Hi,
On 2022-07-15 16:42:14 -0400, Jonathan S. Katz wrote:
> On 7/15/22 4:36 PM, Tomas Vondra wrote:
> > > > If there does not appear to be a clear path forward, we should at least
> > > > document how a user can detect and resolve the issue.
> > >
> > > To me that doesn't really make sense. We have lots of places were performance
> > > changes once you cross some threshold, and lots of those are related to
> > > work_mem.
>
> Yes, but in this case this is nonobvious to a user. A sort that may be
> performing just fine on a pre-PG15 version is suddenly degraded, and the
> user has no guidance as to why or how to remediate.
We make minor changes affecting thresholds at which point things spill to disk
etc *all the time*. Setting the standard that all of those need to be
documented seems not wise to me. Both because of the effort for us, and
because it'll end up being a morass of documentation that nobody can make use
off, potentially preventing people from upgrading because some minor perf
changes in some edge cases sound scary.
I'm fairly certain there were numerous other changes with such effects. We
just don't know because there wasn't as careful benchmarking.
> > > We don't, e.g., provide tooling to detect when performance in aggregation
> > > regresses due to crossing work_mem and could be fixed by a tiny increase in
> > > work_mem.
> > >
> >
> > Yeah. I find it entirely reasonable to tell people to increase work_mem
> > a bit to fix this. The problem is knowing you're affected :-(
>
> This is the concern the RMT discussed. Yes it is reasonable to say "increase
> work_mem to XYZ", but how does a user know how to detect and apply that?
The won't. Nor will they with any reasonable documentation we can give. Nor
will they know in any of the other cases some threshold is crossed.
> This is the part that is worrisome, especially because we don't have any
> sense of what the overall impact will be.
>
> Maybe it's not much, but we should document that there is the potential for
> a regression.
-1
Greetings,
Andres Freund