Thread: Expand applicability of aggregate's sortop optimization
Hi, As you may know, aggregates like SELECT MIN(unique1) FROM tenk1; are rewritten as SELECT unique1 FROM tenk1 ORDER BY unique1 USING < LIMIT 1; by using the optional sortop field in the aggregator. However, this optimization is disabled for clauses that in itself have an ORDER BY clause such as `MIN(unique1 ORDER BY <anything>), because <anything> can cause reordering of distinguisable values like 1.0 and 1.00, which then causes measurable differences in the output. In the general case, that's a good reason to not apply this optimization, but in some cases, we could still apply the index optimization. One of those cases is fixed in the attached patch: if we order by the same column that we're aggregating, using the same order class as the aggregate's sort operator (i.e. the aggregate's sortop is in the same btree opclass as the ORDER BY's sort operator), then we can still use the index operation: The sort behaviour isn't changed, thus we can apply the optimization. PFA the small patch that implements this. Note that we can't blindly accept just any ordering by the same column: If we had an opclass that sorted numeric values by the length of the significant/mantissa, then that'd provide a different (and distinct) ordering from that which is expected by the normal min()/max() aggregates for numeric, which could cause us to return arguably incorrect results for the aggregate expression. Alternatively, the current code could be changed to build indexed paths that append the SORT BY paths to the aggregate's sort operator, but that'd significantly increase complexity here. Kind regards, Matthias van de Meent Neon (https://neon.tech)
Attachment
Matthias van de Meent <boekewurm+postgres@gmail.com> writes:
> PFA the small patch that implements this.
I don't have enough knowledge to have an opinion on most of the patch
other than it looks okay at a glance, but the list API usage could be
updated to more modern variants:
> diff --git a/src/backend/optimizer/plan/planagg.c b/src/backend/optimizer/plan/planagg.c
> index afb5445b77..d8479fe286 100644
> --- a/src/backend/optimizer/plan/planagg.c
> +++ b/src/backend/optimizer/plan/planagg.c
> @@ -253,6 +253,16 @@ can_minmax_aggs(PlannerInfo *root, List **context)
> if (list_length(aggref->args) != 1)
> return false; /* it couldn't be MIN/MAX */
>
> + /*
> + * We might implement the optimization when a FILTER clause is present
> + * by adding the filter to the quals of the generated subquery. For
> + * now, just punt.
> + */
> + if (aggref->aggfilter != NULL)
> + return false;
> +
> + curTarget = (TargetEntry *) linitial(aggref->args);
This could be linitial_node(TargetEntry, aggref->args).
> + if (list_length(aggref->aggorder) > 1)
> + return false;
> +
> + orderClause = castNode(SortGroupClause, linitial(aggref->aggorder));
This could be linitial_node(SortGroupClause, aggref->aggorder).
> + if (orderClause->sortop != aggsortop)
> + {
> + List *btclasses;
> + ListCell *cell;
> + bool match = false;
> +
> + btclasses = get_op_btree_interpretation(orderClause->sortop);
> +
> + foreach(cell, btclasses)
> + {
> + OpBtreeInterpretation *interpretation;
> + interpretation = (OpBtreeInterpretation *) lfirst(cell);
This could be foreach_ptr(OpBtreeInterpretation, interpretation, btclasses),
which also eliminates the need for the explicit `ListCell *` variable
and lfirst() call.
