The following bug has been logged online:
Bug reference: 4952
Logged by: Jeff Janes
Email address: firstname.lastname@example.org
PostgreSQL version: 8.4.0
Operating system: Linux 2.4.21-15.0.3.ELsmp
Description: commit_delay ignored because CountActiveBackends always
I've found that commit_delay never has an effect.
By instrumenting src/backend/access/transam/xact.c, I see that this is
because CountActiveBackends always returns zero.
This seems to be a bug introduced by:
into file src/backend/storage/ipc/procarray.c
I believe the source of the bug is the addition to that file of
+ if (proc != NULL)
The sense of this test should be inverted, as it is NULLs, not non-nulls
that need to be skipped.
Due to this test all truly active backends get skipped, so
CountActiveBackends() always returns zero. Also, I believe without evidence
that the change fails to correct the (hard to reproduce) bug it was
originally introduced to correct, as a proc that is NULL does not get
skipped, and goes on to be dereferenced.
If I change this code to:
+ if (proc == NULL)
Then I find that commit_delay now does have an effect.
(The effect is to make "pgbench -c 15" slower when commit_delay is set to
the max value of 100000. This is what I thought would happen, and was
surprised when it originally did not.)