← Back to team overview

maria-developers team mailing list archive

Fix for TokuDB and parallel replication


Parallel replication so far did not work well with TokuDB, as some people
who tried it found out. I have now pushed to 10.1 some patches to solve the
problems. There are two main fixes:

1. Fix some races where a waiting transaction would miss its wakeup and get
a lock timeout on a waiting row lock, even though the lock was released by
the holding transaction. This fix is due to great work by Rich Prohaska.
This problem is actually not specific to replication, normal transactions on
a master will experience it too. But it hurts replication a lot, since
replication must commit transactions in order, so one stalled transaction
stalls all following transactions as well.

2. Implement the conflict detection and handling necessary for optimistic
parallel replication to work. This basically implements
thd_rpl_deadlock_check() and hton->kill_query methods. This should solve the
problems where optimistic parallel replication with TokuDB breaks with lock
wait timeouts.

If someone wants to test it, I have made a tree available with just these
fixes on top of MariaDB 10.1.19:


The fix should appear in 10.1.20 eventually.

The first part of the patch has also been submitted by Rich to
upstream. When this is (hopefully) merged upstream, and upstream merged into
MariaDB, the MariaDB version of the fix should be replaced with the Percona
one. I tried making the MariaDB version of the fix identical to Rich's pull
request and keeping it in a separate commit, so this should hopefully be
simple to do when the time comes.

 - Kristian.

Follow ups