]> git.ipfire.org Git - thirdparty/postgresql.git/commit
Fix race with timeline selection in logical decoding during promotion
authorMichael Paquier <michael@paquier.xyz>
Thu, 11 Jun 2026 08:28:57 +0000 (17:28 +0900)
committerMichael Paquier <michael@paquier.xyz>
Thu, 11 Jun 2026 08:28:57 +0000 (17:28 +0900)
commiteb4e7224a1c6f0058d708cdfda7326bbf884a871
tree9d7ea9342eba523c17e6c46203f54da8250b0f4d
parent987440b33a511482232c59a190cc16ae4feff9aa
Fix race with timeline selection in logical decoding during promotion

During promotion, there is a window where RecoveryInProgress() returns
true but the WAL segments of the old timeline have already been removed.
A logical decoding could pick up the old timeline in this window when
reading a page, failing with the following error:
ERROR: requested WAL segment ... has already been removed

This issue does not lead to any data correctness issue, as retrying to
decode the data works in follow-up decoding attempts.  It impacts
availability, though.  Other WAL page read callbacks have a similar
issue, this commit takes care of what should be the noisiest code path:
logical decoding with START_REPLICATION in a WAL sender.

A TAP test, based on an injection point waiting in the startup process
after the segments have been removed/recycled, is added.  This part is
backpatched down to v17.

This issue has been causing sporadic failures in the buildfarm, and
was reproducible manually.  This issue happens since logical decoding on
standbys exists, down to v16.

Reported-by: Alexander Lakhin <exclusion@gmail.com>
Author: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
Reviewed-by: Hayato Kuroda <kuroda.hayato@fujitsu.com>
Reviewed-by: Xuneng Zhou <xunengzhou@gmail.com>
Discussion: https://postgr.es/m/7daef094-abf3-4672-bc23-3df4763b16a3@gmail.com
Backpatch-through: 16
src/backend/access/transam/xlog.c
src/backend/replication/walsender.c
src/test/recovery/t/035_standby_logical_decoding.pl