Fix ready. To be merged after 8.0.18 release.

Short:

Assertion is wrong. It should check if there are any temporary tables that should be logged to binlog, not if there are any temporary tables.

Long:

This is handling of corner case when:

1. gtid_next is set to GTID value

2. DROP statement with temporary and non-temporary tables is executed

3. row-based logging is enabled

Theoretically such transaction should not be logged by master. It would be split into 2 separate transactions.

However there are still at least 2 cases when it can happen:

1. Simulate replicated transaction by invoking DROP with 2 such tables directly on slave

2. Create t1, t2 as non-temporary tables on master, then replicate to slave. Next drop t1, t2 from slave and create t1, t2 on slave only using stored procedure. This scenario is strange, however rpl_gtid_split_statements.test performs such scenario.

If slave uses SBL, case 1 above will issue the error (2 tables would be logged to binlog, but as separate transactions using the same GTID, which is not allowed).

If slave uses SBL, case 2 will still work, as temporary tables created by stored procedures are not logged even if SBL is used.

If slave uses RBL, case 1 will cause only non-temporary table to be logged, temporary table will be just dropped without logging.

8.0 fix: https://github.com/percona/percona-server/pull/3466

Verified still exists in 8.0.16

8.0 triage - Fix if Time.

Thank you for the core analysis.