Hi,
We just had a customer hit this issue. I kind of wonder whether this
shouldn't be backpatched: Currently the execution on the primary is
O(NBuffers * log(ndrels)) whereas it's O(NBuffers * ndrels) on the
standby - with a lot higher constants to boot. That means it's very
easy to get into situations where the standy starts to lag behind very significantly.
> --- a/src/backend/access/transam/twophase.c
> +++ b/src/backend/access/transam/twophase.c
> @@ -1445,6 +1445,7 @@ FinishPreparedTransaction(const char *gid, bool isCommit)
> int ndelrels;
> SharedInvalidationMessage *invalmsgs;
> int i;
> + SMgrRelation *srels = NULL;
>
> /*
> * Validate the GID, and lock the GXACT to ensure that two backends do not
> @@ -1534,13 +1535,16 @@ FinishPreparedTransaction(const char *gid, bool isCommit)
> delrels = abortrels;
> ndelrels = hdr->nabortrels;
> }
> +
> + srels = palloc(sizeof(SMgrRelation) * ndelrels);
> for (i = 0; i < ndelrels; i++)
> - {
> - SMgrRelation srel = smgropen(delrels[i], InvalidBackendId);
> + srels[i] = smgropen(delrels[i], InvalidBackendId);
>
> - smgrdounlink(srel, false);
> - smgrclose(srel);
> - }
> + smgrdounlinkall(srels, ndelrels, false);
> +
> + for (i = 0; i < ndelrels; i++)
> + smgrclose(srels[i]);
> + pfree(srels);
This code is now duplicated three times - shouldn't we just add a
function that encapsulates dropping relations in a commit/abort record?
Greetings,
Andres Freund
