opt: add rule to decorrelate union operators

[DrewKimball]

opt: improve EnsureKey custom function

This commit makes a small improvement to the EnsureKey custom function
(used in decorrelation rules), so that it can add passthrough columns to
a Project operator in an effort to find a key. This can prevent rules
from adding unnecessary Ordinality operators to the query plan.

Epic: None

Release note: None

opt: improve exists subquery hoisting

This commit makes a small improvement to the subquery-hoisting rules so
that hoisting an EXISTS subquery can often avoid projecting a new
column to check for NULL values. This can allow other optimization rules
to match later on.

Epic: None

Release note: None

opt: add rule to decorrelate unions in EXISTS subqueries

This commit adds a new rule TryDecorrelateUnion, which matches on a
Union or UnionAll operator in the input of a ScalarGroupBy. The
ScalarGroupBy must have "any-not-null" semantics, meaning it produces
an arbitrary non-null value from each input column.

If these conditions are satisfied, the Union operator is replaced by an
InnerJoin between two ScalarGroupBy operators. A Project coalesces
columns from each side of the join to produce the final aggregated values.

This transformation does not itself decorrelate the Union operators, but
it does make it easier for other rules to do so.

Release note: None

Epic: None

[blathers-crl]

Your pull request contains more than 1000 changes. It is strongly encouraged to split big PRs into smaller chunks.

_{🦉 Hoot! I am a Blathers, a bot for CockroachDB. My owner is dev-inf.}

[cockroach-teamcity]

This change is 

[mgartner]

Very nice!!!

Reviewed 6 of 6 files at r1, 10 of 10 files at r2, 3 of 3 files at r3, all commit messages.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @DrewKimball)

---

pkg/sql/opt/norm/rules/decorrelate.opt line 410 at r3 (raw file):

# TryDecorrelateUnion replaces a Union/UnionAll beneath a ScalarGroupBy with an
# InnerJoin between two ScalarGroupBy operators. A Project operator coalesces

It's really a cross-join isn't it? A diagram of the transformation in this comment, either in an abbreviated opt plan format or SQL, might be helpful in explaining the rule, but up to you.

---

pkg/sql/opt/norm/testdata/rules/decorrelate line 7595 at r3 (raw file):

 │              └── j
 └── projections
      └── CASE WHEN COALESCE("?column?", "?column?") IS NOT NULL THEN 1 ELSE 0 END

nit: I realize it'll blow up the size of these plans, but a more verbose output might be nice, e.g., to see that the correct columns are coalesced here.

[DrewKimball]

TFTR!

Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @DrewKimball)

---

pkg/sql/opt/norm/rules/decorrelate.opt line 410 at r3 (raw file):

Previously, mgartner (Marcus Gartner) wrote…

It's really a cross-join isn't it? A diagram of the transformation in this comment, either in an abbreviated opt plan format or SQL, might be helpful in explaining the rule, but up to you.

Done.

---

pkg/sql/opt/norm/testdata/rules/decorrelate line 7595 at r3 (raw file):

Previously, mgartner (Marcus Gartner) wrote…

nit: I realize it'll blow up the size of these plans, but a more verbose output might be nice, e.g., to see that the correct columns are coalesced here.

Done. Turns out there's a show-columns directive.

[DrewKimball]

bors r+

[craig]

Build succeeded:

• linux_arm64_build
• macos_amd64_build
• macos_arm64_build
• check_generated_code
• unit_tests
• docker_image_amd64
• local_roachtest
• linux_amd64_fips_build
• acceptance
• lint
• local_roachtest_fips
• windows_build
• linux_amd64_build
• examples_orms

[mgartner]

pkg/sql/opt/norm/testdata/rules/decorrelate line 7801 at r6 (raw file):

# No-op because there's an Intersect instead of a Union.
norm expect-not=TryDecorrelateUnion format=(hide-all,show-columns)
SELECT *, CASE WHEN EXISTS (SELECT 1 FROM xy WHERE x = k INTERSECT SELECT 1 FROM uv WHERE v = k) THEN 1 ELSE 0 END FROM a;

Post-merge thought: all these tests have a single column projected in the subquery. Should we have some tests with multiple columns, expressions on those columns, and column references instead of constants? E.g.:

... EXISTS (SELECT 1, 2 ... UNION ...)

... EXISTS (SELECT x ... UNION ...)

... EXISTS (SELECT x + y ... UNION ...)

... EXISTS (SELECT x, y ... UNION ...