- ilmari
On Wed, 8 May 2024 at 22:13, Matthias van de Meent <boekewurm+postgres@gmail.com> wrote: > As you may know, aggregates like SELECT MIN(unique1) FROM tenk1; are > rewritten as SELECT unique1 FROM tenk1 ORDER BY unique1 USING < LIMIT > 1; by using the optional sortop field in the aggregator. > However, this optimization is disabled for clauses that in itself have > an ORDER BY clause such as `MIN(unique1 ORDER BY <anything>), because > <anything> can cause reordering of distinguisable values like 1.0 and > 1.00, which then causes measurable differences in the output. In the > general case, that's a good reason to not apply this optimization, but > in some cases, we could still apply the index optimization. I wonder if we should also consider as an alternative to this to just have an aggregate support function, similar to SupportRequestOptimizeWindowClause that just nullifies the aggorder / aggdistinct fields for Min/Max aggregates on types where there's no possible difference in output when calling the transition function on rows in a different order. Would that apply in enough cases for you? I think it would rule out Min(numeric) and Max(numeric). We were careful not to affect the number of decimal places in the numeric output when using the moving aggregate inverse transition infrastructure for WindowFuncs, so I agree we should maintain an ability to control the aggregate transition order for numeric. (See do_numeric_discard's maxScale if check) I don't think floating point types have the same issues here. At least +1.0 is greater than -1.0. Are there any strange collation rules that would cause issues if we did this with the text types? David
On Thu, 9 May 2024 at 12:26, David Rowley <dgrowleyml@gmail.com> wrote: > I wonder if we should also consider as an alternative to this to just > have an aggregate support function, similar to > SupportRequestOptimizeWindowClause that just nullifies the aggorder / > aggdistinct fields for Min/Max aggregates on types where there's no > possible difference in output when calling the transition function on > rows in a different order. > > Would that apply in enough cases for you? One additional thought is that the above method would also help eliminate redundant sorting in queries with a GROUP BY clause. Whereas, the can_minmax_aggs optimisation is not applied in that case. David
On Thu, 9 May 2024 at 13:08, David Rowley <dgrowleyml@gmail.com> wrote: > One additional thought is that the above method would also help > eliminate redundant sorting in queries with a GROUP BY clause. > Whereas, the can_minmax_aggs optimisation is not applied in that case. Another argument for using this method is that SupportRequestOptimizeAggref could allow future unrelated optimisations such as swapping count(<non-nullable-col>) for count(*). Where <non-nullable-col> is a NOT NULL column and isn't nullable by any outer join. Doing that could speed some queries up quite a bit as it may mean fewer columns to deform from the tuple. You could imagine a fact table with many columns and a few dimensions, often the dimension columns that you'd expect to use in GROUP BY would appear before the columns you'd aggregate on. David
On 5/8/24 17:13, Matthias van de Meent wrote: > As you may know, aggregates like SELECT MIN(unique1) FROM tenk1; are > rewritten as SELECT unique1 FROM tenk1 ORDER BY unique1 USING < LIMIT > 1; by using the optional sortop field in the aggregator. > However, this optimization is disabled for clauses that in itself have > an ORDER BY clause such as `MIN(unique1 ORDER BY <anything>), because > <anything> can cause reordering of distinguisable values like 1.0 and > 1.00, which then causes measurable differences in the output. In the > general case, that's a good reason to not apply this optimization, but > in some cases, we could still apply the index optimization. Thanks for the job! I guess you could be more brave and push down also FILTER statement. > > One of those cases is fixed in the attached patch: if we order by the > same column that we're aggregating, using the same order class as the > aggregate's sort operator (i.e. the aggregate's sortop is in the same > btree opclass as the ORDER BY's sort operator), then we can still use > the index operation: The sort behaviour isn't changed, thus we can > apply the optimization. > > PFA the small patch that implements this. > > Note that we can't blindly accept just any ordering by the same > column: If we had an opclass that sorted numeric values by the length > of the significant/mantissa, then that'd provide a different (and > distinct) ordering from that which is expected by the normal > min()/max() aggregates for numeric, which could cause us to return > arguably incorrect results for the aggregate expression. As I see, the code: aggsortop = fetch_agg_sort_op(aggref->aggfnoid); if (!OidIsValid(aggsortop)) return false; /* not a MIN/MAX aggregate */ used twice and can be evaluated earlier to avoid duplicated code. Also, I'm unsure about the necessity of looking through the btree classes. Maybe just to check the commutator to the sortop, like in the diff attached? Or could you provide an example to support your approach? -- regards, Andrei Lepikhov
Attachment
