]> granicus.if.org Git - postgresql/commit
Prevent sharing transition states between ordered-set aggregates.
authorTom Lane <tgl@sss.pgh.pa.us>
Thu, 12 Oct 2017 02:18:01 +0000 (22:18 -0400)
committerTom Lane <tgl@sss.pgh.pa.us>
Thu, 12 Oct 2017 02:18:01 +0000 (22:18 -0400)
commit96cfc7e19ae1b70597a90487262cfb9f4531d846
treecd639130f55403f3cd60ce69302bc3cb49d7fb11
parent0da46d75e31ddfa9180345a14d720814e36922fa
Prevent sharing transition states between ordered-set aggregates.

This ought to work, but the built-in OSAs are not capable of coping,
because their final-functions destructively modify their transition
state (specifically, the contained tuplesort object).  That was fine
when those functions were written, but commit 804163bc2 moved the
goalposts without telling orderedsetaggs.c.

We should fix the built-in OSAs to support this, but it will take
a little work, especially if we don't want to sacrifice performance
in the normal non-shared-state case.  Given that it took a year after
9.6 release for anyone to notice this bug, we should not prioritize
sharable-state over nonsharable-state performance.  And a proper fix
is likely to be more complicated than we'd want to back-patch, too.

Therefore, let's just put in this stop-gap patch to prevent nodeAgg.c
from choosing to use shared state for OSAs.  We can revert it in HEAD
when we get a better fix.

Report from Lukas Eder, diagnosis by me, patch by David Rowley.
Back-patch to 9.6 where the problem was introduced.

Discussion: https://postgr.es/m/CAB4ELO5RZhOamuT9Xsf72ozbenDLLXZKSk07FiSVsuJNZB861A@mail.gmail.com
src/backend/executor/nodeAgg.c
src/test/regress/expected/aggregates.out
src/test/regress/sql/aggregates.sql