On Thu, Aug 25, 2016 at 12:16 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I'm not following. The only insertions happening in this test case
> are coming from various clients doing the same INSERT ON CONFLICT UPDATE.
> If one of them has successfully inserted "42" into the arbiter index,
> how is it that other ones are not seeing that?
It's complicated. One backend may completely overtake another in this
race. We don't give up early within ExecInsertIndexTuples(), because
there isn't any useful distinction made between the speculative
insertion case, and the deferred constraint case. The ordering, as
such, is therefore irrelevant.
Fortunately, I posted a fix, of sorts, more than a year ago:
https://commitfest.postgresql.org/10/403/
It never occurred to me to make this argument for it.
There is a separate question of how to make the ordering avoid
problems if it did matter (if that patch ever got committed). I think
it would fix this exact test case, but only accidentally, because the
executor gets a list of IndexInfo structs from the relcache in a
consistent though fairly insignificant order (ordered by OID). I don't
think you can change that property within the relcache, because the
ordering must be consistent (to avoid deadlocks, perhaps elsewhere).
Maybe we should change the ordering of those IndexInfo structs to
something more suitable, but it must be immutable (it cannot hinge
upon the details of one particular DML statement). I actually also
wrote a patch to prefer insertion into the primary key first, which
also went nowhere (I gave up on that one, to be fair).
--
Peter Geoghegan