Re: Avoiding pin scan during btree vacuum

Robert Haas wrote:
> On Wed, Oct 19, 2016 at 6:30 PM, Alvaro Herrera
> <alvherre(at)2ndquadrant(dot)com> wrote:
> > Robert Haas wrote:
> >> On Mon, Jan 4, 2016 at 10:30 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> >> >> This seems like a might subtle thing to backpatch. If we really want to
> >> >> go there, ISTM that the relevant code should stew in an unreleased
> >> >> branch for a while, before being backpatched.
> >> >
> >> > I'm definitely -1 on back-patching such a thing. Put it in HEAD for
> >> > awhile. If it survives six months or so then we could discuss it again.
> >>
> >> I agree with Tom.
> >
> > Okay, several months have passed with this in the development branch and
> > now seems a good time to backpatch this all the way back to 9.4.
> >
> > Any objections?
>
> Although the code has now been in the tree for six months, it's only
> been in a released branch for three weeks, which is something about
> which to think.

The objection above was "stew in an unreleased branch", which it has.

> I guess what's really bothering me is that I don't think this is a bug
> fix. It seems like a performance optimization. And I think that we
> generally have a policy that we don't back-patch performance
> optimizations. Of course, there is some fuzziness because when the
> performance gets sufficiently bad, we sometimes decide that it amounts
> to a bug, as in 73cc7d3b240e1d46b1996382e5735a820f8bc3f7. Maybe this
> case qualifies for similar treatment, but I'm not sure.

I have seen a number of situations in which the standby strangely lags
behind seemingly randomly sometimes for very long periods of time,
without any apparent cause, without any possible remedy, and it makes
the standby unusable. If the user happens to monitor the lag they may
notice it and some may decide not to run certain queries. In other
cases the I/O load is so bad that queries that otherwise run without a
hitch are stuck for long.

In my opinion, this is a serious problem.

> Why are you talking about back-patching this to 9.4 rather than all
> supported branches?

As far as I understand this is dependent on catalog scans being MVCC,
so it cannot be backpatched any further than that.

--
Álvaro Herrera https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services