Re: out-of-order XID insertion in KnownAssignedXids

On Thu, Oct 11, 2018 at 08:04:11PM +0300, Konstantin Knizhnik wrote:
> Proposed patch is attached.

The problem I have with this patch doing the duplication removal and
qsort work in LogCurrentRunningXacts is that it would still lock
ProcArrayLock until the WAL record has been written with wal_level =
logical. With many sessions and 2PC transactions running around, this
would be a performance impact for any deployments every time a
checkpoint happens or every time the bgwriter decide to log a standby
snapshot. Hence I would go instead with the attached, which does the
legwork at recovery, which is a one-time code path as you mentioned.
Okay, this makes the recovery a bit longer but that's way better than
impacting all deployments of Postgres, even those not using 2PC when
normally running. And as the sorting phase already happens we just need
to do something like the attached.

One thing that we could also do for HEAD is to add in
RunningTransactionsData if the transaction comes from a 2PC entry, and
allow recovery to do more sanity checks. This would require a WAL
format change. The proposed patch needs to be back-patched down to
9.3.
--
Michael