On Sat, Oct 3, 2020 at 5:44 PM Tomas Vondra
<tomas.von...@2ndquadrant.com> wrote:
>
> On Sat, Oct 03, 2020 at 10:50:06AM -0400, James Coleman wrote:
> >On Fri, Oct 2, 2020 at 11:16 PM James Coleman <jtc...@gmail.com> wrote:
> >>
> >> On Fri, Oct 2, 2020 at 7:07 PM James Coleman <jtc...@gmail.com> wrote:
> >> >
> >> > On Fri, Oct 2, 2020 at 6:28 PM Tomas Vondra
> >> > <tomas.von...@2ndquadrant.com> wrote:
> >> > >
> >> > > On Fri, Oct 02, 2020 at 05:45:52PM -0400, James Coleman wrote:
> >> > > >On Fri, Oct 2, 2020 at 4:56 PM Tomas Vondra
> >> > > ><tomas.von...@2ndquadrant.com> wrote:
> >> > > >>
> >> > > >> ...
> >> > > >>
> >> > > >> More importanly, it does not actually fix the issue - it does fix
> >> > > >> that
> >> > > >> particular query, but just replacing the DISTINCT with either ORDER
> >> > > >> BY
> >> > > >> or GROUP BY make it fail again :-(
> >> > > >>
> >> > > >> Attached is a simple script I used, demonstrating these issues for
> >> > > >> the
> >> > > >> three cases that expect to have ressortgroupref != 0 (per the
> >> > > >> comment
> >> > > >> before TargetEntry in plannodes.h).
> >> > > >
> >> > > >So does checking for volatile expressions (if you happened to test
> >> > > >that) solve all the cases? If you haven't tested that yet, I can try
> >> > > >to do that this evening.
> >> > > >
> >> > >
> >> > > Yes, it does fix all the three queries in the SQL script.
> >> > >
> >> > > The question however is whether this is the root issue, and whether
> >> > > it's
> >> > > the right way to fix it. For example - volatility is not the only
> >> > > reason
> >> > > that may block qual pushdown. If you look at qual_is_pushdown_safe, it
> >> > > also blocks pushdown of leaky functions in security_barrier views. So I
> >> > > wonder if that could cause failures too, somehow. But I haven't managed
> >> > > to create such example.
> >> >
> >> > I was about to say that the issue here is slightly different from qual
> >> > etc. pushdown, since we're not concerned about quals here, and so I
> >> > wonder where we determine what target list entries to put in a given
> >> > scan path, but then I realized that implies (maybe!) a simpler
> >> > solution. Instead of duplicating checks on target list entries would
> >> > be safe, why not check directly in get_useful_pathkeys_for_relation()
> >> > whether or not the pathkey has a target list entry?
> >> >
> >> > I haven't been able to try that out yet, and so maybe I'm missing
> >> > something, but I need to step out for a bit, so I'll have to look at
> >> > it later.
> >>
> >> I've started poking at this, but haven't yet found a workable
> >> solution. See the attached patch which "solves" the problem by
> >> breaking putting the sort under the gather merge, but it at least
> >> demonstrates conceptually what I think we're interested in doing.
> >>
> >> The issue currently is that the comparison of expressions fails -- in
> >> our query, the first column selected shows up as a Var (with the same
> >> contents) but is a different pointer in the em_expr and the reltarget
> >> exprs list. I don't yet see a good way to get the equivalence class
> >> for a PathTarget entry.
> >
> >Hmm, I think I was looking to do is the attached patch. I didn't
> >realize until I did a lot of reading through source that we have an
> >equal() function that can compare exprs. That (plus the realization in
> >[1] the originally reported backtrace wasn't where the error actually
> >came from) convinced me that what we need is to confirm not only that
> >the all of the vars in the ec member come from the relids in the rel
> >but also that the expr is actually represented in the target of the
> >rel.
