mirror of
https://github.com/MariaDB/server.git
synced 2025-12-11 14:55:44 +01:00
674c9b0261
This commit carries changes to Innodb for the user XA's binlog-coordinated
ordered commit. It's the part III of the series aimed at
improving on parallel slave performance and providing crash-recovery.
As the part I has unified binlog coordinated completion (commit and
rollback) of the XA transactions with the normal ones a remained
task is to optimize flush to disk call away from the XA completion execution
in engine like it's done to the normal transaction.
Namely innobase_{rollback,commit}_by_xid()^\footnote{Unlike the normal
transaction case, XA-rollback also needs a similar addressing} now
raise a flag blocking the flush before innobase_commit_low().
The flag is restored upon that, which of course what the normal
transaction does in its execution path past the low-commit.
Tests demonstrate the changes are effectual, e.g
binlog_xa_prepared_disconnect.test " failed ", to require it to
adapted to possible loss of the commit/rollback upon crash.
Binlog recovery in such scenarios is going to be recovered in the part
IV of MDEV-33168. The feasibility of successful recovery of
XA-"COMPLETE" is based on a plain observation, that in order to decide
whether to commit an XA or not it's sufficient to follow
the normal transaction recovery rule: when xid is found in binlog
the transaction gets completed (committed in the normal transaction case).
|
||
|---|---|---|
| .. | ||
| archive | ||
| blackhole | ||
| columnstore | ||
| connect | ||
| csv | ||
| example | ||
| federated | ||
| federatedx | ||
| heap | ||
| innobase | ||
| maria | ||
| mroonga | ||
| myisam | ||
| myisammrg | ||
| oqgraph | ||
| perfschema | ||
| rocksdb | ||
| sequence | ||
| sphinx | ||
| spider | ||
| test_sql_discovery | ||