regression test for lp:1100496 runs too slowly

[Seppo Jaakola]

Running the regression test for lp:1100496 shows that the slave node freezes for long periods. The test will eventually pass, but just takes too long to complete.

Running same test against a native MySQL replication pair does not show similar slowness.

[Seppo Jaakola (seppo-jaakola)]

Fix was pushed into wsrep-5.5, in revision: http://bazaar.launchpad.net/~codership/codership-mysql/wsrep-5.5/revision/3934

The problem was due to InnoDB purge thread which would not treat lock types on WSREP_BF correctly when inheriting locks.

In the fix, InnoDB lock manager was refactored to not use WSREP_BF lock mode anymore. Locking priority is implemented by checking running transactions BF state at the time of each lock conflict. We will need another round of refactoring, to not access mysql THD structure for each lock conflict, transaction state can be held in InnoDB trx.

[Seppo Jaakola (seppo-jaakola)]

The problem does not show up in MySQL 5.6 version. This is because of kernel_mutex has been dropped in 5.6. In 5.5 version, InnoDB purge thread holds kernel_mutex, which finally accumulated the issue.

But, nevertheless, refactoring of WSREP_BF lock mode out, looks like good to merge to 5.6 branch as well. I will start work on this task.

[Seppo Jaakola (seppo-jaakola)]

Pushed a cleanup fix for wsrep-5.5: http://bazaar.launchpad.net/~codership/codership-mysql/wsrep-5.5/revision/3939

There was a problem, which made native build (without WITH_WSREP option) to fail

Merged the fix into wsrep-5.6, in revision:
http://bazaar.launchpad.net/~codership/codership-mysql/5.6/revision/4032

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PXC-1562