c->repl_end_off);
+            long long repl_max_writes = g_pserver->repl_backlog_max_writes_per_event > 0 ? g_pserver->repl_backlog_max_writes_per_event : LLONG_MAX;
+
             serverAssert(c->repl_curr_off != -1);
 
-            if (c->repl_curr_off != c->repl_end_off){
+            if (c->repl_curr_off != c->repl_end_off) {
                 long long repl_curr_idx = getReplIndexFromOffset(c->repl_curr_off); 
                 long long nwritten2ndStage = 0; /* How much was written from the start of the replication backlog
                                                 * in the event of a wrap around write */
                 /* normal case with no wrap around */
-                if (repl_end_idx >= repl_curr_idx){
-                    nwritten = connWrite(c->conn, g_pserver->repl_backlog + repl_curr_idx, repl_end_idx - repl_curr_idx);
+                if (repl_end_idx >= repl_curr_idx) {
+                    nwritten = connWrite(c->conn, g_pserver->repl_backlog + repl_curr_idx, std::min(repl_max_writes, repl_end_idx - repl_curr_idx));
                 /* wrap around case */
                 } else {
-                    nwritten = connWrite(c->conn, g_pserver->repl_backlog + repl_curr_idx, g_pserver->repl_backlog_size - repl_curr_idx);
+                    nwritten = connWrite(c->conn, g_pserver->repl_backlog + repl_curr_idx, std::min(repl_max_writes, g_pserver->repl_backlog_size - repl_curr_idx));
                     /* only attempt wrapping if we write the correct number of bytes */
                     if (nwritten == g_pserver->repl_backlog_size - repl_curr_idx){
-                        nwritten2ndStage = connWrite(c->conn, g_pserver->repl_backlog, repl_end_idx);
+                        nwritten2ndStage = connWrite(c->conn, g_pserver->repl_backlog, std::min(repl_max_writes - nwritten, repl_end_idx));
                         if (nwritten2ndStage != -1)
                             nwritten += nwritten2ndStage;
                     }                
@@ -1869,6 +1905,8 @@ int writeToClient(client *c, int handler_installed) {
                 if (nwritten2ndStage == -1) nwritten = -1;
                 if (nwritten == -1)
                     break;
+
+                if (totwritten > g_pserver->repl_backlog_max_writes_per_event) break;
             } else {
                 break;
             }
@@ -1939,7 +1977,7 @@ int writeToClient(client *c, int handler_installed) {
     if (nwritten == -1) {
         if (connGetState(c->conn) != CONN_STATE_CONNECTED) {
             serverLog(LL_VERBOSE,
-                "Error writing to client: %s", connGetLastError(c->conn));
+                "Error writing to client: %s (totwritten: %zd)", connGetLastError(c->conn), totwritten);
             freeClientAsync(c);
             
             return C_ERR;
@@ -2760,9 +2798,9 @@ void readQueryFromClient(connection *conn) {
         return;
     }
 
-    if (cserver.cthreads > 1 || g_pserver->m_pstorageFactory) {
+    if (cserver.cthreads > 1 || g_pserver->m_pstorageFactory || g_pserver->is_overloaded) {
         parseClientCommandBuffer(c);
-        if (g_pserver->enable_async_commands && !serverTL->disable_async_commands && listLength(g_pserver->monitors) == 0 && (aeLockContention() || serverTL->rgdbSnapshot[c->db->id] || g_fTestMode) && !serverTL->in_eval && !serverTL->in_exec) {
+        if (g_pserver->enable_async_commands && !serverTL->disable_async_commands && listLength(g_pserver->monitors) == 0 && (aeLockContention() || serverTL->rgdbSnapshot[c->db->id] || g_fTestMode) && !serverTL->in_eval && !serverTL->in_exec && !g_pserver->is_overloaded) {
             // Frequent writers aren't good candidates for this optimization, they cause us to renew the snapshot too often
             //  so we exclude them unless the snapshot we need already exists.
             // Note: In test mode we want to create snapshots as often as possibl to excercise them - we don't care about perf
@@ -2774,8 +2812,20 @@ void readQueryFromClient(connection *conn) {
                 processInputBuffer(c, false, CMD_CALL_SLOWLOG | CMD_CALL_STATS | CMD_CALL_ASYNC);
             }
         }
-        if (!c->vecqueuedcmd.empty())
-            serverTL->vecclientsProcess.push_back(c);
+        if (!c->vecqueuedcmd.empty()) {
+            if (g_pserver->m_pstorageFactory != nullptr && g_pserver->prefetch_enabled && !(c->flags & CLIENT_BLOCKED)) {
+                serverTL->setclientsPrefetch.insert(c);
+            } else {
+                if (g_pserver->is_overloaded && !(c->flags & (CLIENT_MASTER | CLIENT_SLAVE | CLIENT_PENDING_WRITE | CLIENT_PUBSUB | CLIENT_BLOCKED | CLIENT_IGNORE_OVERLOAD)) && ((random() % 100) < g_pserver->overload_protect_strength)) {
+                    for (unsigned i = 0; i < c->vecqueuedcmd.size(); i++) {
+                        addReply(c, shared.overloaderr);
+                    }
+                    c->vecqueuedcmd.clear();
+                } else {
+                    serverTL->vecclientsProcess.push_back(c);
+                }
+            }
+        }
     } else {
         // If we're single threaded its actually better to just process the command here while the query is hot in the cache
         //  multithreaded lock contention dominates and batching is better
@@ -2789,9 +2839,9 @@ void processClients()
 {
     serverAssert(GlobalLocksAcquired());
 
-    // Note that this function is reentrant and vecclients may be modified by code called from processInputBuffer
+    // Note that this function is reentrant and vecclientsProcess may be modified by code called from processInputBuffer
     while (!serverTL->vecclientsProcess.empty()) {
-        client *c = serverTL->vecclientsProcess.front();
+        client *c = *serverTL->vecclientsProcess.begin();
         serverTL->vecclientsProcess.erase(serverTL->vecclientsProcess.begin());
 
         /* There is more data in the client input buffer, continue parsing it

src/object.cpp

@@ -46,6 +46,7 @@ robj *createObject(int type, void *ptr) {
     char *oB = (char*)zcalloc(sizeof(robj)+mvccExtraBytes, MALLOC_SHARED);
     robj *o = reinterpret_cast<robj*>(oB + mvccExtraBytes);
     
+    new (o) redisObject;
     o->type = type;
     o->encoding = OBJ_ENCODING_RAW;
     o->m_ptr = ptr;
@@ -418,6 +419,7 @@ void decrRefCount(robj_roptr o) {
         case OBJ_NESTEDHASH: freeNestedHashObject(o); break;
         default: serverPanic("Unknown object type"); break;
         }
+        o->~redisObject();
         if (g_pserver->fActiveReplica) {
             zfree(reinterpret_cast<redisObjectExtended*>(o.unsafe_robjcast())-1);
         } else {

src/rdb.cpp

@@ -2542,7 +2542,7 @@ robj *rdbLoadObject(int rdbtype, rio *rdb, sds key, int *error, uint64_t mvcc_ts
          * encoding version in the lower 10 bits of the module ID. */
         void *ptr = mt->rdb_load(&io,moduleid&1023);
         if (io.ctx) {
-            moduleFreeContext(io.ctx);
+            moduleFreeContext(io.ctx, false /* propogate */);
             zfree(io.ctx);
         }
 
@@ -2865,6 +2865,12 @@ public:
 
     size_t endWork() {
         if (!fLaunched) {
+            for (int idb = 0; idb < cserver.dbnum; ++idb) {
+                if (g_pserver->m_pstorageFactory) {
+                    g_pserver->db[idb]->processChangesAsync(cstorageWritesInFlight);
+                }
+            }
+            while (cstorageWritesInFlight > 0);
             return ckeysLoaded;
         }
         if (!vecbatch.empty()) {

src/readwritelock.h

@@ -8,38 +8,46 @@ class readWriteLock {
     int m_readCount = 0;
     int m_writeCount = 0;
     bool m_writeWaiting = false;
-    bool m_notify = true;
+    bool m_multi = true;
 public:
     readWriteLock(const char *name) : m_readLock(name), m_writeLock(name) {}
 
     void acquireRead() {
-        std::unique_lock<fastlock> rm(m_readLock);
-        while (m_writeCount > 0 || m_writeWaiting)
-            m_cv.wait(rm);
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
+        if (m_multi) {
+            rm.lock();
+            while (m_writeCount > 0 || m_writeWaiting)
+                m_cv.wait(rm);
+        }
         m_readCount++;
     }
     
     bool tryAcquireRead() {
         std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
-        if (!rm.try_lock())
-            return false;
-        if (m_writeCount > 0 || m_writeWaiting)
-            return false;
+        if (m_multi) {
+            if (!rm.try_lock())
+                return false;
+            if (m_writeCount > 0 || m_writeWaiting)
+                return false;
+        }
         m_readCount++;
         return true;
     }
 
     void acquireWrite(bool exclusive = true) {
-        std::unique_lock<fastlock> rm(m_readLock);
-        m_writeWaiting = true;
-        while (m_readCount > 0)
-            m_cv.wait(rm);
-        if (exclusive) {
-            /* Another thread might have the write lock while we have the read lock
-               but won't be able to release it until they can acquire the read lock
-               so release the read lock and try again instead of waiting to avoid deadlock */
-            while(!m_writeLock.try_lock())
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
+        if (m_multi) {
+            rm.lock();
+            m_writeWaiting = true;
+            while (m_readCount > 0)
                 m_cv.wait(rm);
+            if (exclusive) {
+                /* Another thread might have the write lock while we have the read lock
+                but won't be able to release it until they can acquire the read lock
+                so release the read lock and try again instead of waiting to avoid deadlock */
+                while(!m_writeLock.try_lock())
+                    m_cv.wait(rm);
+            }
         }
         m_writeCount++;
         m_writeWaiting = false;
@@ -52,32 +60,38 @@ public:
 
     bool tryAcquireWrite(bool exclusive = true) {
         std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
-        if (!rm.try_lock())
-            return false;
-        if (m_readCount > 0)
-            return false;
-        if (exclusive)
-            if (!m_writeLock.try_lock())
+        if (m_multi) {
+            if (!rm.try_lock())
                 return false;
+            if (m_readCount > 0)
+                return false;
+            if (exclusive)
+                if (!m_writeLock.try_lock())
+                    return false;
+        }
         m_writeCount++;
         return true;
     }
 
     void releaseRead() {
-        std::unique_lock<fastlock> rm(m_readLock);
-        m_readCount--;
-        if (m_notify)
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
+        if (m_multi) {
+            rm.lock();
             m_cv.notify_all();
+        }
+        m_readCount--;
     }
 
     void releaseWrite(bool exclusive = true) {
-        std::unique_lock<fastlock> rm(m_readLock);
+        std::unique_lock<fastlock> rm(m_readLock, std::defer_lock);
         serverAssert(m_writeCount > 0);
-        if (exclusive)
-            m_writeLock.unlock();
-        m_writeCount--;
-        if (m_notify)
+        if (m_multi) {
+            rm.lock();
+            if (exclusive)
+                m_writeLock.unlock();
             m_cv.notify_all();
+        }
+        m_writeCount--;
     }
 
     void downgradeWrite(bool exclusive = true) {
@@ -85,8 +99,8 @@ public:
         acquireRead();
     }
 
