[ reviving an old thread ]

On Sun, Apr 5, 2020 at 10:14 AM Tomas Vondra
<[email protected]> wrote:
> For the record, here is the relevant part of the Incremental Sort patch
> series, updating add_partial_path and add_partial_path_precheck to also
> consider startup cost.
>
> The changes in the first two patches are pretty straight-forward, plus
> there's a proposed optimization in the precheck function to only run
> compare_pathkeys if entirely necessary. I'm currently evaluating those
> changes and I'll post the results to the incremental sort thread.

I rediscovered this problem while testing pg_plan_advice, and I think
we should commit something to fix it. Consider the following query
from the regression tests:

explain (costs off) select a,b,sum(c) from t group by 1,2 order by
1,2,3 limit 1;

With the settings in use at the time the plan is executed, the plan
has a cost of 68.11. If you rerun the same query with
enable_seqscan=off, the estimated cost drops to 65.28. This is
obviously a bit of a surprising outcome, since we should be
considering a subset of the original possibilities and should
therefore not be able to achieve a better outcome. What happens is
that the planner is deciding between Finalize GroupAggregate => Gather
Merge => Partial GroupAggregate => Incremental Sort => Parallel Index
Scan and GroupAggregate => Gather Merge => Incremental Sort =>
Parallel Index Scan. When we're computing partial paths for
partially_grouped_rel, we normally construct two: one that does
Incremental Sort => Parallel Index Scan and another that does Sort =>
Parallel Seq Scan. Because add_partial_path() doesn't consider startup
costs, the second path causes the first one to be discarded. Once we
add the Gather Merge node, we do consider startup costs (since it's
not a partial path any more) and now the high startup cost of Sort =>
Parallel Seq Scan makes it lose out. Instead, we abandon the idea of
two-step aggregation altogether and pick a plan that involves one step
aggregation. To me, this is a pretty clear illustration that the
comments that I wrote here are just wrong, the reasoning is wrong, and
what we're doing doesn't make a whole lot of sense. We're throwing
away paths that would turn out to be the winning option if we kept
them on the basis of a specious argument that startup cost doesn't
matter.

So, I started working on Tomas's patches from April 5, 2020. They
don't apply any more, so I had to rebase them. That resulted in a
bunch of regression test failures, and some of those plan changes
didn't make a whole lot of sense. At that point, I realized that when
I added the disabled_nodes field to the Path structure, I failed to
properly adjust the logic in add_partial_path() for the fact that we
now keep path lists sorted first by number of disabled nodes and then
by total cost. After fixing that, there were still two tests failing.
One was the query above, where the plan improves. In the other case, a
Hash Left Join became a Hash Right Join, and the reason that happened
is because the two paths had exactly equivalent total costs, but the
Hash Right Join has a cheaper startup cost. Arguably that's an
improvement, too, but it defeated the purpose of the test case, so
that will need to be adjusted. On further study, I found yet another
bug, which is that add_partial_path_precheck() *also* wasn't properly
adjusted to deal with the disabled_nodes field. In other words,
everything that is wrong here is 100% my fault: some of it I did wrong
when I added parallel query, and the rest of it I did wrong when I
added disabled_nodes.

So attached are a couple of patches to try to improve things. 0001
fixes the lack of a disabled_nodes test in add_partial_path. 0002
rewrites teaches add_partial_path about startup cost, and also
rewrites add_partial_path_precheck to do the cost comparisons in the
same manner as we do in compare_path_costs_fuzzily. This fixes both
the failure of that function to consider disabled_nodes, and also the
failure of that function to consider startup_cost. These are formally
two separate bugs, but I do not really relish the idea of fixing them
separately, because it means rewriting add_partial_path_precheck()
twice, and I think the ending state will be the same as what I have
here, and the intermediate state will be of no use to anyone and
possibly buggy as well. The only argument I can see for trying to
separate the two fixes to add_partial_path_precheck() is of one of the
following two things is true:

(1) We want to back-patch the disable_cost portion of the fix. This
has the possibility of destabilizing plans in released branches, so I
assume this is not going to be very appealing.

(2) We don't actually want the changes to consider startup cost. In
that case, this will definitely need to be changed. There's certainly
an argument that considering the startup cost increases the number of
paths we retain and therefore increases planning cost, but I feel like
that argument should lose out to the counter-argument that what we're
doing now is based on the fiction that the startup cost *can't* matter
in the case of a partial path, which we now know is not true: James
Coleman found that originally when working on Incremental Sort, and we
now have a regression test query that demonstrates it. If we're going
to skip considering startup cost in certain cases to reduce path
explosion, it should be based on a principled design choice, which
doesn't appear to be the case here. Rather, it appears to be based on
vintage-2016 Robert writing some stuff that wasn't true, and then we
kinda just rolled with it.

--
Robert Haas
EDB: http://www.enterprisedb.com

Attachment: v55-0001-Fix-add_partial_path-interaction-with-disabled_n.patch
Description: Binary data

Attachment: v55-0002-Consider-startup-cost-as-a-figure-of-merit-for-p.patch
Description: Binary data

Reply via email to