> >
> >With the gucs changed as I mentioned earlier both of the plans (with
> >and without a volatile call in the 2nd select entry) now look like:
> >
> > Unique
> > -> Gather Merge
> > Workers Planned: 2
> > -> Sort
> > Sort Key: ref_0.i, (md5(ref_0.t))
> > -> Nested Loop
> > -> Parallel Index Scan using ref_0_i_idx on ref_0
> > -> Function Scan on generate_series ref_1
> >
> >Without the gucs changed the minimal repro case now doesn't error, but
> >results in this plan:
> >
> > HashAggregate
> > Group Key: ref_0.i, CASE WHEN pg_rotate_logfile_old() THEN ref_0.t
> >ELSE ref_0.t END
> > -> Nested Loop
> > -> Seq Scan on ref_0
> > -> Function Scan on generate_series ref_1
> >
> >Similarly in your six queries I now only see parallel query showing up
> >in the last one.
> >
>
> OK, that seems reasonable I think.
If we proceed with this patch I'd like to tweak it to check for the
relids subset first on the theory that it will be cheaper than the
expr equality check loop; that change is attached.
> >I created an entirely new function because adding the target expr
> >lookup to the existing find_em_expr_for_rel() function broke a bunch
> >of postgres_fdw tests. That maybe raises questions about whether that
> >code also could have problems in theory/in the future, but I didn't
> >investigate further. In any case we already know it excludes
> >volatile...so maybe it's fine because in practice that's actually a
> >broader exclusion than what we're doing here.
> >
>
> I don't think postgres_fdw needs these new checks, because FDW scan's
> are not allowed in parallel part of the plan - if that changes in the
> future, I'm sure that'll require fixes in plenty other places.
>
> OTOH it's interesting that it triggers those failures - I wonder if we
> could learn something from them? AFAICS those paths can't be built by
> generate_useful_gather_paths (because of the postgres_fdw vs. parallel
> query restrictions), so how do these plans look?
Well the fdw code doesn't trigger the errors, but both
generate_useful_gather_paths() and the fdw code originally relied on
find_em_expr_for_rel(), but now they're diverging. IIRC the fdw code
also uses it for determining what path keys are valid -- in that case
which ones are safe to be sent to the foreign server (so all of the
vars have to be from rels on the foreign server). The fdw code has
additional checks too though, for example no volatile expressions may
be pushed to the remote.
> >This seems to fix the issue, but I'd like feedback on whether it's too
> >strict. We could of course just check em_has_volatile, but I'm
> >wondering if that's simultaneously too strict (by not allowing the
> >volatile expression to be computed in the gather merge supposing
> >there's no join) and too loose (maybe there are other cases we should
> >care about?). It also just strikes me as re-encoding rules that should
> >have already been applied (and thus we should be able to look up in
> >the data we have if it's safe to use the expr/pathkey). Those are
> >mostly intuitions though.
> >
>
> I don't know :-( As I mentioned before, I suspect checking just the
> volatility may not be enough in some cases (leakyness, etc.) and I think
> you're right it may be too strict in other cases.
>
> Not sure I understand which rules you think we're re-encoding, but I
> have a nagging feeling there's a piece of code somewhere earlier in
> the query planner meant to prevent such cases (sort on path without all
> the pathkeys), and that fixing it here is just a band aid :-(
I'm getting at exactly what you're saying below: there's got to be
some existing rules (and presumably code already enforcing it in some
places) at what level in the path tree a given pathkey is
safe/correct, and we don't want to reinvent those rules (or ideally
that code). Verifying the expr is in the rel seems to me to be
intuitively correct, but I wish there was another place in the code
that could validate that.
> I'm sure we're building plans with Sort on top of Index Scan paths, so
> how come those don't fail? How come the non-parallel version of these
> queries don't fail?
Yeah, exactly. Something has to be doing something similar -- I don't
think everywhere else using sort_pathkeys instead of query_pathkeys
really changes the problem. I've read through all of the places we use
generate_useful_gather_paths() as well as root->sort_pathkeys to try
to get a better understanding of this. Most of the places I convinced
myself it's already safe -- for example places like
create_orderd_paths are working with the upper rel, and the full
target list is available. A few other places I added comments (see the
"questions" patch in the attached series) where I'm not sure why we
apply projections *after* we create a sort path; given the issue we're
discussing here I would have thought you'd want to do exactly the
opposite.
I haven't yet tried to read through the code in other places where we
build an explicit sort to see if I can find any hints as to what
existing code does to ensure this situation doesn't occur, but so far
I haven't found anything else obvious.
Is there someone who might be able to point us in the right
direction/validate what we're thinking?
