Re: [HACKERS] ginInsertCleanup called from vacuum could still miss tuples to be deleted

Hi,
Please unsubscribe me.
thanksEric

On Tuesday, November 14, 2017, 2:02:04 AM GMT+8, Peter Geoghegan <pg(at)bowt(dot)ie> wrote:

On Mon, Nov 13, 2017 at 12:25 AM, Masahiko Sawada <sawada(dot)mshk(at)gmail(dot)com> wrote:
> Commit e2c79e14 prevented multiple cleanup process for pending list in
> GIN index. But I think that there is still possibility that vacuum
> could miss tuples to be deleted if someone else is cleaning up the
> pending list.

I've been suspicious of that commit (and related commits) for a while
now [1]. I think that it explains a couple of different problem
reports that we have seen.

> In ginInsertCleanup(), we lock the GIN meta page by LockPage and could
> wait for the concurrent cleaning up process if stats == NULL. And the
> source code comment says that this happen is when ginINsertCleanup is
> called by [auto]vacuum/analyze or gin_clean_pending_list(). I agree
> with this behavior. However, looking at the callers the stats is NULL
> only either if pending list exceeds to threshold during insertions or
> if only analyzing is performed by an autovacum worker or ANALYZE
> command. So I think we should inVacuum = (stats != NULL) instead.
> Also, we might want autoanalyze and ANALYZE command to wait for
> concurrent process as well. Attached patch fixes these two issue. If
> this is a bug we should back-patch to 9.6.

How did you figure this out? Did you just notice that the code wasn't
doing what it claimed to do, or was there a problem that you saw in
production?

[1] https://postgr.es/m/CAH2-WzmtLXbs8+c19t1T=Rj0KyP7vK9q8hQJULgDLdVMuEeeUw@mail.gmail.com
--
Peter Geoghegan