-    void setNotify(bool notify) {
-        m_notify = notify;
+    void setMulti(bool multi) {
+        m_multi = multi;
     }
 
     bool hasReader() {

src/redis-cli.c

@@ -2524,7 +2524,10 @@ cleanup:
 
 static int clusterManagerNodeConnect(clusterManagerNode *node) {
     if (node->context) redisFree(node->context);
-    node->context = redisConnect(node->ip, node->port);
+    struct timeval tv;
+    tv.tv_sec = config.cluster_manager_command.timeout / 1000;
+    tv.tv_usec = (config.cluster_manager_command.timeout % 1000) * 1000;
+    node->context = redisConnectWithTimeout(node->ip, node->port, tv);
     if (!node->context->err && config.tls) {
         const char *err = NULL;
         if (cliSecureConnection(node->context, config.sslconfig, &err) == REDIS_ERR && err) {
@@ -5626,7 +5629,10 @@ static int clusterManagerCommandImport(int argc, char **argv) {
     char *reply_err = NULL;
     redisReply *src_reply = NULL;
     // Connect to the source node.
-    redisContext *src_ctx = redisConnect(src_ip, src_port);
+    struct timeval tv;
+    tv.tv_sec = config.cluster_manager_command.timeout / 1000;
+    tv.tv_usec = (config.cluster_manager_command.timeout % 1000) * 1000;
+    redisContext *src_ctx = redisConnectWithTimeout(src_ip, src_port, tv);
     if (src_ctx->err) {
         success = 0;
         fprintf(stderr,"Could not connect to KeyDB at %s:%d: %s.\n", src_ip,

src/replication.cpp

@@ -188,6 +188,8 @@ int bg_unlink(const char *filename) {
 /* ---------------------------------- MASTER -------------------------------- */
 
 bool createDiskBacklog() {
+    if (g_pserver->repl_backlog_disk != nullptr)
+        return true;    // already exists
     // Lets create some disk backed pages and add them here
     std::string path = "./repl-backlog-temp" + std::to_string(gettid());
 #if (defined __APPLE__ || defined __FreeBSD__)
@@ -281,7 +283,7 @@ void resizeReplicationBacklog(long long newsize) {
             
             if (cserver.repl_backlog_disk_size != 0) {
                 if (newsize > g_pserver->repl_backlog_config_size || cserver.force_backlog_disk) {
-                    if (g_pserver->repl_backlog == g_pserver->repl_backlog_disk)
+                    if (g_pserver->repl_backlog_disk == nullptr || g_pserver->repl_backlog == g_pserver->repl_backlog_disk)
                         return; // Can't do anything more
                     serverLog(LL_NOTICE, "Switching to disk backed replication backlog due to exceeding memory limits");
                     backlog = g_pserver->repl_backlog_disk;
@@ -306,10 +308,25 @@ void resizeReplicationBacklog(long long newsize) {
                 auto cbActiveBacklog = cbPhase1 + g_pserver->repl_backlog_idx;
                 serverAssert(g_pserver->repl_backlog_histlen == cbActiveBacklog);
             }
-            if (g_pserver->repl_backlog != g_pserver->repl_backlog_disk)
+            if (g_pserver->repl_backlog != g_pserver->repl_backlog_disk) {
                 zfree(g_pserver->repl_backlog);
-            else
+            } else {
                 serverLog(LL_NOTICE, "Returning to memory backed replication backlog");
+                auto repl_backlog = g_pserver->repl_backlog_disk;
+                g_pserver->repl_backlog_disk = nullptr;
+                g_pserver->asyncworkqueue->AddWorkFunction([repl_backlog, size = cserver.repl_backlog_disk_size]{
+                    // The kernel doesn't make promises with how it will manage the memory but users really want to
+                    //  see the RSS go down.  So lets encourage that to happen.
+                    madvise(repl_backlog, size, MADV_DONTNEED); // NOTE: This will block until all pages are released
+                    aeAcquireLock();
+                    if (g_pserver->repl_backlog_disk == nullptr && cserver.repl_backlog_disk_size == size) {
+                        g_pserver->repl_backlog_disk = repl_backlog;
+                    } else {
+                        munmap(g_pserver->repl_backlog_disk, size);
+                    }
+                    aeReleaseLock();
+                }, true /*fHiPri*/);
+            }
             g_pserver->repl_backlog = backlog;
             g_pserver->repl_backlog_idx = g_pserver->repl_backlog_histlen;
             if (g_pserver->repl_batch_idxStart >= 0) {
@@ -329,6 +346,17 @@ void resizeReplicationBacklog(long long newsize) {
             /* Next byte we have is... the next since the buffer is empty. */
             g_pserver->repl_backlog_off = g_pserver->master_repl_offset+1;
             g_pserver->repl_backlog_start = g_pserver->master_repl_offset;
+            listIter li;
+            listNode *ln;
+            listRewind(g_pserver->slaves, &li);
+            while ((ln = listNext(&li))) {
+                client *replica = (client*)listNodeValue(ln);
+
+                std::unique_lock<fastlock> ul(replica->lock);
+
+                replica->repl_curr_off = g_pserver->master_repl_offset;
+                replica->repl_end_off = g_pserver->master_repl_offset;
+            }
         }
     }
     g_pserver->repl_backlog_size = newsize;
@@ -1017,8 +1045,6 @@ public:
     void addReplica(client *replica) {
         replicas.push_back(replica);
         replicationSetupSlaveForFullResync(replica,getPsyncInitialOffset());
-        // Optimize the socket for bulk transfer
-        //connDisableTcpNoDelay(replica->conn);
     }
 
     bool isActive() const { return !replicas.empty(); }
@@ -1033,7 +1059,7 @@ public:
             while (checkClientOutputBufferLimits(replica)
               && (replica->flags.load(std::memory_order_relaxed) & CLIENT_CLOSE_ASAP) == 0) {
                 ul.unlock();
-                usleep(0);
+                usleep(1000);   // give 1ms for the I/O before we poll again
                 ul.lock();
             }
         }
@@ -1091,7 +1117,7 @@ public:
         reply->used += size;
     }
 
-    void addLongWithPrefix(long val, char prefix) {
+    void addLongLongWithPrefix(long long val, char prefix) {
         char buf[128];
         int len;
 
@@ -1103,15 +1129,19 @@ public:
     }
 
     void addArrayLen(int len) {
-        addLongWithPrefix(len, '*');
+        addLongLongWithPrefix(len, '*');
     }
 
     void addLong(long val) {
-        addLongWithPrefix(val, ':');
+        addLongLongWithPrefix(val, ':');
+    }
+
+    void addLongLong(long long val) {
+        addLongLongWithPrefix(val, ':');
     }
 
     void addString(const char *s, unsigned long len) {
-        addLongWithPrefix(len, '$');
+        addLongLongWithPrefix(len, '$');
         addData(s, len);
         addData("\r\n", 2);
     }
@@ -1127,11 +1157,24 @@ public:
         }
     }
 
-    void putSlavesOnline() {
+    void putReplicasOnline() {
+        for (auto replica : replicas) {
+            std::unique_lock<fastlock> ul(replica->lock);
+            //  If we put the replica online before the output is drained then it will get immediately closed
+            while (checkClientOutputBufferLimits(replica)
+              && (replica->flags.load(std::memory_order_relaxed) & CLIENT_CLOSE_ASAP) == 0) {
+                ul.unlock();
+                usleep(1000);   // give 1ms for the I/O before we poll again
+                ul.lock();
+            }
+        }
+
+        aeAcquireLock();
         for (auto replica : replicas) {
             replica->replstate = SLAVE_STATE_FASTSYNC_DONE;
             replica->repl_put_online_on_ack = 1;
         }
+        aeReleaseLock();
     }
 
     void abort() {
@@ -1147,22 +1190,19 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
     // TODO: This needs to be on a background thread
     int retval = C_OK;
     serverAssert(GlobalLocksAcquired());
-    serverLog(LL_NOTICE, "Starting storage provider fast full sync with target: %s", "disk");
+    serverLog(LL_NOTICE, "Starting fast full sync with target: %s", "disk");
 
     std::shared_ptr<replicationBuffer> spreplBuf = std::make_shared<replicationBuffer>();
     listNode *ln;
     listIter li;
-    client *replica = nullptr;
     listRewind(g_pserver->slaves, &li);
-    while (replica == nullptr && (ln = listNext(&li))) {
+    while ((ln = listNext(&li))) {
         client *replicaCur = (client*)listNodeValue(ln);
-        if ((replicaCur->slave_capa & SLAVE_CAPA_ROCKSDB_SNAPSHOT) && (replicaCur->replstate == SLAVE_STATE_WAIT_BGSAVE_START)) {
-            replica = replicaCur;
-            spreplBuf->addReplica(replica);
+        if ((replicaCur->slave_capa & SLAVE_CAPA_KEYDB_FASTSYNC) && (replicaCur->replstate == SLAVE_STATE_WAIT_BGSAVE_START)) {
+            spreplBuf->addReplica(replicaCur);
             replicaCur->replstate = SLAVE_STATE_FASTSYNC_TX;
         }
     }
-    serverAssert(replica != nullptr);
 
     spreplBuf->addArrayLen(2);   // Two sections: Metadata and databases
 
@@ -1177,7 +1217,7 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
             spreplBuf->addString(rsi->repl_id, CONFIG_RUN_ID_SIZE);
         spreplBuf->addArrayLen(2);
             spreplBuf->addString("repl-offset", 11);
-            spreplBuf->addLong(rsi->master_repl_offset);
+            spreplBuf->addLongLong(rsi->master_repl_offset);
 
     if (dictSize(g_pserver->lua_scripts)) {
         dictEntry *de;
@@ -1193,13 +1233,13 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
         di = NULL; /* So that we don't release it again on error. */
     }
 
-    std::shared_ptr<std::vector<std::unique_ptr<const StorageCache>>> spvecspsnapshot = std::make_shared<std::vector<std::unique_ptr<const StorageCache>>>();
+    std::shared_ptr<std::vector<const redisDbPersistentDataSnapshot*>> spvecsnapshot = std::make_shared<std::vector<const redisDbPersistentDataSnapshot*>>();
     for (int idb = 0; idb < cserver.dbnum; ++idb) {
-        spvecspsnapshot->emplace_back(g_pserver->db[idb]->CloneStorageCache());
+        spvecsnapshot->emplace_back(g_pserver->db[idb]->createSnapshot(getMvccTstamp(), false));
     }
     aeReleaseLock();
 
-    g_pserver->asyncworkqueue->AddWorkFunction([spreplBuf = std::move(spreplBuf), spvecspsnapshot = std::move(spvecspsnapshot)]{
+    g_pserver->asyncworkqueue->AddWorkFunction([spreplBuf = std::move(spreplBuf), spvecsnapshot = std::move(spvecsnapshot)]{
         int retval = C_OK;
         auto timeStart = ustime();
         auto lastLogTime = timeStart;
@@ -1209,15 +1249,16 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
         // Databases
         replBuf.addArrayLen(cserver.dbnum);
         for (int idb = 0; idb < cserver.dbnum; ++idb) {
-            auto &spsnapshot = (*spvecspsnapshot)[idb];
-            size_t snapshotDeclaredCount = spsnapshot->count();
+            auto &spsnapshot = (*spvecsnapshot)[idb];
+            size_t snapshotDeclaredCount = spsnapshot->size();
             replBuf.addArrayLen(snapshotDeclaredCount);
             size_t count = 0;
-            bool result = spsnapshot->enumerate([&replBuf, &count, &cbData, &lastLogTime, &cbLastUpdate](const char *rgchKey, size_t cchKey, const void *rgchVal, size_t cchVal) -> bool{
+            bool result = spsnapshot->iterate_threadsafe([&replBuf, &count, &cbData, &lastLogTime, &cbLastUpdate](const char *strKey, robj_roptr o) -> bool{
                 replBuf.addArrayLen(2);
 
