Hi,
On 2022-01-18 21:50:07 -0500, Tom Lane wrote:
> I just found one thing making check-world slower than it ought to be:
> src/test/recovery/t/008_fsm_truncation.pl does
>
> $node_primary->append_conf(
> 'postgresql.conf', qq{
> fsync = on
> wal_log_hints = on
> max_prepared_transactions = 5
> autovacuum = off
> });
>
> There is no reason for this script to be overriding Cluster.pm's
> fsync = off setting.
>
> This appears to go back to 917dc7d23 of 2016, so I think it just
> predates our recognition that we should disable fsync in routine
> tests.
Yea, I noticed this too. I was wondering if there's a conceivable reason to
actually want fsyncs, but I couldn't come up with one.
On systems where IO isn't overloaded, the main problem with this test are
elsewhere: It multiple times waits for VACUUMs that are blocked truncating the
table. Which these days takes 5 seconds. Thus the test takes quite a while.
To me VACUUM_TRUNCATE_LOCK_TIMEOUT = 5s seems awfully long. On a system with a
lot of tables that's much more than vacuum will take. So this can easily lead
to using up all autovacuum workers...
> This actually causes parallel check-world to fail altogether on florican's
> host, because the initial fsync of the recovered primary takes more than 3
> minutes when there's conflicting I/O traffic, causing pg_ctl to time out.
Ugh.
I noticed a few other sources of "unnecessary" fsyncs. The most frequent
being the durable_rename() of backup_manifest in pg_basebackup.c. Manifests are
surprisingly large, 135k for a freshly initdb'd cluster.
There's an fsync in walmethods.c:tar_close() that sounds intentional, but I
don't really understand what the comment:
/* Always fsync on close, so the padding gets fsynced */
if (tar_sync(f) < 0)
Greetings,
Andres Freund