On Tue, Oct 2, 2012 at 05:36:28PM -0700, Tianyin Xu wrote:
> Hi, pgsql-docs,
>
> I was confused by the system behavior of PG's write ahead log, specifically the
> relation between commit_delay, commit_siblings, and fsync.
>
> Basically according to the manual (http://www.postgresql.org/docs/9.1/static/
> runtime-config-wal.html), I know:
>
> "commit_delay" specifies the delay before a transaction attempts to flush the
> WAL buffer out to disk.
> "commit_siblings" specifies minimum number of concurrent open transactions to
> require before performing the commit_delay delay.
>
> However, after looking at the code, I figured out these configuration
> parameters are also controlled by fsync.
>
> The only usage of the two parameter I can find is:
>
> /* src/backend/access/transam/xact.c */
>
> static TransactionId
> RecordTransactionCommit(void)
> ...
> if (CommitDelay > 0 &&
> enableFsync &&
> MinimumActiveBackends(CommitSiblings))
> pg_usleep(CommitDelay);
>
> ...
> }
>
> where, CommitDelay stores the parameter of "commit_delay", CommitSiblings
> stores the parameter of "commit_siblings", and enableFsync stores the parameter
> of "fsync".
>
> It seems the two parameters (commit_delay and commit_siblings) only have effect
> when fsync is set to be *on*.
>
> But, unfortunately, such behavior is not covered in the manual. The mistake I
> made is only set the two values but disabling fsync so the expected PG does not
> behave as I expected. I think it's definitely helpful to specify the relations
> between these configuration parameters either in the manual, or in the source
> code by giving warnings logs.
Docs updated for PG 9.3 with the attached patch.
--
Bruce Momjian <bruce@momjian.us> http://momjian.us
EnterpriseDB http://enterprisedb.com
+ It's impossible for everything to be true. +