-                replBuf.addString(rgchKey, cchKey);
-                replBuf.addString((const char *)rgchVal, cchVal);
+                replBuf.addString(strKey, sdslen(strKey));
+                sds strT = serializeStoredObjectAndExpire(o);
+                replBuf.addString(strT, sdslen(strT));
                 ++count;
                 if ((count % 8092) == 0) {
                     auto curTime = ustime();
@@ -1228,7 +1269,8 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
                         lastLogTime = ustime();
                     }
                 }
-                cbData += cchKey + cchVal;
+                cbData += sdslen(strKey) + sdslen(strT);
+                sdsfree(strT);
                 return replBuf.isActive();
             });
 
@@ -1253,9 +1295,7 @@ int rdbSaveSnapshotForReplication(rdbSaveInfo *rsi) {
         auto usec = ustime() - timeStart;
         serverLog(LL_NOTICE, "Transferred %zuMB total (%zuMB data) in %.2f seconds.  (%.2fGbit/s)", spreplBuf->cbWritten()/1024/1024, cbData/1024/1024, usec/1000000.0, (spreplBuf->cbWritten()*8.0)/(usec/1000000.0)/1000000000.0);
         if (retval == C_OK) {
-            aeAcquireLock();
-            replBuf.putSlavesOnline();
-            aeReleaseLock();
+            replBuf.putReplicasOnline();
         }
     });
 
@@ -1300,7 +1340,7 @@ int startBgsaveForReplication(int mincapa) {
     /* Only do rdbSave* when rsiptr is not NULL,
      * otherwise replica will miss repl-stream-db. */
     if (rsiptr) {
-        if (mincapa & SLAVE_CAPA_ROCKSDB_SNAPSHOT && g_pserver->m_pstorageFactory)
+        if (mincapa & SLAVE_CAPA_KEYDB_FASTSYNC && FFastSyncEnabled())
             retval = rdbSaveSnapshotForReplication(rsiptr);
         else if (socket_target)
             retval = rdbSaveToSlavesSockets(rsiptr);
@@ -1479,7 +1519,7 @@ void syncCommand(client *c) {
     }
 
     /* CASE 0: Fast Sync */
-    if ((c->slave_capa & SLAVE_CAPA_ROCKSDB_SNAPSHOT) && g_pserver->m_pstorageFactory) {
+    if (c->slave_capa & SLAVE_CAPA_KEYDB_FASTSYNC && FFastSyncEnabled()) {
         serverLog(LL_NOTICE,"Fast SYNC on next replication cycle");
     /* CASE 1: BGSAVE is in progress, with disk target. */
     } else if (g_pserver->FRdbSaveInProgress() &&
@@ -1670,8 +1710,8 @@ void replconfCommand(client *c) {
                 c->slave_capa |= SLAVE_CAPA_PSYNC2;
             else if (!strcasecmp((const char*)ptrFromObj(c->argv[j+1]), "activeExpire"))
                 c->slave_capa |= SLAVE_CAPA_ACTIVE_EXPIRE;
-            else if (!strcasecmp((const char*)ptrFromObj(c->argv[j+1]), "rocksdb-snapshot-load"))
-                c->slave_capa |= SLAVE_CAPA_ROCKSDB_SNAPSHOT;
+            else if (!strcasecmp((const char*)ptrFromObj(c->argv[j+1]), "keydb-fastsync"))
+                c->slave_capa |= SLAVE_CAPA_KEYDB_FASTSYNC;
 
             fCapaCommand = true;
         } else if (!strcasecmp((const char*)ptrFromObj(c->argv[j]),"ack")) {
@@ -1736,10 +1776,14 @@ void replconfCommand(client *c) {
 
     if (fCapaCommand) {
         sds reply = sdsnew("+OK");
-        if (g_pserver->fActiveReplica)
+        if (g_pserver->fActiveReplica) {
             reply = sdscat(reply, " active-replica");
-        if (g_pserver->m_pstorageFactory && (c->slave_capa & SLAVE_CAPA_ROCKSDB_SNAPSHOT) && !g_pserver->fActiveReplica)
-            reply = sdscat(reply, " rocksdb-snapshot-save");
+        }
+        if ((c->slave_capa & SLAVE_CAPA_KEYDB_FASTSYNC) && FFastSyncEnabled()) {
+            reply = sdscat(reply, " keydb-fastsync-save");
+        } else {
+            c->slave_capa = (c->slave_capa & (~SLAVE_CAPA_KEYDB_FASTSYNC)); // never try to fast sync for this as they won't expect it
+        }
         reply = sdscat(reply, "\r\n");
         addReplySds(c, reply);
     } else {
@@ -2488,7 +2532,7 @@ size_t parseCount(const char *rgch, size_t cch, long long *pvalue) {
     return cchNumeral + 3;
 }
 
-bool readSnapshotBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi) {
+bool readFastSyncBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi) {
     int fUpdate = g_pserver->fActiveReplica || g_pserver->enable_multimaster;
     serverAssert(GlobalLocksAcquired());
     serverAssert(mi->master == nullptr);
@@ -2513,6 +2557,11 @@ bool readSnapshotBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi
         }
     }
 
+    if (mi->repl_state == REPL_STATE_WAIT_STORAGE_IO) {
+        goto LWaitIO;
+    }
+
+    serverAssert(mi->parseState != nullptr);
     for (int iter = 0; iter < 10; ++iter) {
         if (mi->parseState->shouldThrottle())
             return false;
@@ -2523,11 +2572,16 @@ bool readSnapshotBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi
 
         auto nread = connRead(conn, mi->bulkreadBuffer+qblen, readlen);
         if (nread <= 0) {
-            if (connGetState(conn) != CONN_STATE_CONNECTED)
+            if (connGetState(conn) != CONN_STATE_CONNECTED) {
+                serverLog(LL_WARNING,"I/O error trying to sync with MASTER: %s",
+                    (nread == -1) ? strerror(errno) : "connection lost");
                 cancelReplicationHandshake(mi, true);
+            }
             return false;
         }
+        g_pserver->stat_net_input_bytes += nread;
         mi->repl_transfer_lastio = g_pserver->unixtime;
+        mi->repl_transfer_read += nread;
         sdsIncrLen(mi->bulkreadBuffer,nread);
 
         size_t offset = 0;
@@ -2599,8 +2653,9 @@ bool readSnapshotBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi
             rsi.repl_stream_db = mi->parseState->getMetaDataLongLong("repl-stream-db");
             rsi.repl_offset = mi->parseState->getMetaDataLongLong("repl-offset");
             sds str = mi->parseState->getMetaDataStr("repl-id");
-            if (sdslen(str) == CONFIG_RUN_ID_SIZE+1) {
+            if (sdslen(str) == CONFIG_RUN_ID_SIZE) {
                 memcpy(rsi.repl_id, str, CONFIG_RUN_ID_SIZE+1);
+                rsi.repl_id_is_set = 1;
             }
 
             fFinished = true;
@@ -2623,7 +2678,14 @@ bool readSnapshotBulkPayload(connection *conn, redisMaster *mi, rdbSaveInfo &rsi
     if (!fFinished)
         return false;
 
+LWaitIO:
+    if (mi->parseState->hasIOInFlight()) {
+        mi->repl_state = REPL_STATE_WAIT_STORAGE_IO;
+        return false;
+    }
+
     serverLog(LL_NOTICE, "Fast sync complete");
+    serverAssert(!mi->parseState->hasIOInFlight());
     delete mi->parseState;
     mi->parseState = nullptr;
     return true;
@@ -2999,8 +3061,8 @@ void readSyncBulkPayload(connection *conn) {
         return;
     }
 
-    if (mi->isRocksdbSnapshotRepl) {
-        if (!readSnapshotBulkPayload(conn, mi, rsi))
+    if (mi->isKeydbFastsync) {
+        if (!readFastSyncBulkPayload(conn, mi, rsi))
             return;
     } else {
         if (!readSyncBulkPayloadRdb(conn, mi, rsi, usemark))
@@ -3012,17 +3074,15 @@ void readSyncBulkPayload(connection *conn) {
 
     /* Final setup of the connected slave <- master link */
     replicationCreateMasterClient(mi,mi->repl_transfer_s,rsi.repl_stream_db);
-    if (mi->isRocksdbSnapshotRepl) {
+    if (mi->isKeydbFastsync) {
         /* We need to handle the case where the initial querybuf data was read by fast sync */
         /* This should match the work readQueryFromClient would do for a master client */
         mi->master->querybuf = sdscatsds(mi->master->querybuf, mi->bulkreadBuffer);
+        mi->master->pending_querybuf = sdscatsds(mi->master->pending_querybuf, mi->bulkreadBuffer);
+        mi->master->read_reploff += sdslen(mi->bulkreadBuffer);
+
         sdsfree(mi->bulkreadBuffer);
         mi->bulkreadBuffer = nullptr;
-
-        mi->master->pending_querybuf = sdscatlen(mi->master->pending_querybuf,
-            mi->master->querybuf,sdslen(mi->master->querybuf));
-
-        mi->master->read_reploff += sdslen(mi->master->querybuf);
     }
     mi->repl_transfer_s = nullptr;
     mi->repl_state = REPL_STATE_CONNECTED;
@@ -3061,13 +3121,13 @@ void readSyncBulkPayload(connection *conn) {
     }
 
     /* Send the initial ACK immediately to put this replica in online state. */
-    if (usemark || mi->isRocksdbSnapshotRepl) replicationSendAck(mi);
+    if (usemark || mi->isKeydbFastsync) replicationSendAck(mi);
 
     /* Restart the AOF subsystem now that we finished the sync. This
      * will trigger an AOF rewrite, and when done will start appending
      * to the new file. */
     if (g_pserver->aof_enabled) restartAOFAfterSYNC();
-    if (mi->isRocksdbSnapshotRepl)
+    if (mi->isKeydbFastsync)
         readQueryFromClient(conn); // There may be querybuf data we just appeneded
     return;
 }
@@ -3393,15 +3453,15 @@ void parseMasterCapa(redisMaster *mi, sds strcapa)
     char *pchEnd = szStart;
 
     mi->isActive = false;
-    mi->isRocksdbSnapshotRepl = false;
+    mi->isKeydbFastsync = false;
     for (;;)
     {
         if (*pchEnd == ' ' || *pchEnd == '\0') {
             // Parse the word
             if (strncmp(szStart, "active-replica", pchEnd - szStart) == 0) {
                 mi->isActive = true;
-            } else if (strncmp(szStart, "rocksdb-snapshot-save", pchEnd - szStart) == 0) {
-                mi->isRocksdbSnapshotRepl = true;
+            } else if (strncmp(szStart, "keydb-fastsync-save", pchEnd - szStart) == 0) {
+                mi->isKeydbFastsync = true;
             }
             szStart = pchEnd + 1;
         }
@@ -3548,9 +3608,9 @@ retry_connect:
             "capa","psync2",
             "capa","activeExpire",
         };
-        if (g_pserver->m_pstorageFactory && !g_pserver->fActiveReplica && g_pserver->repl_diskless_load != REPL_DISKLESS_LOAD_SWAPDB) {
+        if (FFastSyncEnabled() && g_pserver->repl_diskless_load != REPL_DISKLESS_LOAD_SWAPDB) {
             veccapabilities.push_back("capa");
-            veccapabilities.push_back("rocksdb-snapshot-load");
+            veccapabilities.push_back("keydb-fastsync");
         }
 
         err = sendCommandArgv(conn, veccapabilities.size(), veccapabilities.data(), nullptr);
@@ -3733,7 +3793,7 @@ retry_connect:
     }
 