On 5/9/24 08:08, David Rowley wrote: > On Thu, 9 May 2024 at 12:26, David Rowley <dgrowleyml@gmail.com> wrote: >> I wonder if we should also consider as an alternative to this to just >> have an aggregate support function, similar to >> SupportRequestOptimizeWindowClause that just nullifies the aggorder / >> aggdistinct fields for Min/Max aggregates on types where there's no >> possible difference in output when calling the transition function on >> rows in a different order. >> >> Would that apply in enough cases for you? > > One additional thought is that the above method would also help > eliminate redundant sorting in queries with a GROUP BY clause. > Whereas, the can_minmax_aggs optimisation is not applied in that case. I generally like the idea of a support function. But as I can see, the can_minmax_aggs() rejects if any of the aggregates don't pass the checks. The prosupport feature is designed to be applied to each function separately. How do you think to avoid it? Also, I don't clearly understand the case you mentioned here - does it mean that you want to nullify orders for other aggregate types if they are the same as the incoming order? -- regards, Andrei Lepikhov
On Wed, 17 Jul 2024 at 05:29, Andrei Lepikhov <lepihov@gmail.com> wrote: > > On 5/8/24 17:13, Matthias van de Meent wrote: > > As you may know, aggregates like SELECT MIN(unique1) FROM tenk1; are > > rewritten as SELECT unique1 FROM tenk1 ORDER BY unique1 USING < LIMIT > > 1; by using the optional sortop field in the aggregator. > > However, this optimization is disabled for clauses that in itself have > > an ORDER BY clause such as `MIN(unique1 ORDER BY <anything>), because > > <anything> can cause reordering of distinguisable values like 1.0 and > > 1.00, which then causes measurable differences in the output. In the > > general case, that's a good reason to not apply this optimization, but > > in some cases, we could still apply the index optimization. > > Thanks for the job! I guess you could be more brave and push down also > FILTER statement. While probably not impossible, I wasn't planning on changing this code with new optimizations; just expanding the applicability of the current optimizations. Note that the "aggfilter" clause was not new, but moved up in the code to make sure we use this local information to bail out (if applicable) before trying to use the catalogs for bail-out information. > > > > One of those cases is fixed in the attached patch: if we order by the > > same column that we're aggregating, using the same order class as the > > aggregate's sort operator (i.e. the aggregate's sortop is in the same > > btree opclass as the ORDER BY's sort operator), then we can still use > > the index operation: The sort behaviour isn't changed, thus we can > > apply the optimization. > > > > PFA the small patch that implements this. > > > > Note that we can't blindly accept just any ordering by the same > > column: If we had an opclass that sorted numeric values by the length > > of the significant/mantissa, then that'd provide a different (and > > distinct) ordering from that which is expected by the normal > > min()/max() aggregates for numeric, which could cause us to return > > arguably incorrect results for the aggregate expression. > > As I see, the code: > aggsortop = fetch_agg_sort_op(aggref->aggfnoid); > if (!OidIsValid(aggsortop)) > return false; /* not a MIN/MAX aggregate */ > > used twice and can be evaluated earlier to avoid duplicated code. The code is structured like this to make sure we only start accessing catalogs once we know that all other reasons to bail out from this optimization indicate we can apply the opimization. You'll notice that I've tried to put the cheapest checks that only use caller-supplied information first, and catalog accesses only after that. If the fetch_agg_sort_op clause would be deduplicated, it would either increase code complexity to handle both aggref->aggorder paths, or it would increase the cost of planning MAX(a ORDER BY b) because of the newly added catalog access. > Also, I'm unsure about the necessity of looking through the btree > classes. Maybe just to check the commutator to the sortop, like in the > diff attached? Or could you provide an example to support your approach? I think it could work, but I'd be hesitant to rely on that, as commutator registration is optional (useful, but never required for btree operator classes' operators). Looking at the btree operator class, which is the definition of sortability in PostgreSQL, seems more suitable and correct. Kind regards, Matthias van de Meent Neon (https://neon.tech)
On 17/7/2024 16:33, Matthias van de Meent wrote: > On Wed, 17 Jul 2024 at 05:29, Andrei Lepikhov <lepihov@gmail.com> wrote: >> Thanks for the job! I guess you could be more brave and push down also >> FILTER statement. > > While probably not impossible, I wasn't planning on changing this code > with new optimizations; just expanding the applicability of the > current optimizations. Got it>> As I see, the code: >> aggsortop = fetch_agg_sort_op(aggref->aggfnoid); >> if (!OidIsValid(aggsortop)) >> return false; /* not a MIN/MAX aggregate */ >> >> used twice and can be evaluated earlier to avoid duplicated code. > > The code is structured like this to make sure we only start accessing > catalogs once we know that all other reasons to bail out from this > optimization indicate we can apply the opimization. You'll notice that > I've tried to put the cheapest checks that only use caller-supplied > information first, and catalog accesses only after that. > > If the fetch_agg_sort_op clause would be deduplicated, it would either > increase code complexity to handle both aggref->aggorder paths, or it > would increase the cost of planning MAX(a ORDER BY b) because of the > newly added catalog access. IMO it looks like a micro optimisation. But I agree, it is more about code style - let the committer decide what is better.