On Sat, May 30, 2015 at 11:10 AM, Robert Haas <robertmhaas@gmail.com> wrote:
>> Let me share something that people have told me privately but don't want
>> to state publicly (at least with attribution), and that is that we have
>> seen great increases in feature development (often funded), without a
>> corresponding increase development efforts focused on stability. The
>> fact Alvaro has had to almost single-handedly fix multi-xact bug until
>> very recently is testament to that.
>
> It's clear - at least to me - that we need to put more resources into
> stabilizing the new multixact system. This is killing us. If we can't
> stabilize this, people will go use some other database.
+1. I don't grok the MultiXact code as some people do, but even still,
I think problems have been ongoing for so long now that we must change
course. FWIW, my perception from afar is that the problems haven't
really tapered off, and we'd be better off taking a fresh approach.
> Equally importantly, we need to make sure that we never release
> something comparably broken ever again. And that's why I'm not
> sanguine about shipping what we've got without adequate reflection.
As you said, there was a failure to appreciate the interactions with
VACUUM. That should have made us more introspective about what we
didn't know and couldn't know during during 9.3 development, but it
didn't.
> What, in this release, could break things badly? RLS? Grouping sets?
> Heikki's WAL format changes? That last one sounds really scary to me;
> it's painful if not impossible to fix the WAL format in a minor
> release.
I think we actually have learned some lessons here. MultiXacts were a
somewhat unusual case for a couple of reasons that I need not rehash.
In contrast, Heikki's WAL format changes (just for example) are
fundamentally just a restructuring to the existing format. Sure, there
could be bugs, but I think that it's fundamentally different to the
9.3 MultiXact stuff, in that the MultiXact stuff appears to be
stubbornly difficult to stabilize over months and years. That feels
like something that is unlikely to be true for anything that made it
into 9.5.
--
Peter Geoghegan