Re: pgsql: Avoid duplicate XIDs at recovery when building initial snapshot

On Tue, Oct 23, 2018 at 10:43:38AM +0900, Michael Paquier wrote:
> Well, following the same kind of thoughts, txid_current_snapshot() uses
> sort_snapshot() to remove all the duplicates after fetching its data
> from GetSnapshotData(), so wouldn't we want to do something about
> removal of duplicates if dummy PGXACT entries are found while scanning
> the ProcArray also in this case? What I would think we should do is not
> only to patch GetRunningTransactionData() but also GetSnapshotData() so
> as we don't have duplicates also in this case, and do things in such a
> way that both code paths use the same logic, and that we don't need to
> have sort_snapshot() anymore. That would be more costly though...

My apologies it took a bit longer than I thought. I got a patch on my
desk for a couple of days, and finally took the time to finish something
which would address the concerns raised here. As long as we don't reach
more than hundreds of thousands of entries, there is not going to be any
performance impact. So what I do in the attached is to revert 1df21ddb,
and then have GetRunningTransactionData() sort the XIDs in the snapshot
and remove duplicates only if at least one dummy proc entry is found
while scanning, for xids and subxids. This way, there is no need to
impact most of the instance deployments with the extra sort/removal
phase as most don't use two-phase transactions. The sorting done at
recovery when initializing the standby snapshot still needs to happen of
course.

The patch is added to the upcoming CF for review.

Thanks,
--
Michael