On Tue, Apr 23, 2019 at 07:54:52PM -0400, Tom Lane wrote:
> That seems like pretty much of a hack :-(, in particular I'm not
> convinced that we'd not end up with a missing index entry afterwards.
> Maybe it's the only way, but I think first we need to trace down
> exactly why this broke. I remember we had some special-case code
> for reindexing pg_class, maybe something broke that?
Yes, reindex_relation() has some infra to enforce the list of indexes
in the cache for pg_class which has been introduced by a56a016 as far
as it goes.
> It also seems quite odd that it doesn't fail every time; surely it's
> not conditional whether we'll try to insert a new pg_class tuple or not?
> We need to understand that, too. Maybe the old code never really
> worked in all cases? It seems clear that the commit you bisected to
> just allowed a pre-existing misbehavior to become exposed (more easily).
>
> No time to look closer right now.
Yeah, there is a fishy smell underneath which comes from 9.6. When
testing with 9.5 or older a database creation does not create any
crash on a subsequent reindex. Not sure I'll have the time to look at
that more today, perhaps tomorrow depending on the odds.
--
Michael