Re: Add 64-bit XIDs into PostgreSQL 15

On Mon, Nov 28, 2022 at 1:30 PM Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> What is the purpose of using 64-bit XIDs, if not to avoid having to
> stop the world when we run short of XIDs?

I agree that the xidStopLimit mechanism was designed with the specific
goal of preventing "true" physical XID wraparound that results in
wrong answers to queries. It was not written with the intention of
limiting the accumulation of freeze debt, which would have to use
units like unfrozen heap pages to make any sense. That is a historic
fact -- no question.

I think that it is nevertheless quite possible that just refusing to
allocate any more XIDs could make sense with 64-bit XIDs, where we
don't strictly have to to keep the system fully functional. That might
still be the lesser evil, in that other world. The cutoff itself would
depend on many workload details, I suppose.

Imagine if we actually had 64-bit XIDs -- let's assume for a moment
that it's a done deal. This raises a somewhat awkward question: do you
just let the system get further and further behind on freezing,
forever? We can all agree that 2 billion XIDs is very likely the wrong
time to start refusing new XIDs -- because it just isn't designed with
debt in mind. But what's the right time, if any? How much debt is too
much?

At the very least these seem like questions that deserve serious consideration.

> I'd say that if this patch, or any patch with broadly similar goals,
> fails to remove xidStopLimit, it might as well not exist.

Well, it could in principle be driven by lots of different kinds of
information, and make better decisions by actually targeting freeze
debt in some way. An "enhanced version of xidStopLimit with 64-bit
XIDs" could kick in far far later than it currently would. Obviously
that has some value.

I'm not claiming to know how to build this "better xidStopLimit
mechanism", by the way. I'm not seriously proposing it. Mostly I'm
just saying that the question "where do you draw the line if not at 2
billion XIDs?" is a very pertinent question. It is not a question that
is made any less valid by the fact that we already know that 2 billion
XIDs is pretty far from optimal in almost all cases. Having some limit
based on something is likely to be more effective than having no limit
based on nothing.

Admittedly this argument works a lot better with the failsafe than it
does with xidStopLimit. Both are removed by the patch.

--
Peter Geoghegan