release-22.1: opt: don't add reordered join with extra filters to original memo group #91654
Conversation
|
Thanks for opening a backport. Please check the backport criteria before merging:
If some of the basic criteria cannot be satisfied, ensure that the exceptional criteria are satisfied within.
Add a brief release justification to the body of your PR to justify this backport. Some other things to consider:
|
|
This change is |
There was a problem hiding this comment.
Reviewed 2 of 2 files at r1, all commit messages.
Reviewable status:complete! 1 of 0 LGTMs obtained (waiting on @DrewKimball)
|
unconstrained_non_covering_index_scan_enabled doesn't seem to exist on 22.1 |
There was a problem hiding this comment.
Reviewable status:
pkg/sql/opt/xform/testdata/rules/join_order line 2639 at r1 (raw file):
Previously, rytaft (Rebecca Taft) wrote…
unconstrained_non_covering_index_scan_enabled doesn't seem to exist on 22.1
Hm, neither do the other settings, and I'm struggling to reproduce without them. Do you think it would be ok to backport the fix without the test?
|
Previously, DrewKimball (Drew Kimball) wrote…
Is there any way to add another test that shows that your change works? You might be able to get the same effect for this test by adding hints to force specific indexes / join algorithms |
|
Previously, rytaft (Rebecca Taft) wrote…
Also, not sure how difficult it is to backport these settings, but it seems like it would be useful to have these settings available on 22.1... |
There was a problem hiding this comment.
Reviewable status:
pkg/sql/opt/xform/testdata/rules/join_order line 2639 at r1 (raw file):
Previously, rytaft (Rebecca Taft) wrote…
Also, not sure how difficult it is to backport these settings, but it seems like it would be useful to have these settings available on 22.1...
That's a good point, I think we've run into this before. I'll look into backporting the optimizer testing settings.
DrewKimball
force-pushed
the
backport22.1-88779
branch
2 times, most recently
from
2a1cad8 to
e4a9735
Compare
The `JoinOrderBuilder` builds reordered join plans from the bottom up. It expects filters to be pushed down as far as possible at each step, and that transitive closure has been calculated over Inner Join equality filters (e.g. `a=b` and `b=c` => `a=c`). It also reuses the original matched joins when possible to avoid duplicate work by adding to the original memo groups. This could previously cause filters to be dropped in the case when the original join tree did not compute transitive closure and push filters down as far as possible. More specifically, the `JoinOrderBuilder` could add new reordered joins with new filters synthesized and pushed down as far as possible to an original memo group that didn't have one of those filters. Subsequent joins would then expect the filter to be part of the memo group, and so it wouldn't be added later on in the plan. In the rare case when the expression without the filter was chosen, this could manifest as a dropped filter in the final plan. This was rare because dropping a filter usually does not produce a lower-cost plan. As an example, take this original join tree: ``` (xy join ab on true) join uv on x = u and a = u; ``` Here it is possible to sythesize and push down a `x = a` filter, and so the `JoinOrderBuilder` would do this and add it to the group: ``` group (xy join ab on true), (xy join ab on x = a) ``` Later joins would use this group as an input, an expect the `x = a` filter to be present. If costing happened to choose the first expression in the group, we would end up choosing a plan like this: ``` (xy join ab on true) join uv on x = u ``` Where the `a = u` filter isn't included in the top-level join because it would be redundant to add it when `x = u` and `x = a` are already present. This is a bit of a simplification, but is essentially the problem fixed by this commit. This commit adds a check to the `JoinOrderBuilder` to identify cases where filters (including ones sythesized from the transitive closure) weren't pushed all the way down in the original join tree. When this is true, none of the originally matched joins can be reused when reordered joins are built except for the root join. This solution may perform some duplicate work when filters aren't pushed down, but it shouldn't matter because this case is rare (and should be avoided whenever possible). Fixes cockroachdb#88659 Release note (bug fix): Fixed a bug introduced in 20.2 that could cause filters to be dropped from a query plan with many joins in rare cases.
DrewKimball
force-pushed
the
backport22.1-88779
branch
from
e4a9735 to
60718c0
Compare
|
TFTR |
Backport 1/1 commits from #88779.
/cc @cockroachdb/release
The
JoinOrderBuilderbuilds reordered join plans from the bottom up.It expects filters to be pushed down as far as possible at each step, and
that transitive closure has been calculated over Inner Join equality filters
(e.g.
a=bandb=c=>a=c). It also reuses the original matched joinswhen possible to avoid duplicate work by adding to the original memo groups.
This could previously cause filters to be dropped in the case when the
original join tree did not compute transitive closure and push filters down
as far as possible. More specifically, the
JoinOrderBuildercould add newreordered joins with new filters synthesized and pushed down as far as possible
to an original memo group that didn't have one of those filters. Subsequent
joins would then expect the filter to be part of the memo group, and so it
wouldn't be added later on in the plan. In the rare case when the expression
without the filter was chosen, this could manifest as a dropped filter in the
final plan. This was rare because dropping a filter usually does not produce
a lower-cost plan.
As an example, take this original join tree:
Here it is possible to sythesize and push down a
x = afilter, and so theJoinOrderBuilderwould do this and add it to the group:Later joins would use this group as an input, an expect the
x = afilter tobe present. If costing happened to choose the first expression in the group,
we would end up choosing a plan like this:
Where the
a = ufilter isn't included in the top-level join because itwould be redundant to add it when
x = uandx = aare already present.This is a bit of a simplification, but is essentially the problem fixed
by this commit.
This commit adds a check to the
JoinOrderBuilderto identify cases wherefilters (including ones sythesized from the transitive closure) weren't
pushed all the way down in the original join tree. When this is true, none
of the originally matched joins can be reused when reordered joins are
built except for the root join. This solution may perform some duplicate
work when filters aren't pushed down, but it shouldn't matter because this
case is rare (and should be avoided whenever possible).
Fixes #88659
Release note (bug fix): Fixed a bug introduced in 20.2 that could cause
filters to be dropped from a query plan with many joins in rare cases.
Release justification: low-risk fix for rare optimizer correctness bug