MDEV-37229: Set proper trx isolation level before every applied event #512
Conversation
denis-protivensky
force-pushed
the
10.6-MDEV-37229
branch
from
ae78717 to
f632a63
Compare
denis-protivensky
force-pushed
the
10.6-MDEV-37229
branch
from
f632a63 to
937511c
Compare
| } | ||
| } | ||
|
|
||
| /* Statement-based replication requires InnoDB repeatable read |
There was a problem hiding this comment.
I wonder if this is a correct fix. It may make the test pass for sure, but the purpose of the wsrep_forced_binlog_format is to force given binlog format on master node. Here it is used in applier context. There is no guarantee that this variable would have the same value on all of the nodes.
Would there be a better way check whether the isolation level should be changed, e.g. by looking at event type? If the type is QUERY_EVENT, use repeatable read, otherwise read committed.
This fix also does not address the fact that the isolation level is left to repeatable read after IST (any idea why is that?), maybe that should be at least mentioned in commit message.
There was a problem hiding this comment.
I don't remember the details of IST, I'll have to recall that. Concerning the QUERY_EVENT, is it only used for statement-based replication?
There was a problem hiding this comment.
In my understanding Query event is also used in TOI
|
Closing in favor of MariaDB#4422 |
Apparently, applier thread performing IST would be left with REPEATABLE_READ isolation level.
Every applier thread in Galera should run with READ_COMMITTED transaction isolation level to prevent applying issues caused by InnoDB gap locks.
The exception is statement-based replication, where REPEATABLE_READ is required by the server code.