* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> I wrote:
> > Peter Geoghegan <pg@heroku.com> writes:
> >> On Fri, May 8, 2015 at 11:59 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> >>> Ooops. But shouldn't that have failed 100% of the time in a CCA build?
> >>> Or is the candidates list fairly noncritical?
>
> >> The candidates list is absolutely critical.
>
> > Oh, I was confusing CCA with RELCACHE_FORCE_RELEASE, which does something
> > a bit different.
>
> Actually, looking closer, the quoted code is simply not broken without
> RELCACHE_FORCE_RELEASE: without that, neither heap_close nor index_close
> will do anything that could cause a cache flush. So while it's certainly
> good pratice to move that lappend_oid call up, it does not explain the
> observed symptoms. We still need some more investigation here.
Couldn't a cache flush request come from another backend? Although this
isn't being run in a parallel group, is it? Maybe a delayed signal that
happens to show up late at just the right time? Dunno if we've ever
actually seen that but the thought occured to me.
Thanks!
Stephen