race condition in sync rep

I believe I've figured out why synchronous replication has such
terrible performance with fsync=off: it has a nasty race condition.
It may happen - if the standby responds very quickly - that the
standby acks the commit record and awakens waiters before the
committing backend actually begins to wait. There's no cross-check
for this: the committing backend waits unconditionally, with no regard
to whether the necessary ACK has already arrived. At this point we
may be in for a very long wait: another ACK will be required to
release waiters, and that may not be immediately forthcoming. I had
thought that the next ACK (after at most wal_receiver_status_interval)
would do the trick, but it appears to be even worse than that: by
making the standby win the race, I was easily able to get the master
to hang for over a minute, and it only got released when I committed
another transaction. Had I been sufficiently patient, the next
checkpoint probably would have done the trick.

Of course, with fsync=off on the standby, it's much easier for the
standby to win the race.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company