On Tue, Mar 29, 2022 at 5:01 PM Andres Freund <andres@anarazel.de> wrote:
> I think it's reasonably rare because in cases there'd be corruption, we'd
> typically not even have written them out / throw them away explicitly - we
> only read stats when starting without crash recovery.
>
> So the "expected" case of corruption afaicts solely is a OS crash just after
> the shutdown checkpoint completed?
Can we prevent that case from occurring, so that there are no expected cases?
> > And maybe we should have, inside the stats system, something that
> > keeps track of when the stats file was last recreated from scratch because
> > of a corruption event, separately from when it was last intentionally reset.
>
> That would be easy to add. Don't think we have a good place to show the
> information right now - perhaps just new functions not part of any view?
I defer to you on where to put it.
> I can think of these different times:
>
> - Last time stats were removed due to starting up in crash recovery
> - Last time stats were created from scratch, because no stats data file was
> present at startup
> - Last time stats were thrown away due to corruption
> - Last time a subset of stats were reset using one of the pg_reset* functions
>
> Makes sense?
Yes. Possibly that last could be broken in to two: when all stats were
last reset, when some stats were last reset.
> > Does redo update the stats?
>
> With "update" do you mean generate new stats? In the shared memory stats patch
> it triggers stats to be dropped, on HEAD it just resets all stats at startup.
>
> Redo itself doesn't generate stats, but bgwriter, checkpointer, backends do.
Well, I guess what I'm trying to figure out is what happens if we run
in recovery for a long time -- say, a year -- and then get promoted.
Do we have reasons to expect that the stats will be accurate enough to
use at that point, or will they be way off?
I don't have a great understanding of how this all works, but if
running recovery for a long time is going to lead to a situation where
the stats progressively diverge from reality, then preserving them
doesn't seem as valuable as if they're going to be more or less
accurate.
--
Robert Haas
EDB: http://www.enterprisedb.com
