]> granicus.if.org Git - postgresql/commit
Avoid catalog lookups in RelationAllowsEarlyPruning().
authorThomas Munro <tmunro@postgresql.org>
Wed, 28 Aug 2019 01:37:03 +0000 (13:37 +1200)
committerThomas Munro <tmunro@postgresql.org>
Wed, 28 Aug 2019 05:59:50 +0000 (17:59 +1200)
commiteeb68c1cddf3cdbfbd543a7aa24e6cbbe0e31b16
tree58b04dd7161a012f01e9c65940d6f7385b37cda4
parent19bfa15a8274a8429883d08ab4c5b3881ac65f07
Avoid catalog lookups in RelationAllowsEarlyPruning().

RelationAllowsEarlyPruning() performed a catalog scan, but is used
in two contexts where that was a bad idea:

1.  In heap_page_prune_opt(), which runs very frequently in some large
    scans.  This caused major performance problems in a field report
    that was easy to reproduce.

2.  In TestForOldSnapshot(), which runs while we hold a buffer content
    lock.  It's not clear if this was guaranteed to be free of buffer
    deadlock risk.

The check was introduced in commit 2cc41acd8 and defended against a
real problem: 9.6's hash indexes have no page LSN and so we can't
allow early pruning (ie the snapshot-too-old feature).  We can remove
the check from all later releases though: hash indexes are now logged,
and there is no way to create UNLOGGED indexes on regular logged
tables.

If a future release allows such a combination, it might need to put
a similar check in place, but it'll need some more thought.

Back-patch to 10.

Author: Thomas Munro
Reviewed-by: Tom Lane, who spotted the second problem
Discussion: https://postgr.es/m/CA%2BhUKGKT8oTkp5jw_U4p0S-7UG9zsvtw_M47Y285bER6a2gD%2Bg%40mail.gmail.com
Discussion: https://postgr.es/m/CAA4eK1%2BWy%2BN4eE5zPm765h68LrkWc3Biu_8rzzi%2BOYX4j%2BiHRw%40mail.gmail.com
src/backend/utils/cache/relcache.c
src/include/utils/rel.h
src/include/utils/snapmgr.h