Re: Ideas for a relcache test mode about missing invalidations

At Thu, 2 Aug 2018 14:40:50 -0700, Peter Geoghegan <pg(at)bowt(dot)ie> wrote in <CAH2-WznJ8K1HXLTKOY53QBw2dtY0z53xKB-0RpqsJQBS0TkrdQ(at)mail(dot)gmail(dot)com>
pg> On Thu, Aug 2, 2018 at 3:18 AM, Kyotaro HORIGUCHI
pg> <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> wrote:
pg> >> [1] http://archives.postgresql.org/message-id/CAKoxK%2B5fVodiCtMsXKV_1YAKXbzwSfp7DgDqUmcUAzeAhf%3DHEQ%40mail.gmail.com
pg> >> [2] https://www.postgresql.org/message-id/12259.1532117714@sss.pgh.pa.us
pg> >
pg> > As for [1], it is not a issue on invalidation. It happens also if
pg> > the relation has any index and even drop is not needed. The
pg> > following steps are sufficient.
pg> >
pg> > create table t( pk serial, t text );
pg> > insert into t( t ) values( 'hello' ), ('world');
pg> > create index idx_fake0 on t (pk);
pg> > create index idx_fake on t ( f_fake( pk ) ); -- ERROR
pg>
pg> The fact that there was a weird error wasn't really what we cared
pg> about there. If the user is doing something that's clearly
pg> unreasonable or nonsensical, then they cannot expect much from the
pg> error message (maybe we could do better, but it's not a priority).

Hmm. I don't think it's unreasonable or nonsensical, but I don't
argue it here.

pg> What we really cared about was the fact that it was possible to make a
pg> backend's relcache irrecoverably corrupt. That should never be allowed
pg> to happen, even when the user is determined to do something
pg> unreasonable.

I reread through the thread and IUCC, drop_index() is sending
inval on the owing relation and invalidation happens (that is,
RelationCacheInvalidateEntry is called for the owner
relation.). Relcache for the owner is actually being rebuit
during the following create index. At least the problem mentioned
in [1] is happening using a fresh relcache created after invoking
the index creation. The cause is RelationGetIndexList collects
all "indislive" indexes, including the idx_fake before population,
which can be fixed by the attached patch..

I'm I still misunderstanding something?

# Anyway, I'll make another thread for the another issue.

[1]: https://www.postgresql.org/message-id/20180628150209.n2qch5jtn3vt2xaa%40alap3.anarazel.de

regards.

--
Kyotaro Horiguchi
NTT Open Source Software Center