>> Also, I'm unsure about the necessity of looking through the btree >> classes. Maybe just to check the commutator to the sortop, like in the >> diff attached? Or could you provide an example to support your approach? > > I think it could work, but I'd be hesitant to rely on that, as > commutator registration is optional (useful, but never required for > btree operator classes' operators). Looking at the btree operator > class, which is the definition of sortability in PostgreSQL, seems > more suitable and correct. Hm, I dubious about that. Can you provide an example which my variant will not pass but your does that correctly? -- regards, Andrei Lepikhov
On Wed, 17 Jul 2024 at 17:12, Andrei Lepikhov <lepihov@gmail.com> wrote: > I generally like the idea of a support function. But as I can see, the > can_minmax_aggs() rejects if any of the aggregates don't pass the > checks. The prosupport feature is designed to be applied to each > function separately. How do you think to avoid it? You wouldn't avoid it. The prosupport function would be called once for each Aggref in the query. Why is that a problem? > Also, I don't clearly understand the case you mentioned here - does it > mean that you want to nullify orders for other aggregate types if they > are the same as the incoming order? No, I mean unconditionally nullify Aggref->aggorder and Aggref->aggdistinct for aggregate functions where ORDER BY / DISTINCT in the Aggref makes no difference to the result. I think that's ok for max() and min() for everything besides NUMERIC. For aggorder, we'd have to *not* optimise sum() and avg() for floating point types as that could change the result. sum() and avg() for INT2, INT4 and INT8 seems fine. I'd need to check, but I think sum(numeric) is ok too as the dscale should end up the same regardless of the order. Obviously, aggdistinct can't be changed for sum() and avg() on any type. It seems also possible to adjust count(non-nullable-var) into count(*), which, if done early enough in planning could help significantly by both reducing evaluation during execution, but also possibly reduce tuple deformation if that Var has a higher varattno than anything else in the relation. That would require checking varnullingrels is empty and the respective RelOptInfo's notnullattnums mentions the Var. David
On 17/7/2024 16:33, Matthias van de Meent wrote: > On Wed, 17 Jul 2024 at 05:29, Andrei Lepikhov <lepihov@gmail.com> wrote: >> As I see, the code: >> aggsortop = fetch_agg_sort_op(aggref->aggfnoid); >> if (!OidIsValid(aggsortop)) >> return false; /* not a MIN/MAX aggregate */ >> >> used twice and can be evaluated earlier to avoid duplicated code. > > The code is structured like this to make sure we only start accessing > catalogs once we know that all other reasons to bail out from this > optimization indicate we can apply the opimization. You'll notice that > I've tried to put the cheapest checks that only use caller-supplied > information first, and catalog accesses only after that. After additional research I think I get the key misunderstanding why you did so: As I see, the checks: if (list_length(aggref->aggorder) > 1) return false; if (orderClause->tleSortGroupRef != curTarget->ressortgroupref) return false; not needed at all. You already have check: if (list_length(aggref->args) != 1) and this tells us, that if we have ordering like MIN(x ORDER BY <smth>), this <smth> ordering contains only aggregate argument x. Because if it contained some expression, the transformAggregateCall() would add this expression to agg->args by calling the transformSortClause() routine. The tleSortGroupRef is just exactly ressortgroupref - no need to recheck it one more time. Of course, it is suitable only for MIN/MAX aggregates, but we discuss only them right now. Am I wrong? If you want, you can place it as assertions (see the diff in attachment). -- regards, Andrei Lepikhov
Attachment
On Wed, 17 Jul 2024 at 16:09, Andrei Lepikhov <lepihov@gmail.com> wrote: > > On 17/7/2024 16:33, Matthias van de Meent wrote: > > On Wed, 17 Jul 2024 at 05:29, Andrei Lepikhov <lepihov@gmail.com> wrote: > >> Thanks for the job! I guess you could be more brave and push down also > >> FILTER statement. > > > > While probably not impossible, I wasn't planning on changing this code > > with new optimizations; just expanding the applicability of the > > current optimizations. > Got it>> As I see, the code: > >> aggsortop = fetch_agg_sort_op(aggref->aggfnoid); > >> if (!OidIsValid(aggsortop)) > >> return false; /* not a MIN/MAX aggregate */ > >> > >> used twice and can be evaluated earlier to avoid duplicated code. > > > > The code is structured like this to make sure we only start accessing > > catalogs once we know that all other reasons to bail out from this > > optimization indicate we can apply the opimization. You'll notice that > > I've tried to put the cheapest checks that only use caller-supplied > > information first, and catalog accesses only after that. > > > > If the fetch_agg_sort_op clause would be deduplicated, it would either > > increase code complexity to handle both aggref->aggorder paths, or it > > would increase the cost of planning MAX(a ORDER BY b) because of the > > newly added catalog access. > IMO it looks like a micro optimisation. But I agree, it is more about > code style - let the committer decide what is better.