[MXS-4137] Add native support for conditional group commit tuning for parallel replication Created: 2022-05-17 Updated: 2023-12-15 |
|
| Status: | Open |
| Project: | MariaDB MaxScale |
| Component/s: | mariadbmon |
| Affects Version/s: | None |
| Fix Version/s: | 24.02 |
| Type: | New Feature | Priority: | Major |
| Reporter: | Rob Schwyzer | Assignee: | Joe Cotellese |
| Resolution: | Unresolved | Votes: | 2 |
| Labels: | None | ||
| Description |
| Comments |
| Comment by markus makela [ 2023-04-30 ] |
rob.schwyzer@mariadb.com does this only affect replicas where non-replication traffic generates writes? In other words, if the replicas are read-only, would this feature have any effect on them? Write performance on replicas seems like something that's not at all relevant for asynchronous replication that's managed by MaxScale: a replica never receives writes and thus one would assume that any group commit ordering and/or delays wouldn't affect it. If this is not the case and you're requesting something else, can you give a clear example where the problem occurs? Elkin Would you be able to confirm whether binlog_commit_wait_count and binlog_commit_wait_usec have an effect on transactions commited by replication? If these configuration variables affect the replicas that are used in this read-only behavior, I'd say that this is something that should be fixed in the server. If this behavior does not affect replication in any significant way, like you mentioned, this seems like something that's already implemented in the form of promotion and demotion scripts and this feature would offer a more limited set of actions to take (i.e. no dynamic SQL). Perhaps a better way to deal with this issue would be to make the SQL scripts more easy to edit from the GUI? Having to invent a new mechanism for doing something that's already doable in a more versatile and generally applicable way doesn't seem very useful at this point. I do see the value in "best practices" but this does not require this kind of a feature to be implemented and could be done with something as simple as a default SQL script for promotion and demotion. |
| Comment by Rob Schwyzer [ 2023-05-01 ] |
This would be the critical bit of knowledge, yeah. The concern would be the impact to local binlogging on replicas. I'll message Faisal to recheck this though to ensure there's not another use-case from his customers. |
| Comment by Andrei Elkin [ 2023-05-03 ] |
|
markus makela, the variables may be relevant the parallel (read-only) slave, not for the serial one. I am not yet fully confident in their effect 'cos on slave there's a separate grouping to commit in the master's binlog order. The two groupings overlap so the user has to careful at playing around with the vars. |