MDEV-38636 @@sql_slave_skip_counter resets to 1 when skipping events by @@gtid_ignore_duplicates
#4593
Conversation
There was a problem hiding this comment.
This section is after the block where this driver SQL thread dismisses queues events to its workers.
There was a problem hiding this comment.
We might be able to remove the @@gtid_ignore_duplicates code on the Parallel side as well now, but I don’t dare, and so leave it to MDEV-30458 Consolidate Serial Replica to Parallel Replica with 1 Worker Thread.
sql/log_event_server.cc
Outdated
| rli->slave_skip_counter)); | ||
| if ((server_id == global_system_variables.server_id && | ||
| !(rli->replicate_same_server_id || (flags & LOG_EVENT_ACCEPT_OWN_F))) || | ||
| rli->gtid_skip_flag || |
There was a problem hiding this comment.
Parallel mode should (at least I expect) handle Relay_log_info::gtid_skip_flag separately, so gtid_skip_flag-skipped events should never reach the amended code.
There was a problem hiding this comment.
I replaced Relay_log_info::gtid_skip_flag with rpl_group_info::gtid_ignore_duplicate_state; my expectation still applies.
…s by `@@gtid_ignore_duplicates` MDEV-5804 implemented `@@gtid_ignore_duplicates` on serial replication by overwriting `@@sql_slave_skip_counter` to 1, which MDEV-4937 reënabled in GTID replication. To avoid conflicts at their root, this commit reïmplements `@@gtid_ignore_duplicates` so it no longer depends on `@@sql_slave_skip_counter`. The new implementation is located next to `@@replicate_events_marked_for_skip` and makes use of the existing `rpl_group_info::gtid_ignore_duplicate_state` infrastructure, which was only used for updating states, not by `@@gtid_ignore_duplicates` itself. Reviewed-by: Kristian Nielsen <knielsen@knielsen-hq.org>
@@sql_slave_skip_counterincreases to 1 when skipping events by@@gtid_ignore_duplicates#4585MDEV-5804 implemented
@@gtid_ignore_duplicateson serial replication by overwriting@@sql_slave_skip_counterto 1, which MDEV-4937 reënabled in GTID replication.To avoid conflicts at their root, this commit reïmplements
@@gtid_ignore_duplicatesso it no longer depends on@@sql_slave_skip_counter. The new implementation is located next to@@slave_skip_counterand utilizes the existingrpl_group_info::gtid_ignore_duplicate_stateinfrastructure, which was only used to update states, not by@@gtid_ignore_duplicatesitself.