     /* Prepare a suitable temp file for bulk transfer */
-    if (!useDisklessLoad()) {
+    if (!useDisklessLoad() && !mi->isKeydbFastsync) {
         while(maxtries--) {
             auto dt = std::chrono::system_clock::now().time_since_epoch();
             auto dtMillisecond = std::chrono::duration_cast<std::chrono::milliseconds>(dt);
@@ -3862,6 +3922,10 @@ int cancelReplicationHandshake(redisMaster *mi, int reconnect) {
         delete mi->parseState;
         mi->parseState = nullptr;
     }
+    if (mi->bulkreadBuffer) {
+        sdsfree(mi->bulkreadBuffer);
+        mi->bulkreadBuffer = nullptr;
+    }
 
     if (mi->repl_state == REPL_STATE_TRANSFER) {
         replicationAbortSyncTransfer(mi);
@@ -4327,7 +4391,6 @@ void replicationCacheMaster(redisMaster *mi, client *c) {
     if (c->flags & CLIENT_MULTI) discardTransaction(c);
     listEmpty(c->reply);
     c->sentlen = 0;
-    c->sentlenAsync = 0;
     c->reply_bytes = 0;
     c->bufpos = 0;
     resetClient(c);
@@ -4766,6 +4829,10 @@ void replicationCron(void) {
     {
         redisMaster *mi = (redisMaster*)listNodeValue(lnMaster);
 
+        if (mi->repl_state == REPL_STATE_WAIT_STORAGE_IO && !mi->parseState->hasIOInFlight()) {
+            readSyncBulkPayload(mi->repl_transfer_s);
+        }
+
         std::unique_lock<decltype(mi->master->lock)> ulock;
         if (mi->master != nullptr)
             ulock = decltype(ulock)(mi->master->lock);
@@ -4782,7 +4849,8 @@ void replicationCron(void) {
 
         /* Bulk transfer I/O timeout? */
         if (mi->masterhost && mi->repl_state == REPL_STATE_TRANSFER &&
-            (time(NULL)-mi->repl_transfer_lastio) > g_pserver->repl_timeout)
+            (time(NULL)-mi->repl_transfer_lastio) > g_pserver->repl_timeout &&
+            !(g_pserver->loading & LOADING_REPLICATION))
         {
             serverLog(LL_WARNING,"Timeout receiving bulk data from MASTER... If the problem persists try to set the 'repl-timeout' parameter in keydb.conf to a larger value.");
             cancelReplicationHandshake(mi,true);

src/sds.h

@@ -418,7 +418,10 @@ public:
     sdsstring &operator=(const sdsstring &other)
     {
         sdsfree(m_str);
-        m_str = sdsdup(other.m_str);
+        if (other.m_str != nullptr)
+            m_str = sdsdup(other.m_str);
+        else
+            m_str = nullptr;
         return *this;
     }
 

src/server.cpp

@@ -1961,15 +1961,6 @@ void getExpansiveClientsInfo(size_t *in_usage, size_t *out_usage) {
     *out_usage = o;
 }
 
-int closeClientOnOverload(client *c) {
-    if (g_pserver->overload_closed_clients > MAX_CLIENTS_SHED_PER_PERIOD) return false;
-    if (!g_pserver->is_overloaded) return false;
-    // Don't close masters, replicas, or pub/sub clients
-    if (c->flags & (CLIENT_MASTER | CLIENT_SLAVE | CLIENT_PENDING_WRITE | CLIENT_PUBSUB | CLIENT_BLOCKED)) return false;
-    freeClient(c);
-    ++g_pserver->overload_closed_clients;
-    return true;
-}
 
 /* This function is called by serverCron() and is used in order to perform
  * operations on clients that are important to perform constantly. For instance
@@ -2041,7 +2032,6 @@ void clientsCron(int iel) {
             if (clientsCronTrackExpansiveClients(c, curr_peak_mem_usage_slot)) goto LContinue;
             if (clientsCronTrackClientsMemUsage(c)) goto LContinue;
             if (closeClientOnOutputBufferLimitReached(c, 0)) continue; // Client also free'd
-            if (closeClientOnOverload(c)) continue;
         LContinue:
             fastlock_unlock(&c->lock);
         }        
@@ -2137,7 +2127,7 @@ void databasesCron(bool fMainThread) {
                 ::dict *dict = g_pserver->db[rehash_db]->dictUnsafeKeyOnly();
                 /* Are we async rehashing? And if so is it time to re-calibrate? */
                 /* The recalibration limit is a prime number to ensure balancing across threads */
-                if (rehashes_per_ms > 0 && async_rehashes < 131 && !cserver.active_defrag_enabled && cserver.cthreads > 1 && dictSize(dict) > 2048 && dictIsRehashing(dict) && !g_pserver->loading) {
+                if (g_pserver->enable_async_rehash && rehashes_per_ms > 0 && async_rehashes < 131 && !cserver.active_defrag_enabled && cserver.cthreads > 1 && dictSize(dict) > 2048 && dictIsRehashing(dict) && !g_pserver->loading && aeLockContention() > 1) {
                     serverTL->rehashCtl = dictRehashAsyncStart(dict, rehashes_per_ms * ((1000 / g_pserver->hz) / 10));  // Estimate 10% CPU time spent in lock contention
                     if (serverTL->rehashCtl)
                         ++async_rehashes;
@@ -2601,8 +2591,13 @@ int serverCron(struct aeEventLoop *eventLoop, long long id, void *clientData) {
 
     /* Check for CPU Overload */
     run_with_period(10'000) {
+        if (g_pserver->is_overloaded) {
+            if (g_pserver->overload_protect_strength < 100)
+                g_pserver->overload_protect_strength *= 2;
+        } else {
+            g_pserver->overload_protect_strength = g_pserver->overload_protect_tenacity;
+        }
         g_pserver->is_overloaded = false;
-        g_pserver->overload_closed_clients = 0;
         static clock_t last = 0;
         if (g_pserver->overload_protect_threshold > 0) {
             clock_t cur = clock();
@@ -2610,6 +2605,7 @@ int serverCron(struct aeEventLoop *eventLoop, long long id, void *clientData) {
             perc /= cserver.cthreads;
             perc *= 100.0;
             serverLog(LL_WARNING, "CPU Used: %.2f", perc);
+            g_pserver->last_overload_cpu_reading = static_cast<float>(perc);
             if (perc > g_pserver->overload_protect_threshold) {
                 serverLog(LL_WARNING, "\tWARNING: CPU overload detected.");
                 g_pserver->is_overloaded = true;
@@ -2842,6 +2838,20 @@ void beforeSleep(struct aeEventLoop *eventLoop) {
 
     locker.arm();
 
+    for (auto *tok : serverTL->setStorageTokensProcess) {
+        tok->db->processStorageToken(tok);
+    }
+    serverTL->setStorageTokensProcess.clear();
+
+    if (g_pserver->m_pstorageFactory != nullptr && !serverTL->setclientsPrefetch.empty()) {
+        g_pserver->db[0]->prefetchKeysFlash(serverTL->setclientsPrefetch);
+        for (client *c : serverTL->setclientsPrefetch) {
+            if (!(c->flags & CLIENT_BLOCKED))
+                serverTL->vecclientsProcess.push_back(c);
+        }
+        serverTL->setclientsPrefetch.clear();
+    }
+
     /* end any snapshots created by fast async commands */
     for (int idb = 0; idb < cserver.dbnum; ++idb) {
         if (serverTL->rgdbSnapshot[idb] != nullptr && serverTL->rgdbSnapshot[idb]->FStale()) {
@@ -3056,10 +3066,8 @@ void afterSleep(struct aeEventLoop *eventLoop) {
         serverAssert(serverTL->gcEpoch.isReset());
         serverTL->gcEpoch = g_pserver->garbageCollector.startEpoch();
 
-        aeAcquireLock();
         for (int idb = 0; idb < cserver.dbnum; ++idb)
             g_pserver->db[idb]->trackChanges(false);
-        aeReleaseLock();
 
         serverTL->disable_async_commands = false;
     }
@@ -3119,7 +3127,8 @@ void createSharedObjects(void) {
         "-NOREPLICAS Not enough good replicas to write.\r\n")));
     shared.busykeyerr = makeObjectShared(createObject(OBJ_STRING,sdsnew(
         "-BUSYKEY Target key name already exists.\r\n")));
-    
+    shared.overloaderr = makeObjectShared(createObject(OBJ_STRING,sdsnew(
+        "-OVERLOAD KeyDB is overloaded.\r\n")));
 
     /* The shared NULL depends on the protocol version. */
     shared.null[0] = NULL;
@@ -3253,6 +3262,7 @@ void initMasterInfo(redisMaster *master)
     master->repl_state = REPL_STATE_NONE;
     master->repl_down_since = 0; /* Never connected, repl is down since EVER. */
     master->mvccLastSync = 0;
+    master->repl_transfer_fd = -1;
 }
 
 void initServerConfig(void) {
@@ -4039,7 +4049,7 @@ void initServer(void) {
     g_pserver->cron_malloc_stats.allocator_active = 0;
     g_pserver->cron_malloc_stats.allocator_resident = 0;
     g_pserver->cron_malloc_stats.sys_available = 0;
-    g_pserver->cron_malloc_stats.sys_total = g_pserver->force_eviction_percent ? getMemTotal() : 0;
+    g_pserver->cron_malloc_stats.sys_total = getMemTotal();
     g_pserver->lastbgsave_status = C_OK;
     g_pserver->aof_last_write_status = C_OK;
     g_pserver->aof_last_write_errno = 0;
@@ -4150,6 +4160,16 @@ void InitServerLast() {
 
     g_pserver->asyncworkqueue = new (MALLOC_LOCAL) AsyncWorkQueue(cserver.cthreads);
 
+    if (g_pserver->m_pstorageFactory != nullptr) {
+        g_pserver->asyncreadworkqueue = new (MALLOC_LOCAL) AsyncWorkQueue(cserver.cthreads);
+
+        //Process one write/commit at a time to ensure consistency
+        g_pserver->asyncwriteworkqueue = new (MALLOC_LOCAL) AsyncWorkQueue(1);
+    } else {
+        g_pserver->asyncreadworkqueue = nullptr;
+        g_pserver->asyncwriteworkqueue = nullptr;
+    }
+
     // Allocate the repl backlog
     
 }
@@ -5102,8 +5122,8 @@ int processCommand(client *c, int callFlags) {
 
             sds str = sdsempty();
             for (int j = 0; j < numkeys; j++) {
-                sdscatsds(str, (sds)ptrFromObj(c->argv[keyindex[j]]));
-                sdscat(str, " ");
+                str = sdscatsds(str, (sds)ptrFromObj(c->argv[keyindex[j]]));
+                str = sdscat(str, " ");
             }
         
             if (numkeys > 0)
@@ -5663,7 +5683,8 @@ sds genRedisInfoString(const char *section) {
             "lru_clock:%u\r\n"
             "executable:%s\r\n"
             "config_file:%s\r\n"
-            "availability_zone:%s\r\n",
+            "availability_zone:%s\r\n"
+            "features:%s\r\n",
             KEYDB_SET_VERSION,
             redisGitSHA1(),
             strtol(redisGitDirty(),NULL,10) > 0,
@@ -5690,7 +5711,8 @@ sds genRedisInfoString(const char *section) {
             lruclock,
             cserver.executable ? cserver.executable : "",
             cserver.configfile ? cserver.configfile : "",
-            g_pserver->sdsAvailabilityZone);
+            g_pserver->sdsAvailabilityZone,
+            "cluster_mget");
     }
 
