[PATCH v4 0/3] Replica rejoin

Vladimir Davydov vdavydov.dev at gmail.com
Sat Jul 21 15:38:11 MSK 2018

After this patch set is applied, an instance will try to detect if
it fell too much behind its peers in the cluster and so needs to be
rebootstrapped. If it does, it will skip local recovery and instead
proceed to bootstrap from a remote master. Old files (xlog, snap)
are not deleted during rebootstrap. They will be removed by gc as


Changes in v4:
 - Rebase on top of the latest 1.10 and remove merged patches.
 - Log everything that affects rebootstrap decision making.
 - Rebootstrap an instance only if it can't follow *all* masters,
   not just *any* of them, as it used to be.

v3: https://www.freelists.org/post/tarantool-patches/PATCH-v3-0011-Replica-rejoin

Changes in v3:
 - Remove merged patches, add some new ones.
 - Rebase on top of the latest 1.10: this required patching gc to make
   it track vclocks instead of signatures so that it could report the
   vclock of the oldest xlog stored on the instance.
 - Follow-up on the recently committed patch for recovery subsystem: add
   some comments and remove double scanning of the WAL directory.
 - Introduce a new IPROTO command, IPROTO_REQUEST_STATUS, to be used
   instead of IPROTO_REQUEST_VOTE; send a map in reply to this command.
   Rationale: a map is more flexible and can be extended. In particular,
   we can use the very same message for inquiring the oldest vclock
   stored on the master to detect if a replica needs to be rejoined,
   instead of introducing a new IPROTO command, as we did in v2.
 - Do NOT rebootstrap a replica if it has some data that is absent on
   the master. Rationale: we don't want to lose ANY data by rejoining a
   replica; besides, if a replica's vclock is incomparable with the
   master's, xdir_scan may break.

v2: https://www.freelists.org/post/tarantool-patches/PATCH-v2-0011-Replica-rejoin

Changes in v2:
 - Implement rebootstrap support for vinyl engine.
 - Call recover_remaining_wals() explicitly after recovery_stop_local()
   as suggested by @kostja.
 - Add comment to memtx_engine_new() explaining why we need to init
   INSTANCE_UUID before proceeding to local recovery.

v1: https://www.freelists.org/post/tarantool-patches/RFC-PATCH-0012-Replica-rejoin

Vladimir Davydov (3):
  replication: rebootstrap instance on startup if it fell behind
  vinyl: simplify vylog recovery from backup
  vinyl: implement rebootstrap support

 src/box/box.cc                           |   9 ++
 src/box/relay.cc                         |   3 +
 src/box/replication.cc                   |  59 +++++++
 src/box/replication.h                    |   9 ++
 src/box/vy_log.c                         | 190 +++++++++++++++++------
 src/box/vy_log.h                         |  34 ++++
 src/errinj.h                             |   1 +
 test/box/errinj.result                   |   6 +-
 test/replication/replica_rejoin.result   | 250 ++++++++++++++++++++++++++++++
 test/replication/replica_rejoin.test.lua |  91 +++++++++++
 test/vinyl/replica_rejoin.lua            |  13 ++
 test/vinyl/replica_rejoin.result         | 257 +++++++++++++++++++++++++++++++
 test/vinyl/replica_rejoin.test.lua       |  88 +++++++++++
 test/vinyl/suite.ini                     |   2 +-
 14 files changed, 965 insertions(+), 47 deletions(-)
 create mode 100644 test/replication/replica_rejoin.result
 create mode 100644 test/replication/replica_rejoin.test.lua
 create mode 100644 test/vinyl/replica_rejoin.lua
 create mode 100644 test/vinyl/replica_rejoin.result
 create mode 100644 test/vinyl/replica_rejoin.test.lua


More information about the Tarantool-patches mailing list