Re: Race condition in TransactionIdIsInProgress

On Fri, 11 Feb 2022 at 08:48, Simon Riggs <simon(dot)riggs(at)enterprisedb(dot)com> wrote:
>
> On Fri, 11 Feb 2022 at 06:11, Andres Freund <andres(at)anarazel(dot)de> wrote:
>
> > Looks lik syncrep will make this a lot worse, because it can drastically
> > increase the window between the TransactionIdCommitTree() and
> > ProcArrayEndTransaction() due to the SyncRepWaitForLSN() inbetween. But at
> > least it might make it easier to write tests exercising this scenario...
>
> Agreed
>
> TransactionIdIsKnownCompleted(xid) is only broken because the single
> item cache is set too early in some cases. The single item cache is
> important for performance, so we just need to be more careful about
> setting the cache.

Something like this... fix_cachedFetchXid.v1.patch prevents the cache
being set, but this fails! Not worked out why, yet.

just_remove_TransactionIdIsKnownCompleted_call.v1.patch
just removes the known offending call, passes make check, but IMHO
leaves the same error just as likely by other callers.

--
Simon Riggs http://www.EnterpriseDB.com/