     /* Clients */
@@ -5805,7 +5827,9 @@ sds genRedisInfoString(const char *section) {
             "lazyfree_pending_objects:%zu\r\n"
             "lazyfreed_objects:%zu\r\n"
             "storage_provider:%s\r\n"
-            "available_system_memory:%s\r\n",
+            "available_system_memory:%s\r\n"
+            "maxstorage:%llu\r\n"
+            "storage_used:%lu\r\n",
             zmalloc_used,
             hmem,
             g_pserver->cron_malloc_stats.process_rss,
@@ -5851,7 +5875,9 @@ sds genRedisInfoString(const char *section) {
             lazyfreeGetPendingObjectsCount(),
             lazyfreeGetFreedObjectsCount(),
             g_pserver->m_pstorageFactory ? g_pserver->m_pstorageFactory->name() : "none",
-            available_system_mem
+            available_system_mem,
+            g_pserver->maxstorage,
+            g_pserver->m_pstorageFactory ? g_pserver->m_pstorageFactory->totalDiskspaceUsed() : 0
         );
         freeMemoryOverheadData(mh);
     }
@@ -6269,13 +6295,15 @@ sds genRedisInfoString(const char *section) {
         "used_cpu_sys_children:%ld.%06ld\r\n"
         "used_cpu_user_children:%ld.%06ld\r\n"
         "server_threads:%d\r\n"
-        "long_lock_waits:%" PRIu64 "\r\n",
+        "long_lock_waits:%" PRIu64 "\r\n"
+        "last_overload_cpu_reading:%.2f\r\n",
         (long)self_ru.ru_stime.tv_sec, (long)self_ru.ru_stime.tv_usec,
         (long)self_ru.ru_utime.tv_sec, (long)self_ru.ru_utime.tv_usec,
         (long)c_ru.ru_stime.tv_sec, (long)c_ru.ru_stime.tv_usec,
         (long)c_ru.ru_utime.tv_sec, (long)c_ru.ru_utime.tv_usec,
         cserver.cthreads,
-        fastlock_getlongwaitcount());
+        fastlock_getlongwaitcount(),
+        g_pserver->last_overload_cpu_reading);
 #ifdef RUSAGE_THREAD
         struct rusage m_ru;
         getrusage(RUSAGE_THREAD, &m_ru);
@@ -7411,7 +7439,12 @@ static void validateConfiguration()
         exit(EXIT_FAILURE);
     }
 
-    g_pserver->repl_backlog_config_size = g_pserver->repl_backlog_size; // this is normally set in the update logic, but not on initial config
+    if (g_pserver->repl_diskless_load == REPL_DISKLESS_LOAD_SWAPDB && g_pserver->m_pstorageFactory) {
+        serverLog(LL_WARNING, "SWAPDB is not implemented when using a storage provider.");
+        exit(EXIT_FAILURE);
+    }
+
+    g_pserver->repl_backlog_size = g_pserver->repl_backlog_config_size; // this is normally set in the update logic, but not on initial config
 }
 
 int iAmMaster(void) {

src/server.h

@@ -64,6 +64,7 @@
 #include <string>
 #include <mutex>
 #include <unordered_set>
+#include <arpa/inet.h>
 #ifdef __cplusplus
 extern "C" {
 #include <lua.h>
@@ -125,9 +126,25 @@ typedef long long ustime_t; /* microsecond time type. */
 #define OVERLOAD_PROTECT_PERIOD_MS 10'000 // 10 seconds
 #define MAX_CLIENTS_SHED_PER_PERIOD (OVERLOAD_PROTECT_PERIOD_MS / 10)  // Restrict to one client per 10ms
 
+#define IPV4_BITS 32
+#define IPV6_BITS 128
+
 extern int g_fTestMode;
 extern struct redisServer *g_pserver;
 
+class TCleanup {
+    std::function<void()> fn;
+
+public:
+    TCleanup(std::function<void()> fn)
+        : fn(fn)
+    {}
+
+    ~TCleanup() {
+        fn();
+    }
+};
+
 struct redisObject;
 class robj_roptr
 {
@@ -546,6 +563,7 @@ extern int configOOMScoreAdjValuesDefaults[CONFIG_OOM_COUNT];
                                           RDB without replication buffer. */
 #define CLIENT_FORCE_REPLY (1ULL<<44) /* Should addReply be forced to write the text? */
 #define CLIENT_AUDIT_LOGGING (1ULL<<45) /* Client commands required audit logging */
+#define CLIENT_IGNORE_OVERLOAD (1ULL<<46) /* Client that should not be disconnected by overload protection */
 
 /* Client block type (btype field in client structure)
  * if CLIENT_BLOCKED flag is set. */
@@ -557,7 +575,8 @@ extern int configOOMScoreAdjValuesDefaults[CONFIG_OOM_COUNT];
 #define BLOCKED_ZSET 5    /* BZPOP et al. */
 #define BLOCKED_PAUSE 6   /* Blocked by CLIENT PAUSE */
 #define BLOCKED_ASYNC 7
-#define BLOCKED_NUM 8     /* Number of blocked states. */
+#define BLOCKED_STORAGE 8
+#define BLOCKED_NUM 9     /* Number of blocked states. */
 
 /* Client request types */
 #define PROTO_REQ_INLINE 1
@@ -593,6 +612,7 @@ typedef enum {
     REPL_STATE_RECEIVE_PSYNC_REPLY, /* Wait for PSYNC reply */
     /* --- End of handshake states --- */
     REPL_STATE_TRANSFER,        /* Receiving .rdb from master */
+    REPL_STATE_WAIT_STORAGE_IO,
     REPL_STATE_CONNECTED,       /* Connected to master */
 } repl_state;
 
@@ -620,7 +640,7 @@ typedef enum {
 #define SLAVE_CAPA_EOF (1<<0)    /* Can parse the RDB EOF streaming format. */
 #define SLAVE_CAPA_PSYNC2 (1<<1) /* Supports PSYNC2 protocol. */
 #define SLAVE_CAPA_ACTIVE_EXPIRE (1<<2) /* Will the slave perform its own expirations? (Don't send delete) */
-#define SLAVE_CAPA_ROCKSDB_SNAPSHOT (1<<3)
+#define SLAVE_CAPA_KEYDB_FASTSYNC (1<<3)
 
 /* Synchronous read timeout - replica side */
 #define CONFIG_REPL_SYNCIO_TIMEOUT 5
@@ -966,6 +986,7 @@ struct redisObjectExtended {
 
 typedef struct redisObject {
     friend redisObject *createEmbeddedStringObject(const char *ptr, size_t len);
+    friend redisObject *createObject(int type, void *ptr);
 protected:
     redisObject() {}
 
@@ -1210,6 +1231,8 @@ public:
     bool keycacheIsEnabled();
 
     void prefetchKeysAsync(client *c, struct parsed_command &command);
+    void prefetchKeysFlash(const std::unordered_set<client*> &setc);
+    void processStorageToken(StorageToken *tok);
 
     bool FSnapshot() const { return m_spdbSnapshotHOLDER != nullptr; }
 
@@ -1219,6 +1242,7 @@ public:
 
     dict_iter find_cached_threadsafe(const char *key) const;
 
+    static void storageLoadCallback(struct aeEventLoop *el, struct StorageToken *token);
     static void activeExpireCycleCore(int type);
 
 protected:
@@ -1368,6 +1392,8 @@ struct redisDb : public redisDbPersistentDataSnapshot
     using redisDbPersistentData::keycacheIsEnabled;
     using redisDbPersistentData::dictUnsafeKeyOnly;
     using redisDbPersistentData::prefetchKeysAsync;
+    using redisDbPersistentData::prefetchKeysFlash;
+    using redisDbPersistentData::processStorageToken;
     using redisDbPersistentData::prepOverwriteForSnapshot;
     using redisDbPersistentData::FRehashing;
     using redisDbPersistentData::FTrackingChanges;
@@ -1609,7 +1635,6 @@ struct client {
     unsigned long long reply_bytes; /* Tot bytes of objects in reply list. */
     size_t sentlen;         /* Amount of bytes already sent in the current
                                buffer or object being sent. */
-    size_t sentlenAsync;    /* same as sentlen buf for async buffers (which are a different stream) */
     time_t ctime;           /* Client creation time. */
     long duration;          /* Current command duration. Used for measuring latency of blocking/non-blocking cmds */
     time_t lastinteraction; /* Time of the last interaction, used for timeout */
@@ -1744,7 +1769,7 @@ struct sharedObjectsStruct {
     *emptyarray, *wrongtypeerr, *nokeyerr, *syntaxerr, *sameobjecterr,
     *outofrangeerr, *noscripterr, *loadingerr, *slowscripterr, *bgsaveerr,
     *masterdownerr, *roslaveerr, *execaborterr, *noautherr, *noreplicaserr,
-    *busykeyerr, *oomerr, *plus, *messagebulk, *pmessagebulk, *subscribebulk,
+    *busykeyerr, *oomerr, *overloaderr, *plus, *messagebulk, *pmessagebulk, *subscribebulk,
     *unsubscribebulk, *psubscribebulk, *punsubscribebulk, *del, *unlink,
     *rpop, *lpop, *lpush, *rpoplpush, *lmove, *blmove, *zpopmin, *zpopmax,
     *emptyscan, *multi, *exec, *left, *right, *hset, *srem, *xgroup, *xclaim,  
@@ -1870,7 +1895,7 @@ struct redisMaster {
     long long master_initial_offset;           /* Master PSYNC offset. */
 
     bool isActive = false;
-    bool isRocksdbSnapshotRepl = false;
+    bool isKeydbFastsync = false;
     int repl_state;          /* Replication status if the instance is a replica */
     off_t repl_transfer_size; /* Size of RDB to read from master during sync. */
     off_t repl_transfer_read; /* Amount of RDB read from master during sync. */
@@ -2207,6 +2232,8 @@ struct redisServerThreadVars {
     int propagate_in_transaction = 0;  /* Make sure we don't propagate nested MULTI/EXEC */
     int client_pause_in_transaction = 0; /* Was a client pause executed during this Exec? */
     std::vector<client*> vecclientsProcess;
+    std::unordered_set<client*> setclientsPrefetch;
+    std::unordered_set<StorageToken*> setStorageTokensProcess;
     dictAsyncRehashCtl *rehashCtl = nullptr;
 
     int getRdbKeySaveDelay();
@@ -2305,6 +2332,7 @@ struct redisServer {
     rax *errors;                /* Errors table */
     int activerehashing;        /* Incremental rehash in serverCron() */
     int active_defrag_running;  /* Active defragmentation running (holds current scan aggressiveness) */
+    int enable_async_rehash = 1;    /* Should we use the async rehash feature? */
     int cronloops;              /* Number of times the cron function run */
     char runid[CONFIG_RUN_ID_SIZE+1];  /* ID always different at every exec. */
     int sentinel_mode;          /* True if this instance is a Sentinel. */
@@ -2545,6 +2573,7 @@ struct redisServer {
     int repl_diskless_sync_delay;   /* Delay to start a diskless repl BGSAVE. */
     std::atomic <long long> repl_lowest_off; /* The lowest offset amongst all replicas
                                                 -1 if there are no replicas */
+    long long repl_backlog_max_writes_per_event; /* Caps the output buffer sent to the replica at once */
     /* Replication (replica) */
     list *masters;
     int enable_multimaster; 
@@ -2686,6 +2715,7 @@ struct redisServer {
 
     int fActiveReplica;                          /* Can this replica also be a master? */
     int fWriteDuringActiveLoad;                  /* Can this active-replica write during an RDB load? */
+    int fEnableFastSync = false;
 
