opt: add FD equivalency for semi-passthrough projection columns #43465
Conversation
|
This change is |
There was a problem hiding this comment.
Reviewable status:
complete! 1 of 0 LGTMs obtained (waiting on @andy-kimball, @justinj, @RaduBerinde, and @rytaft)
pkg/sql/opt/memo/logical_props_builder.go, line 272 at r1 (raw file):
} } // Also add any column that is a direct reference to a non-null column. The
I'd add a test for this new case into memo/testdata/logprops/project.
In certain cases, we have synthesized projections that are simple variable expressions (I hereby name these semi-passthrough). This can happen for example if a non-trivial expression gets constant folded to a single variable (perhaps after placeholder replacement). The optimizer can also construct these to assign different ColumnIDs. This change adds the equivalency between the semi-passthrough column and the one it refers to, and also uses it to mark columns as non-null. This improves the plan when we are ordering on such a semi-passthrough column (as long as both columns are projected). There will be a follow-up fix that makes the `CanProvideOrdering` logic better to handle this case even when the original column is not projected. Fixes cockroachdb#43360.
There was a problem hiding this comment.
TFTR!
Reviewable status:
pkg/sql/opt/memo/logical_props_builder.go, line 272 at r1 (raw file):
Previously, andy-kimball (Andy Kimball) wrote…
I'd add a test for this new case into
memo/testdata/logprops/project.
Done. Added a test for the null column and one for the FD.
|
bors r+ |
|
Hm, I don't think I pushed that last update bors r- |
Canceled |
RaduBerinde
force-pushed
the
proj-eq-fd
branch
from
03aa4e1
to
6377847
Compare
|
bors r+ |
43465: opt: add FD equivalency for semi-passthrough projection columns r=RaduBerinde a=RaduBerinde In certain cases, we have synthesized projections that are simple variable expressions (I hereby name these semi-passthrough). This can happen for example if a non-trivial expression gets constant folded to a single variable (perhaps after placeholder replacement). The optimizer can also construct these to assign different ColumnIDs. This change adds the equivalency between the semi-passthrough column and the one it refers to, and also uses it to mark columns as non-null. This improves the plan when we are ordering on such a semi-passthrough column (as long as both columns are projected). There will be a follow-up fix that makes the `CanProvideOrdering` logic better to handle this case even when the original column is not projected. Fixes #43360. Co-authored-by: Radu Berinde <radu@cockroachlabs.com>
Build succeeded |
|
Release note (performance improvement): the optimizer generates better execution plans in some cases where there is an ORDER BY expression that simplifies to a simple variable reference. |
In certain cases, we have synthesized projections that are simple
variable expressions (I hereby name these semi-passthrough). This can
happen for example if a non-trivial expression gets constant folded to
a single variable (perhaps after placeholder replacement). The
optimizer can also construct these to assign different ColumnIDs.
This change adds the equivalency between the semi-passthrough column
and the one it refers to, and also uses it to mark columns as
non-null. This improves the plan when we are ordering on such a
semi-passthrough column (as long as both columns are
projected). There will be a follow-up fix that makes the
CanProvideOrderinglogic better to handle this case even when theoriginal column is not projected.
Fixes #43360.