On Thu, May 16, 2019 at 2:03 PM Andres Freund <andres@anarazel.de> wrote:
Hi,
On 2019-05-16 13:59:47 -0700, Melanie Plageman wrote: > On Wed, May 15, 2019 at 10:32 PM Ashwin Agrawal <aagrawal@pivotal.io> wrote: > > > > > The second index would help to hold the session after inserting the tuple > > in unique index but before completing the speculative insert. Hence, helps > > to create the condition easily. I believe order of index insertion is > > helping here that unique index is inserted and then non-unique index is > > inserted too. > > > > > Oh, cool. I didn't know that execution order would be guaranteed for which > index > to insert into first.
It's not *strictly* speaking *always* well defined. The list of indexes is sorted by the oid of the index - so once created, it's consistent. But when the oid assignment wraps around, it'd be the other way around. But I think it's ok to disregard that - it'll never happen in regression tests run against a new cluster, and you'd have to run tests against an installed cluster for a *LONG* time for a *tiny* window where the wraparound would happen precisely between the creation of the two indexes.
Makes sense?
Yep, thanks.
I guess we could make that case a tiny bit easier to diagnose in the extremely unlikely case it happens by having a step that outputs SELECT 'index_a'::regclass::int8 < 'index_b'::regclass::int8;
Good idea.
I squashed the changes I suggested in previous emails, Ashwin's patch, my suggested updates to that patch, and the index order check all into one updated patch attached.