James
From e7b94d4efda41becb2f28123eca2db2a88d0d4d0 Mon Sep 17 00:00:00 2001
From: James Coleman <jtc...@gmail.com>
Date: Sat, 3 Oct 2020 10:35:47 -0400
Subject: [PATCH v2 1/2] Fix get_useful_pathkeys_for_relation for volatile
exprs below joins
It's not sufficient to find an ec member whose Vars are all in the rel
we're working with; volatile expressions in particular present a case
where we also need to know if the rel's target contains the expression
or not before we can assume the pathkey for that ec is safe to use at
this point in the query. If the pathkey expr is volatile and we're
below a join, then the expr can't appear in the target until we're above
the join, so we can't sort with it yet.
---
src/backend/optimizer/path/allpaths.c | 11 ++++---
src/backend/optimizer/path/equivclass.c | 43 +++++++++++++++++++++++++
src/backend/optimizer/plan/planner.c | 9 ++++--
src/include/optimizer/paths.h | 1 +
4 files changed, 57 insertions(+), 7 deletions(-)
diff --git a/src/backend/optimizer/path/allpaths.c b/src/backend/optimizer/path/allpaths.c
index b399592ff8..53050b825b 100644
--- a/src/backend/optimizer/path/allpaths.c
+++ b/src/backend/optimizer/path/allpaths.c
@@ -2760,7 +2760,8 @@ get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel)
/*
* Considering query_pathkeys is always worth it, because it might allow
* us to avoid a total sort when we have a partially presorted path
- * available.
+ * available or to push the total sort into the parallel portion of the
+ * query.
*/
if (root->query_pathkeys)
{
@@ -2773,17 +2774,17 @@ get_useful_pathkeys_for_relation(PlannerInfo *root, RelOptInfo *rel)
EquivalenceClass *pathkey_ec = pathkey->pk_eclass;
/*
- * We can only build an Incremental Sort for pathkeys which
- * contain an EC member in the current relation, so ignore any
+ * We can only build a sort for pathkeys which
+ * contain an EC member in the current relation's target, so ignore any
* suffix of the list as soon as we find a pathkey without an EC
- * member the relation.
+ * member in the relation.
*
* By still returning the prefix of the pathkeys list that does
* meet criteria of EC membership in the current relation, we
* enable not just an incremental sort on the entirety of
* query_pathkeys but also incremental sort below a JOIN.
*/
- if (!find_em_expr_for_rel(pathkey_ec, rel))
+ if (!find_em_expr_for_rel_target(pathkey_ec, rel))
break;
npathkeys++;
diff --git a/src/backend/optimizer/path/equivclass.c b/src/backend/optimizer/path/equivclass.c
index b68a5a0ec7..444c805dfb 100644
--- a/src/backend/optimizer/path/equivclass.c
+++ b/src/backend/optimizer/path/equivclass.c
@@ -794,6 +794,49 @@ find_em_expr_for_rel(EquivalenceClass *ec, RelOptInfo *rel)
return NULL;
}
+/*
+ * Find an equivalence class member expression that is in the
+ * the indicated relation's target.
+ */
+Expr *
+find_em_expr_for_rel_target(EquivalenceClass *ec, RelOptInfo *rel)
+{
+ ListCell *lc_em;
+
+ foreach(lc_em, ec->ec_members)
+ {
+ EquivalenceMember *em = lfirst(lc_em);
+ PathTarget *target = rel->reltarget;
+ ListCell *lc_expr;
+
+ /*
+ * First all fo the Vars in the equivalence member have to be taken
+ * entirely from this relation.
+ */
+ if (bms_is_subset(em->em_relids, rel->relids) &&
+ !bms_is_empty(em->em_relids))
+ {
+ /*
+ * Then we must verify that the equivalence member's expr is
+ * generated in the relation's target.
+ */
+ foreach(lc_expr, target->exprs)
+ {
+ /*
+ * If there is more than one equivalence member matching these
+ * requirements we'll be content to choose any one of them.
+ */
+ if (equal(lfirst(lc_expr), em->em_expr))
+ return em->em_expr;
+ }
+
+ }
+ }
+
+ /* We didn't find any suitable equivalence class expression */
+ return NULL;
+}
+
/*
* generate_base_implied_equalities
* Generate any restriction clauses that we can deduce from equivalence
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index f331f82a6c..a48721d2fc 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -7417,7 +7417,9 @@ apply_scanjoin_target_to_paths(PlannerInfo *root,
* current reltarget is. We don't do this in the case where the
* target is parallel-safe, since we will be able to generate superior
* paths by doing it after the final scan/join target has been
- * applied.
