Hi hackers,
Please find attached a patch proposal to $SUBJECT.
Indeed, we have seen occurrences in [1] that some slots were
not invalidated (while we expected vacuum to remove dead rows
leading to slots invalidation on the standby).
Though we don't have strong evidences that this
was due to transactions holding back global xmin (as vacuum did
not run in verbose mode), suspicion is high enough (as Tom pointed
out that the test is broken on its face (see [1])).
The proposed patch:
- set autovacuum = off on the primary (as autovacuum is the usual suspect
for holding global xmin).
- Ensure that vacuum is able to remove dead rows before launching
the slots invalidation tests.
- If after 10 attempts the vacuum is still not able to remove the dead
rows then the slots invalidation tests are skipped: that should be pretty
rare, as currently the majority of the tests are green (with only one attempt).
While at it, the patch also addresses the nitpicks mentioned by Robert in [2].
[1]:
https://www.postgresql.org/message-id/flat/OSZPR01MB6310CFFD7D0DCD60A05DB1C3FD4A9%40OSZPR01MB6310.jpnprd01.prod.outlook.com#71898e088d2a57564a1bd9c41f3e6f36
[2]: https://www.postgresql.org/message-id/CA%2BTgmobHGpU2ZkChgKifGDLaf_%2BmFA7njEpeTjfyNf_msCZYew%40mail.gmail.com
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
