Andrei Elkin writes:

...

...

--- a/sql/rpl_rli.cc +++ b/sql/rpl_rli.cc @@ -1030,7 +1030,7 @@ void Relay_log_info::inc_group_relay_log_pos(ulonglong log_pos, rgi->last_master_timestamp > last_master_timestamp) last_master_timestamp= rgi->last_master_timestamp; } - else + else if (!mi->using_gtid || seen_gtid_log_list_event) { /* Non-parallel case. */ group_relay_log_pos= event_relay_log_pos;

Agree, this looks fine. The dump thread should always be sending the gtid_list event. And yes, the log_pos of the events received until then are not useful to set for the slave.

Right, so this seems a good approach for a fix.

Maybe I should push the workaround to 10.5 to help the effort to remove sporadic failures, and then file a separate bug about the underlying problem and your proposed solution?

Sounds reasonable. Thank you!

Filed as https://jira.mariadb.org/browse/MDEV-33996 - Kristian.