     // Format:
     //  Lower 20 bits: a counter incrementing for each command executed in the same millisecond
@@ -2693,6 +2723,8 @@ struct redisServer {
     uint64_t mvcc_tstamp;
 
     AsyncWorkQueue *asyncworkqueue;
+    AsyncWorkQueue *asyncreadworkqueue;
+    AsyncWorkQueue *asyncwriteworkqueue;
 
     /* System hardware info */
     size_t system_memory_size;  /* Total memory in system as reported by OS */
@@ -2711,7 +2743,95 @@ struct redisServer {
     int tls_rotation;
 
     std::set<sdsstring> tls_auditlog_blocklist; /* Certificates that can be excluded from audit logging */
+    std::set<sdsstring> tls_overload_ignorelist; /* Certificates that are be excluded load shedding */
     std::set<sdsstring> tls_allowlist;
+    class IPV4 {
+        struct in_addr m_ip;
+        struct in_addr m_mask;
+
+        int bitsFromMask() const {
+            uint32_t mask = ntohl(m_mask.s_addr);
+            int bits = 0;
+            while (mask > 0) {
+                bits += mask & 1;
+                mask >>= 1;
+            }
+            return bits;
+        }
+
+    public:
+        IPV4(struct in_addr ip, struct in_addr mask) : m_ip(ip), m_mask(mask) {};
+        bool match(struct in_addr ip) const
+        {
+            return (ip.s_addr & m_mask.s_addr) == m_ip.s_addr;
+        }
+
+        sds getString() const
+        {
+            sds result = sdsempty();
+            char buf[INET_ADDRSTRLEN];
+            inet_ntop(AF_INET, &m_ip, buf, sizeof(buf));
+            result = sdscat(result, buf);
+            int bits = bitsFromMask();
+            if (bits != IPV4_BITS) {
+                result = sdscat(result, "/");
+                result = sdscat(result, std::to_string(bits).c_str());
+            }
+            return result;
+        }
+
+        bool operator<(const IPV4& rhs) const
+        {
+            return memcmp(&m_ip, &rhs.m_ip, sizeof(m_ip)) < 0 || (memcmp(&m_ip, &rhs.m_ip, sizeof(m_ip)) == 0 && memcmp(&m_mask, &rhs.m_mask, sizeof(m_mask)) < 0);
+        }
+    };
+    class IPV6 {
+        struct in6_addr m_ip;
+        struct in6_addr m_mask;
+
+        int bitsFromMask() const {
+            int bits = 0;
+            for (unsigned int i = 0; i < sizeof(struct in6_addr); i++) {
+                uint8_t mask = m_mask.s6_addr[i];
+                while (mask > 0) {
+                    bits += mask & 1;
+                    mask >>= 1;
+                }
+            }
+            return bits;
+        }
+    public:
+        IPV6(struct in6_addr ip, struct in6_addr mask) : m_ip(ip), m_mask(mask) {};
+        bool match(struct in6_addr ip) const
+        {
+            for (unsigned int i = 0; i < sizeof(struct in6_addr); i++) {
+                if ((ip.s6_addr[i] & m_mask.s6_addr[i]) != m_ip.s6_addr[i])
+                    return false;
+            }
+            return true;
+        }
+
+        sds getString() const
+        {
+            sds result = sdsempty();
+            char buf[INET6_ADDRSTRLEN];
+            inet_ntop(AF_INET6, &m_ip, buf, sizeof(buf));
+            result = sdscat(result, buf);
+            int bits = bitsFromMask();
+            if (bits != IPV6_BITS) {
+                result = sdscat(result, "/");
+                result = sdscat(result, std::to_string(bits).c_str());
+            }
+            return result;
+        }
+
+        bool operator<(const IPV6& rhs) const
+        {
+            return memcmp(&m_ip, &rhs.m_ip, sizeof(m_ip)) < 0 || (memcmp(&m_ip, &rhs.m_ip, sizeof(m_ip)) == 0 && memcmp(&m_mask, &rhs.m_mask, sizeof(m_mask)) < 0);
+        }
+    };
+    std::set<IPV4> overload_ignorelist;
+    std::set<IPV6> overload_ignorelist_ipv6;
     redisTLSContextConfig tls_ctx_config;
 
     /* cpu affinity */
@@ -2754,12 +2874,14 @@ struct redisServer {
 
     sds sdsAvailabilityZone;
     int overload_protect_threshold = 0;
+    int overload_protect_tenacity = 0;
+    int overload_protect_strength = 0;
+    float last_overload_cpu_reading = 0.0f;
     int is_overloaded = 0;
-    int overload_closed_clients = 0;
 
-        int module_blocked_pipe[2]; /* Pipe used to awake the event loop if a
-                            client blocked on a module command needs
-                            to be processed. */
+    int module_blocked_pipe[2]; /* Pipe used to awake the event loop if a
+                                client blocked on a module command needs
+                                to be processed. */
 
     bool FRdbSaveInProgress() const { return g_pserver->rdbThreadVars.fRdbThreadActive; }
 };
@@ -2917,7 +3039,7 @@ int moduleGetCommandKeysViaAPI(struct redisCommand *cmd, robj **argv, int argc,
 moduleType *moduleTypeLookupModuleByID(uint64_t id);
 void moduleTypeNameByID(char *name, uint64_t moduleid);
 const char *moduleTypeModuleName(moduleType *mt);
-void moduleFreeContext(struct RedisModuleCtx *ctx);
+void moduleFreeContext(struct RedisModuleCtx *ctx, bool propogate = true);
 void unblockClientFromModule(client *c);
 void moduleHandleBlockedClients(int iel);
 void moduleBlockedClientTimedOut(client *c);
@@ -3180,6 +3302,7 @@ void trimStringObjectIfNeeded(robj *o);
 robj *deserializeStoredObject(const void *data, size_t cb);
 std::unique_ptr<expireEntry> deserializeExpire(const char *str, size_t cch, size_t *poffset);
 sds serializeStoredObject(robj_roptr o, sds sdsPrefix = nullptr);
+sds serializeStoredObjectAndExpire(robj_roptr o);
 
 #define sdsEncodedObject(objptr) (objptr->encoding == OBJ_ENCODING_RAW || objptr->encoding == OBJ_ENCODING_EMBSTR)
 
@@ -3969,6 +4092,10 @@ inline int ielFromEventLoop(const aeEventLoop *eventLoop)
     return iel;
 }
 
+inline bool FFastSyncEnabled() {
+    return g_pserver->fEnableFastSync && !g_pserver->fActiveReplica;
+}
+
 inline int FCorrectThread(client *c)
 {
     return (c->conn == nullptr)

src/snapshot.cpp

@@ -587,8 +587,12 @@ bool redisDbPersistentDataSnapshot::iterate_threadsafe_core(std::function<bool(c
                 if (!fKeyOnly)
                 {
                     size_t offset = 0;
-                    deserializeExpire((const char*)data, cbData, &offset);
+                    std::unique_ptr<expireEntry> spexpire = deserializeExpire((const char*)data, cbData, &offset);
                     o = deserializeStoredObject(reinterpret_cast<const char*>(data)+offset, cbData-offset);
+                    o->SetFExpires(spexpire != nullptr);
+                    if (spexpire != nullptr) {
+                        o->expire = std::move(*spexpire);
+                    }
                 }
                 fContinue = fn(sdsKey, o);
                 if (o != nullptr)

src/storage/rocksdb.cpp

@@ -25,7 +25,7 @@ bool FInternalKey(const char *key, size_t cch)
 std::string getPrefix(unsigned int hashslot)
 {
     char *hash_char = (char *)&hashslot;
-    return std::string(hash_char + (sizeof(unsigned int) - 2), 2);
+    return std::string(hash_char, 2);
 }
 
 std::string prefixKey(const char *key, size_t cchKey)
@@ -149,11 +149,22 @@ size_t RocksDBStorageProvider::clear()
 
     rocksdb::ColumnFamilyHandle *handle = nullptr;
     rocksdb::ColumnFamilyOptions cf_options(m_pfactory->RocksDbOptions());
+    cf_options.level_compaction_dynamic_level_bytes = true;
     m_spdb->CreateColumnFamily(cf_options, strName, &handle);
     m_spcolfamily = std::shared_ptr<rocksdb::ColumnFamilyHandle>(handle);
 
     if (!status.ok())
         throw status.ToString();
+   
+    status = m_spdb->DropColumnFamily(m_spexpirecolfamily.get());
+    strName = m_spexpirecolfamily->GetName();
+
+    m_spdb->CreateColumnFamily(cf_options, strName, &handle);
+    m_spexpirecolfamily = std::shared_ptr<rocksdb::ColumnFamilyHandle>(handle);
+
+    if (!status.ok())
+        throw status.ToString();
+
     m_count = 0;
     return celem;
 }
@@ -191,23 +202,26 @@ bool RocksDBStorageProvider::enumerate_hashslot(callback fn, unsigned int hashsl
     std::string prefix = getPrefix(hashslot);
     std::unique_ptr<rocksdb::Iterator> it = std::unique_ptr<rocksdb::Iterator>(m_spdb->NewIterator(ReadOptions(), m_spcolfamily.get()));
     size_t count = 0;
-    for (it->Seek(prefix.c_str()); it->Valid(); it->Next()) {
+    serverAssert(prefix.size() >= 2);
+    bool full_iter = true;
+    for (it->Seek(rocksdb::Slice(prefix.data(), prefix.size())); it->Valid(); it->Next()) {
         if (FInternalKey(it->key().data(), it->key().size()))
             continue;
-        if (strncmp(it->key().data(),prefix.c_str(),2) != 0)
+        if (it->key().size() < 2 || memcmp(it->key().data(),prefix.data(),2) != 0)
             break;
         ++count;
         bool fContinue = fn(it->key().data()+2, it->key().size()-2, it->value().data(), it->value().size());
-        if (!fContinue)
+        if (!fContinue) {
+            full_iter = false;
             break;
+        }
     }
-    bool full_iter = !it->Valid() || (strncmp(it->key().data(),prefix.c_str(),2) != 0);
     if (full_iter && count != g_pserver->cluster->slots_keys_count[hashslot])
     {
-        printf("WARNING: rocksdb hashslot count mismatch");
+        serverLog(LL_WARNING, "WARNING: rocksdb hashslot %d count mismatch %zu vs expected %lu", hashslot, count, g_pserver->cluster->slots_keys_count[hashslot]);
     }
-    assert(!full_iter || count == g_pserver->cluster->slots_keys_count[hashslot]);
-    assert(it->status().ok()); // Check for any errors found during the scan
+    serverAssert(!full_iter || count == g_pserver->cluster->slots_keys_count[hashslot]);
+    serverAssert(it->status().ok()); // Check for any errors found during the scan
     return full_iter;
 }
 
@@ -242,21 +256,14 @@ void RocksDBStorageProvider::removeExpire(const char *key, size_t cchKey, long l
         throw status.ToString();
 }
 