>> Also, I'm unsure > about the necessity of looking through the btree > >> classes. Maybe just to check the commutator to the sortop, like in the > >> diff attached? Or could you provide an example to support your approach? > > > > I think it could work, but I'd be hesitant to rely on that, as > > commutator registration is optional (useful, but never required for > > btree operator classes' operators). Looking at the btree operator > > class, which is the definition of sortability in PostgreSQL, seems > > more suitable and correct. > Hm, I dubious about that. Can you provide an example which my variant > will not pass but your does that correctly? Here is one: """ CREATE OPERATOR @@> ( function=int4gt, leftarg=int4, rightarg=int4 ); CREATE OPERATOR @@>= ( function=int4ge, leftarg=int4, rightarg=int4 ); CREATE OPERATOR @@= ( function=int4eq, leftarg=int4, rightarg=int4 ); CREATE OPERATOR @@<= ( function=int4le, leftarg=int4, rightarg=int4 ); CREATE OPERATOR @@< ( function=int4lt, leftarg=int4, rightarg=int4 ); CREATE OPERATOR CLASS my_int_ops FOR TYPE int USING btree AS OPERATOR 1 @<@, OPERATOR 2 @<=@, OPERATOR 3 @=@, OPERATOR 4 @>=@, OPERATOR 5 @>@, FUNCTION 1 btint4cmp; CREATE AGGREGATE my_int_max ( BASETYPE = int4, SFUNC = int4larger, STYPE = int4, SORTOP = @>@ ); CREATE TABLE my_table AS SELECT id::int4 FROM generate_series(1, 10000) id; CREATE INDEX ON my_table (id my_int_ops); SELECT my_int_max(id ORDER BY id USING @<@ ) from my_table; """ Because the @<@ and @>@ operators are not registered as commutative, it couldn't apply the optimization in your patch, while the btree operator check does allow it to apply the optimization. Aside: Arguably, checking for commutator operators would not be incorrect when looking at it from "applied operators" point of view, but if that commutative operator isn't registered as opposite ordering of the same btree opclass, then we'd probably break some assumptions of some aggregate's sortop - it could be registered with another opclass, and that could cause us to select a different btree opclass (thus: ordering) than is indicated to be required by the aggregate; the thing we're trying to protect against here. Kind regards, Matthias van de Meent Neon (https://neon.tech)
On 18/7/2024 19:49, Matthias van de Meent wrote: > On Wed, 17 Jul 2024 at 16:09, Andrei Lepikhov <lepihov@gmail.com> wrote: >> >> On 17/7/2024 16:33, Matthias van de Meent wrote: >>> On Wed, 17 Jul 2024 at 05:29, Andrei Lepikhov <lepihov@gmail.com> wrote: > Because the @<@ and @>@ operators are not registered as commutative, > it couldn't apply the optimization in your patch, while the btree > operator check does allow it to apply the optimization. Ok, I got it. And next issue: I think it would be better to save cycles than to free some piece of memory, so why not to break the foreach cycle if you already matched the opfamily? Also, in the patch attached I added your smoothed test to the aggregates.sql > > Aside: Arguably, checking for commutator operators would not be > incorrect when looking at it from "applied operators" point of view, > but if that commutative operator isn't registered as opposite ordering > of the same btree opclass, then we'd probably break some assumptions > of some aggregate's sortop - it could be registered with another > opclass, and that could cause us to select a different btree opclass > (thus: ordering) than is indicated to be required by the aggregate; > the thing we're trying to protect against hereYes, I also think if someone doesn't register < as a commutator to >, it may mean they do it intentionally: may be it is a bit different sortings? - this subject is too far from my experience and I can agree with your approach. -- regards, Andrei Lepikhov
Attachment
On 18/7/2024 14:49, Matthias van de Meent wrote: > Aside: Arguably, checking for commutator operators would not be > incorrect when looking at it from "applied operators" point of view, > but if that commutative operator isn't registered as opposite ordering > of the same btree opclass, then we'd probably break some assumptions > of some aggregate's sortop - it could be registered with another > opclass, and that could cause us to select a different btree opclass > (thus: ordering) than is indicated to be required by the aggregate; > the thing we're trying to protect against here. Hi, This thread stands idle. At the same time, the general idea of this patch and the idea of utilising prosupport functions look promising. Are you going to develop this feature further? -- regards, Andrei Lepikhov
On Tue, 8 Oct 2024 at 18:47, Andrei Lepikhov <lepihov@gmail.com> wrote: > Thanks, I suppose CATALOG_VERSION_NO was the only reason for this fail. Please leave the cat version bump out of your patch. It's a waste of time and resource if you plan to post another patch every time a committer bumps the cat version. David