+ * applied. Since the target isn't parallel safe, we don't need
+ * to apply projections to the partial paths before building
+ * gather paths.
*/
generate_useful_gather_paths(root, rel, false);
@@ -7570,7 +7572,10 @@ apply_scanjoin_target_to_paths(PlannerInfo *root,
* if the relation is parallel safe, and we don't do it for child rels to
* avoid creating multiple Gather nodes within the same plan. We must do
* this after all paths have been generated and before set_cheapest, since
- * one of the generated paths may turn out to be the cheapest one.
+ * one of the generated paths may turn out to be the cheapest one. We've
+ * already applied any necessary projections to the partial paths above
+ * so any Gather Merge paths will be able to make use of path keys in
+ * requested target..
*/
if (rel->consider_parallel && !IS_OTHER_REL(rel))
generate_useful_gather_paths(root, rel, false);
diff --git a/src/include/optimizer/paths.h b/src/include/optimizer/paths.h
index 10b6e81079..6ddb54f415 100644
--- a/src/include/optimizer/paths.h
+++ b/src/include/optimizer/paths.h
@@ -135,6 +135,7 @@ extern EquivalenceClass *get_eclass_for_sort_expr(PlannerInfo *root,
Relids rel,
bool create_it);
extern Expr *find_em_expr_for_rel(EquivalenceClass *ec, RelOptInfo *rel);
+extern Expr *find_em_expr_for_rel_target(EquivalenceClass *ec, RelOptInfo *rel);
extern void generate_base_implied_equalities(PlannerInfo *root);
extern List *generate_join_implied_equalities(PlannerInfo *root,
Relids join_relids,
--
2.17.1
From 98d98ddab02c36bf7ed2faa614c86f29d487d929 Mon Sep 17 00:00:00 2001
From: James Coleman <jtc...@gmail.com>
Date: Sat, 3 Oct 2020 22:05:08 -0400
Subject: [PATCH v2 2/2] questions
---
src/backend/optimizer/plan/planner.c | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/src/backend/optimizer/plan/planner.c b/src/backend/optimizer/plan/planner.c
index a48721d2fc..45463946b5 100644
--- a/src/backend/optimizer/plan/planner.c
+++ b/src/backend/optimizer/plan/planner.c
@@ -5021,6 +5021,10 @@ create_ordered_paths(PlannerInfo *root,
root->sort_pathkeys,
limit_tuples);
/* Add projection step if needed */
+ /* TODO: why don't we apply the projection to the path
+ * before sorting? Is it because it's already been done
+ * by apply_scanjoin_target_to_paths?
+ */
if (sorted_path->pathtarget != target)
sorted_path = apply_projection_to_path(root, ordered_rel,
sorted_path, target);
@@ -5051,6 +5055,10 @@ create_ordered_paths(PlannerInfo *root,
limit_tuples);
/* Add projection step if needed */
+ /* TODO: why don't we apply the projection to the path
+ * before sorting? Is it because it's already been done
+ * by apply_scanjoin_target_to_paths?
+ */
if (sorted_path->pathtarget != target)
sorted_path = apply_projection_to_path(root, ordered_rel,
sorted_path, target);
@@ -5102,6 +5110,10 @@ create_ordered_paths(PlannerInfo *root,
&total_groups);
/* Add projection step if needed */
+ /* TODO: why can't we apply the projection to the partial
+ * path? Is it because it's already been done if possible
+ * by apply_scanjoin_target_to_paths?
+ */
if (path->pathtarget != target)
path = apply_projection_to_path(root, ordered_rel,
path, target);
@@ -5163,6 +5175,10 @@ create_ordered_paths(PlannerInfo *root,
&total_groups);
/* Add projection step if needed */
+ /* TODO: why can't we apply the projection to the partial
+ * path? Is it because it's already been done if possible
+ * by apply_scanjoin_target_to_paths?
+ */
if (sorted_path->pathtarget != target)
sorted_path = apply_projection_to_path(root, ordered_rel,
sorted_path, target);
--
2.17.1