-std::vector<std::string> RocksDBStorageProvider::getExpirationCandidates(unsigned int count, long long time)
+std::vector<std::string> RocksDBStorageProvider::getExpirationCandidates(unsigned int count)
 {
     std::vector<std::string> result;
     std::unique_ptr<rocksdb::Iterator> it = std::unique_ptr<rocksdb::Iterator>(m_spdb->NewIterator(ReadOptions(), m_spexpirecolfamily.get()));
-    for (it->SeekToFirst(); it->Valid() && (*((long long *)it->key().data()) < time) && (result.size() < count); it->Next()) {
+    for (it->SeekToFirst(); it->Valid() && result.size() < count; it->Next()) {
         if (FInternalKey(it->key().data(), it->key().size()))
             continue;
         result.emplace_back(it->value().data(), it->value().size());
-        rocksdb::Status status;
-        if (m_spbatch)
-            status = m_spbatch->Delete(m_spexpirecolfamily.get(), it->key());
-        else
-            status = m_spdb->Delete(WriteOptions(), m_spexpirecolfamily.get(), it->key());
-        if (!status.ok())
-            throw status.ToString();
     }
     return result;
 }
@@ -277,7 +284,7 @@ std::vector<std::string> RocksDBStorageProvider::getEvictionCandidates(unsigned
         }
     } else {
         std::unique_ptr<rocksdb::Iterator> it = std::unique_ptr<rocksdb::Iterator>(m_spdb->NewIterator(ReadOptions(), m_spexpirecolfamily.get()));
-        for (it->SeekToFirst(); it->Valid() && result.size() < 16; it->Next()) {
+        for (it->SeekToFirst(); it->Valid() && result.size() < count; it->Next()) {
             if (FInternalKey(it->key().data(), it->key().size()))
                 continue;
             result.emplace_back(it->value().data(), it->value().size());
@@ -330,6 +337,31 @@ void RocksDBStorageProvider::endWriteBatch()
     m_lock.unlock();
 }
 
+struct BatchStorageToken : public StorageToken {
+    std::shared_ptr<rocksdb::DB> tspdb;    // Note: This must be first so it is deleted last
+    std::unique_ptr<rocksdb::WriteBatchWithIndex> tspbatch;
+};
+
+StorageToken* RocksDBStorageProvider::begin_endWriteBatch(struct aeEventLoop *el, aePostFunctionTokenProc* callback){
+    BatchStorageToken *tok = new BatchStorageToken();
+    tok->tspbatch = std::move(m_spbatch);
+    tok->tspdb = m_spdb;
+    m_spbatch = nullptr;
+    m_lock.unlock();
+    (*m_pfactory->m_wwqueue)->AddWorkFunction([this, el,callback,tok]{
+        tok->tspdb->Write(WriteOptions(),tok->tspbatch.get()->GetWriteBatch());
+        aePostFunction(el,callback,tok);
+    });
+
+    return tok;
+}
+
+void RocksDBStorageProvider::complete_endWriteBatch(StorageToken* tok){
+    delete tok;
+    tok = nullptr;
+}
+
+
 void RocksDBStorageProvider::batch_lock()
 {
     m_lock.lock();
@@ -354,10 +386,63 @@ bool RocksDBStorageProvider::FKeyExists(std::string& key) const
     return m_spdb->Get(ReadOptions(), m_spcolfamily.get(), rocksdb::Slice(key), &slice).ok();
 }
 
+struct RetrievalStorageToken : public StorageToken {
+    std::unordered_map<std::string, std::unique_ptr<rocksdb::PinnableSlice>> mapkeydata;
+};
+
+StorageToken *RocksDBStorageProvider::begin_retrieve(struct aeEventLoop *el, aePostFunctionTokenProc callback, sds *rgkey, size_t ckey) {
+    RetrievalStorageToken *tok = new RetrievalStorageToken();
+
+    for (size_t ikey = 0; ikey < ckey; ++ikey) {
+        tok->mapkeydata.insert(std::make_pair(std::string(rgkey[ikey], sdslen(rgkey[ikey])), nullptr));
+    }
+
+    auto opts = ReadOptions();
+    opts.async_io = true;
+    (*m_pfactory->m_rwqueue)->AddWorkFunction([this, el, callback, tok, opts]{
+        std::vector<std::string> veckeysStr;
+        std::vector<rocksdb::Slice> veckeys;
+        std::vector<rocksdb::PinnableSlice> vecvals;
+        std::vector<rocksdb::Status> vecstatus;
+        veckeys.reserve(tok->mapkeydata.size());
+        veckeysStr.reserve(tok->mapkeydata.size());
+        vecvals.resize(tok->mapkeydata.size());
+        vecstatus.resize(tok->mapkeydata.size());
+        for (auto &pair : tok->mapkeydata) {
+            veckeysStr.emplace_back(prefixKey(pair.first.data(), pair.first.size()));
+            veckeys.emplace_back(veckeysStr.back().data(), veckeysStr.back().size());
+        }
+
+        m_spdb->MultiGet(ReadOptions(),
+            m_spcolfamily.get(),
+            veckeys.size(), const_cast<const rocksdb::Slice*>(veckeys.data()),
+            vecvals.data(), vecstatus.data());
+
+        auto itrDst = tok->mapkeydata.begin();
+        for (size_t ires = 0; ires < veckeys.size(); ++ires) {
+            if (vecstatus[ires].ok()) {
+                itrDst->second = std::make_unique<rocksdb::PinnableSlice>(std::move(vecvals[ires]));
+            }
+            ++itrDst;
+        }
+        aePostFunction(el, callback, tok);
+    });
+    return tok;
+}
+
+void RocksDBStorageProvider::complete_retrieve(StorageToken *tok, callbackSingle fn) {
+    RetrievalStorageToken *rtok = reinterpret_cast<RetrievalStorageToken*>(tok);
+    for (auto &pair : rtok->mapkeydata) {
+        if (pair.second != nullptr) {
+            fn(pair.first.data(), pair.first.size(), pair.second->data(), pair.second->size());
+        }
+    }
+    delete rtok;
+}
 bool RocksDBStorageProvider::FExpireExists(std::string& key) const
 {
     rocksdb::PinnableSlice slice;
     if (m_spbatch)
         return m_spbatch->GetFromBatchAndDB(m_spdb.get(), ReadOptions(), m_spexpirecolfamily.get(), rocksdb::Slice(key), &slice).ok();
     return m_spdb->Get(ReadOptions(), m_spexpirecolfamily.get(), rocksdb::Slice(key), &slice).ok();
-}
+}

src/storage/rocksdb.h

@@ -32,11 +32,15 @@ public:
     virtual void insert(const char *key, size_t cchKey, void *data, size_t cb, bool fOverwrite) override;
     virtual bool erase(const char *key, size_t cchKey) override;
     virtual void retrieve(const char *key, size_t cchKey, callbackSingle fn) const override;
+
+    virtual StorageToken *begin_retrieve(struct aeEventLoop *el, aePostFunctionTokenProc callback, sds *rgkey, size_t ckey);
+    virtual void complete_retrieve(StorageToken *tok, callbackSingle fn);
+
     virtual size_t clear() override;
     virtual bool enumerate(callback fn) const override;
     virtual bool enumerate_hashslot(callback fn, unsigned int hashslot) const override;
 
-    virtual std::vector<std::string> getExpirationCandidates(unsigned int count, long long time) override;
+    virtual std::vector<std::string> getExpirationCandidates(unsigned int count) override;
     virtual std::vector<std::string> getEvictionCandidates(unsigned int count) override;
     virtual void setExpire(const char *key, size_t cchKey, long long expire) override;
     virtual void removeExpire(const char *key, size_t cchKey, long long expire) override;
@@ -45,6 +49,8 @@ public:
 
     virtual void beginWriteBatch() override;
     virtual void endWriteBatch() override;
+    virtual StorageToken* begin_endWriteBatch(struct aeEventLoop *el, aePostFunctionTokenProc* proc);
+    virtual void complete_endWriteBatch(StorageToken *tok);
 
     virtual void bulkInsert(char **rgkeys, size_t *rgcbkeys, char **rgvals, size_t *rgcbvals, size_t celem) override;
 

src/storage/rocksdbfactor_internal.h

@@ -1,5 +1,6 @@
 #pragma once
 #include "rocksdb.h"
+#include "../AsyncWorkQueue.h"
 
 class RocksDBStorageFactory : public IStorageFactory
 {
@@ -11,7 +12,10 @@ class RocksDBStorageFactory : public IStorageFactory
     bool m_fCreatedTempFolder = false;
 
 public:
-    RocksDBStorageFactory(const char *dbfile, int dbnum, const char *rgchConfig, size_t cchConfig);
+    AsyncWorkQueue **m_rwqueue;
+    AsyncWorkQueue **m_wwqueue;
+
+    RocksDBStorageFactory(const char *dbfile, int dbnum, const char *rgchConfig, size_t cchConfig, AsyncWorkQueue **rwqueue, AsyncWorkQueue **wwqueue);
     ~RocksDBStorageFactory();
 
     virtual IStorage *create(int db, key_load_iterator iter, void *privdata) override;

src/storage/rocksdbfactory.cpp

@@ -35,9 +35,9 @@ rocksdb::Options DefaultRocksDBOptions() {
     return options;
 }
 
-IStorageFactory *CreateRocksDBStorageFactory(const char *path, int dbnum, const char *rgchConfig, size_t cchConfig)
+IStorageFactory *CreateRocksDBStorageFactory(const char *path, int dbnum, const char *rgchConfig, size_t cchConfig, AsyncWorkQueue **rwqueue, AsyncWorkQueue **wwqueue)
 {
-    return new RocksDBStorageFactory(path, dbnum, rgchConfig, cchConfig);
+    return new RocksDBStorageFactory(path, dbnum, rgchConfig, cchConfig, rwqueue, wwqueue);
 }
 
 rocksdb::Options RocksDBStorageFactory::RocksDbOptions()
@@ -52,8 +52,8 @@ rocksdb::Options RocksDBStorageFactory::RocksDbOptions()
     return options;
 }
 
-RocksDBStorageFactory::RocksDBStorageFactory(const char *dbfile, int dbnum, const char *rgchConfig, size_t cchConfig)
-    : m_path(dbfile)
+RocksDBStorageFactory::RocksDBStorageFactory(const char *dbfile, int dbnum, const char *rgchConfig, size_t cchConfig, AsyncWorkQueue **rwqueue, AsyncWorkQueue **wwqueue)
+    : m_path(dbfile), m_rwqueue(rwqueue), m_wwqueue(wwqueue)
 {
     dbnum++; // create an extra db for metadata
     // Get the count of column families in the actual database

src/storage/rocksdbfactory.h

@@ -1,3 +1,3 @@
 #pragma once
 
-class IStorageFactory *CreateRocksDBStorageFactory(const char *path, int dbnum, const char *rgchConfig, size_t cchConfig);
+class IStorageFactory *CreateRocksDBStorageFactory(const char *path, int dbnum, const char *rgchConfig, size_t cchConfig, AsyncWorkQueue **rwqueue, AsyncWorkQueue **wwqueue);

src/storage/teststorageprovider.h

@@ -29,7 +29,7 @@ public:
     virtual bool enumerate_hashslot(callback fn, unsigned int hashslot) const override;
     virtual size_t count() const override;
 
