mirror of
https://github.com/wekan/wekan.git
synced 2026-11-07 02:52:33 -05:00
Lauri Ojansivu
880ec2d707
With several WeKan servers on one database, every server applied the continuous
backup setting, so each of them tailed the oplog and wrote its own archives into
the same target directory (#6861, item 3).
The leader holds a lease: one continuousBackupLease document with its id and an
expiry. Taking it is one atomic operation (an insert while there is none, else
an update that matches only an unheld, own or expired lease), the leader renews
it every 15 seconds and it lapses after 45, so another server takes over within
a minute when the leader stops. A server that loses the lease stops its stream.
Startup, saving the settings and a change saved on another server all converge
through that loop, and a restore or a fetch from the cloud takes the lease and
waits one renewal so it is the only writer. One server is always its own leader.
Checked against the real driver on FerretDB (SQLite): 24 clients at once give
exactly one leader; renewal, expiry, loss of the old leader and release behave.
tests/continuousBackupLease.test.cjs pins the same and the wiring in
server/continuousBackup.js. docs/Backup/Continuous-Backup.md no longer says
single-server only.
Thanks to Raymond-Voogt and xet7 !
Refs #6861, not closed: the remaining items of the issue were fixed in 4e28465964.
