On Thu, 2010-05-13 at 20:13 +0900, Fujii Masao wrote:
> On Thu, May 13, 2010 at 8:05 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> > On Thu, 2010-05-13 at 19:32 +0900, Fujii Masao wrote:
> >> On Thu, May 13, 2010 at 7:22 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> >> > On Thu, 2010-05-13 at 19:08 +0900, Fujii Masao wrote:
> >> >
> >> >> I was able to reproduce such a hang by not executing another
> >> >> transaction after rollback. In this case, walsender cannot replicate
> >> >> the rollback since it's not in the disk.
> >> >
> >> > WALWriter is not active?
> >>
> >> WALWriter is active, but unfortunately it doesn't flush all of the WAL
> >> records in wal_buffers. Please see XLogBackgroundFlush().
> >
> > Yes, I wrote it. It flushes after at most 3 cycles, stated in comments.
>
> Yeah, what is worse is that RecordTransactionAbort() doesn't update the
> XLogCtl->asyncCommitLSN, so rollback might not be flushed even after at
> 3 cycles.
Well spotted. Fix applied, thanks both.
-- Simon Riggs www.2ndQuadrant.com