-    virtual std::vector<std::string> getExpirationCandidates(unsigned int, long long) override { return std::vector<std::string>(); }
+    virtual std::vector<std::string> getExpirationCandidates(unsigned int) override { return std::vector<std::string>(); }
     virtual std::vector<std::string> getEvictionCandidates(unsigned int) override { return std::vector<std::string>(); }
     virtual void setExpire(const char *, size_t, long long) override {}
     virtual void removeExpire(const char *, size_t, long long) override {}

src/tls.cpp

@@ -555,19 +555,6 @@ void tlsSetCertificateFingerprint(tls_connection* conn, X509 * cert) {
 #define ASN1_STRING_get0_data ASN1_STRING_data 
 #endif 
 
-class TCleanup {
-    std::function<void()> fn;
-
-public:
-    TCleanup(std::function<void()> fn)
-        : fn(fn)
-    {}
-
-    ~TCleanup() {
-        fn();
-    }
-};
-
 bool tlsCheckCertificateAgainstAllowlist(tls_connection* conn, std::set<sdsstring> allowlist, const char** commonName){
     if (allowlist.empty()){
         // An empty list implies acceptance of all
@@ -640,6 +627,18 @@ bool tlsCheckCertificateAgainstAllowlist(tls_connection* conn, std::set<sdsstrin
     return false;
 }
 
+bool tlsCertificateIgnoreLoadShedding(tls_connection* conn) {
+    const char* cn = "";
+    if (tlsCheckCertificateAgainstAllowlist(conn, g_pserver->tls_overload_ignorelist, &cn)) {
+        // Certificate is in exclusion list, no need to audit log
+        serverLog(LL_NOTICE, "Loadshedding: disabled for %s", conn->c.fprint);
+        return true;
+    } else {
+        serverLog(LL_NOTICE, "Loadshedding: enabled for %s", conn->c.fprint);
+        return false;
+    }
+}
+
 bool tlsCertificateRequiresAuditLogging(tls_connection* conn){
     const char* cn = "";
     if (tlsCheckCertificateAgainstAllowlist(conn, g_pserver->tls_auditlog_blocklist, &cn)) {
@@ -892,6 +891,9 @@ void tlsHandleEvent(tls_connection *conn, int mask) {
                     if (tlsCertificateRequiresAuditLogging(conn)){
                         conn->c.flags |= CONN_FLAG_AUDIT_LOGGING_REQUIRED;
                     }
+                    if (tlsCertificateIgnoreLoadShedding(conn)){
+                        conn->c.flags |= CONN_FLAG_IGNORE_OVERLOAD;
+                    }
                 }
             }
 
@@ -1069,6 +1071,12 @@ static int connTLSWrite(connection *conn_, const void *data, size_t data_len) {
 
     if (conn->c.state != CONN_STATE_CONNECTED) return -1;
     ERR_clear_error();
+
+     if (data_len > std::numeric_limits<int>::max()) {
+        // OpenSSL expects length to be 32-bit int
+        data_len = std::numeric_limits<int>::max();
+    }
+
     ret = SSL_write(conn->ssl, data, data_len);
 
     if (ret <= 0) {

tests/cluster/run.tcl

@@ -10,6 +10,16 @@ source ../../support/cluster.tcl ; # Redis Cluster client.
 set ::instances_count 20 ; # How many instances we use at max.
 set ::tlsdir "../../tls"
 
+# Check if we compiled with flash
+set status [catch {exec ../../../src/keydb-server --is-flash-enabled}]
+if {$status == 0} {
+    puts "KeyDB was built with FLASH, including FLASH tests"
+    set ::flash_enabled 1
+} else {
+    puts "KeyDB was not built with FLASH.  Excluding FLASH tests"
+    set ::flash_enabled 0
+}
+
 proc main {} {
     parse_options
     spawn_instance redis $::redis_base_port $::instances_count {

tests/cluster/tests/17-diskless-load-swapdb.tcl

@@ -14,6 +14,7 @@ test "Cluster is writable" {
     cluster_write_test 0
 }
 
+if {!$::flash_enabled} {
 test "Right to restore backups when fail to diskless load " {
     set master [Rn 0]
     set replica [Rn 1]
@@ -84,3 +85,4 @@ test "Right to restore backups when fail to diskless load " {
     assert_equal {1} [$replica get $slot0_key]
     assert_equal $slot0_key [$replica CLUSTER GETKEYSINSLOT 0 1] "POST RUN"
 }
+}

tests/integration/replication-psync-flash.tcl

@@ -10,7 +10,7 @@
 # checked for consistency.
 if {$::flash_enabled} {
     proc test_psync {descr duration backlog_size backlog_ttl delay cond mdl sdl reconnect} {
-        start_server [list tags {"repl"} overrides [list storage-provider {flash .rocks.db.m} repl-backlog-size 64m]] {
+        start_server [list tags {"repl"} overrides [list storage-provider {flash ./rocks.db.m} repl-backlog-size 64m]] {
             start_server [list tags {flash} overrides [list storage-provider {flash ./rocks.db} delete-on-evict no storage-flush-period 10]] {
 
                 set master [srv -1 client]
@@ -114,6 +114,10 @@ if {$::flash_enabled} {
 
     foreach mdl {no yes} {
         foreach sdl {disabled swapdb} {
+            if {$::flash_enabled && $sdl == "swapdb"} {
+                continue
+            }
+
             test_psync {no reconnection, just sync} 6 1000000 3600 0 {
             } $mdl $sdl 0
 

tests/integration/replication-psync-multimaster.tcl

@@ -128,6 +128,10 @@ proc test_psync {descr duration backlog_size backlog_ttl delay cond mdl sdl reco
 
 foreach mdl {no yes} {
     foreach sdl {disabled swapdb} {
+        if {$::flash_enabled && $sdl == "swapdb"} {
+            continue
+        }
+
         test_psync {no reconnection, just sync} 6 1000000 3600 0 {
         } $mdl $sdl 0
 

tests/integration/replication-psync.tcl

@@ -119,6 +119,10 @@ proc test_psync {descr duration backlog_size backlog_ttl delay cond mdl sdl reco
 
 foreach mdl {no yes} {
     foreach sdl {disabled swapdb} {
+        if {$::flash_enabled && $sdl == "swapdb"} {
+            continue
+        }
+
         test_psync {no reconnection, just sync} 6 1000000 3600 0 {
         } $mdl $sdl 0
 

tests/integration/replication.tcl

@@ -252,6 +252,10 @@ start_server {tags {"repl"}} {
 
 foreach mdl {no yes} {
     foreach sdl {disabled swapdb} {
+        if {$::flash_enabled && $sdl == "swapdb"} {
+            # swapdb not compatible with flash
+            continue
+        }
         start_server {tags {"repl"}} {
             set master [srv 0 client]
             $master config set repl-diskless-sync $mdl
@@ -383,6 +387,7 @@ start_server {tags {"repl"}} {
     }
 }
 
+if {!$::flash_enabled} {
 test {slave fails full sync and diskless load swapdb recovers it} {
     start_server {tags {"repl"}} {
         set slave [srv 0 client]
@@ -547,6 +552,7 @@ test {diskless loading short read} {
         }
     }
 }
+}
 
 # get current stime and utime metrics for a thread (since it's creation)
 proc get_cpu_metrics { statfile } {
@@ -578,12 +584,13 @@ proc compute_cpu_usage {start end} {
     return [ list $pucpu $pscpu ]
 }
 
-
+if {!$::flash_enabled} {
 # test diskless rdb pipe with multiple replicas, which may drop half way
 start_server {tags {"repl"}} {
     set master [srv 0 client]
     $master config set repl-diskless-sync yes
     $master config set repl-diskless-sync-delay 1
+    $master config set enable-keydb-fastsync no
     set master_host [srv 0 host]
     set master_port [srv 0 port]
     set master_pid [srv 0 pid]
@@ -727,6 +734,8 @@ start_server {tags {"repl"}} {
             }
         }
     }
+
+    $master config set enable-keydb-fastsync yes
 }
 
 if 0 {
@@ -811,6 +820,7 @@ test "diskless replication read pipe cleanup" {
     }
 }
 }
+}
 
 test {replicaof right after disconnection} {
     # this is a rare race condition that was reproduced sporadically by the psync2 unit.
@@ -867,8 +877,10 @@ test {replicaof right after disconnection} {
 test {Kill rdb child process if its dumping RDB is not useful} {
     start_server {tags {"repl"}} {
         set slave1 [srv 0 client]
+        $slave1 config set enable-keydb-fastsync no
         start_server {} {
             set slave2 [srv 0 client]
+            $slave2 config set enable-keydb-fastsync no
             start_server {} {
                 set master [srv 0 client]
                 set master_host [srv 0 host]

tests/unit/flash.tcl

@@ -81,21 +81,6 @@ if {$::flash_enabled} {
             }
         }
 
-        test { EXPIRE reduces used disk space } {
-            r flushall
-            r set testkey foo ex 10
-            r debug flush-storage
-            set used [r debug get-storage-usage]
-            assert_equal {1} [r dbsize]
-            wait_for_condition 50 1000 {
-                [r dbsize] == 0
-            } else {
-                fail "key is not expired"
-            }
-            set after [r debug get-storage-usage]
-            assert {$after < $used}
-        }
-
         test { SUBKEY EXPIRE persists after cache flush } {
             r flushall
             r sadd testkey foo bar baz

tests/unit/introspection.tcl

@@ -162,15 +162,18 @@ start_server {tags {"introspection"}} {
             aof_rewrite_cpulist
             time-thread-priority
             bgsave_cpulist
-	    storage-cache-mode
-	    storage-provider-options
-	    use-fork
+            storage-cache-mode
+            storage-provider-options
+            use-fork
             multi-master
             active-replica
             bind
             set-proc-title
             repl-backlog-disk-reserve
-	    tls-allowlist
+            tls-allowlist
+            tls-auditlog-blocklist
+            tls-overload-ignorelist
+            overload-ignorelist
             db-s3-object
         }
 

tests/unit/maxmemory.tcl

@@ -182,6 +182,8 @@ proc test_slave_buffers {test_name cmd_count payload_len limit_memory pipeline}
                 fail "Replication not started."
             }
 
+            after 100
+
             # measure used memory after the slave connected and set maxmemory
             set orig_used [s -1 used_memory]
             set orig_client_buf [s -1 mem_clients_normal]

tests/unit/moduleapi/load.tcl

@@ -1,7 +1,7 @@
 set testmodule [file normalize tests/modules/load.so]
 
 if {$::flash_enabled} {
-    start_server [list tags [list "modules"] overrides [list storage-provider {flash ./rocks.db.master.load.test} databases 256 loadmodule $testmodule]] {
+    start_server [list tags [list "modules"] overrides [list storage-provider {flash ./rocks.db.master.load.test} loadmodule $testmodule]] {
         test "Module is notified of keys loaded from flash" {
             r flushall
             r set foo bar

tests/unit/tls-auditlog.tcl

@@ -51,7 +51,7 @@ if {$::tls} {
         }
     }
 
-    test {TLS: Able to connect while with SAN having a comprehensive list} {
+    test {TLS Audit Log: Able to connect while with SAN having a comprehensive list} {
         start_server {tags {"tls"} overrides {tls-auditlog-blocklist {dummy.keydb.dev san2.keydb.dev other.keydb.dev}}} {
             catch {r PING} e
             assert_match {PONG} $e
@@ -65,7 +65,7 @@ if {$::tls} {
         }
     }
 
-    test {TLS Audit LogTLS: Able to connect while with SAN having a comprehensive list with wildcards} {
+    test {TLS Audit Log: Able to connect while with SAN having a comprehensive list with wildcards} {
         start_server {tags {"tls"} overrides {tls-auditlog-blocklist {dummy.* san*.dev other.*}}} {
             catch {r PING} e
             assert_match {PONG} $e