sql: fix JSON fetch value operator evaluation #55316

mgartner · 2020-10-08T02:31:00Z

sql: fix JSON fetch value operator evaluation

This commit fixes incorrect evaluation of the JSON fetch value operator,
->. There were two bugs causing incorrect evaluation.

The first issue was a result of optimizer normalization rules that
produced unequivalent scalar expressions when converting -> to @>.
The rules would convert expressions like j -> '"a"' = 1 to
j @> '{"a": 1}'. This is invalid because -> results in NULL when
the LHS does not contain the RHS key, but the resulting @> expression
is always either true or false, never NULL. These normalization
rules have been removed.

These two rules existed to provide inverted index-acceleration for
queries with ->, because the optimizer can only index-accerlate @>
operators during exploration. As a result of their removal, queries with
-> operators are no longer index-accelerated. This will be remedied in
a future commit.

The second issue was a bug in the vectorized overload of the ->
operator. Previously, when the operator evaluated to NULL with two
non-NULL inputs, the resulting NULL would not be tracked by the
Nulls struct.

Fixes #49143

Release note (bug fix): Previously, the JSON fetch value operator, ->,
would evaluate incorrectly in some cases. This has been fixed.

opt: index-accelerate equalities with JSON fetch expressions

This commit allows inverted indexes to be scanned to satisfy query
filters in the form: j->'a' = '1'. The optimizer had previously
supported this by normalizing these expressions into expressions with
JSON containment operators, @>, but it lost this support when these
normalization rules were found to produce inequivalent expressions.

Note that query filters of several forms, which were previously
accelerated via the incorrect normalization rules, are not accelerated
as part of this commit, such as:

j->'a' @> '{"x": "y"}
j->'a'->'b' = '"c"'

Release note: None

cockroach-teamcity · 2020-10-08T02:31:09Z

This change is

mgartner · 2020-10-08T02:39:29Z

pkg/sql/opt/exec/execbuilder/testdata/inverted_index

 │          table                d@primary                    ·       ·
 │          key columns          a                            ·       ·
 └── scan   ·                    ·                            (a)     ·
 ·           estimated row count  110 (missing stats)          ·       ·
 ·           table                d@foo_inv                    ·       ·
 ·           spans                /"a"/"b"-/"a"/"b"/PrefixEnd  ·       ·

+# TODO(mgartner): Add support for building inverted index constraints for chained JSON
+# fetch operators.


I created a ticket to track here: #55317

I also created a ticket to track the j->'a' @> '{"x": "y"} case: #55318

mgartner · 2020-10-08T02:47:35Z

pkg/sql/opt/exec/execbuilder/testdata/inverted_index

@@ -273,33 +273,37 @@ filter     ·                    ·                     (a, b)  ·
 ·          spans                FULL SCAN             ·       ·


+# TODO(mgartner): It should not be required to force the index scan. It is
+# required until the statistics builder treats b->'a' = '"b"' similarly to the
+# containment operator, @>.


I created a ticket to track here: #55319

mgartner · 2020-10-08T02:49:51Z

@yuzefovich please take a look at the vectorized engine changes in the first commit, and let me know if this is the right approach.

rytaft

opt stuff

Reviewed 12 of 12 files at r1, 5 of 5 files at r2.
Reviewable status: complete! 1 of 0 LGTMs obtained (waiting on @RaduBerinde and @yuzefovich)

yuzefovich

The approach looks good to me, thanks. I just have a couple of nits.

Reviewable status: complete! 2 of 0 LGTMs obtained (waiting on @mgartner and @RaduBerinde)

pkg/sql/colexec/proj_const_ops_tmpl.go, line 101 at r1 (raw file):

	}
	projCol := projVec._RET_TYP()
	_outNulls := projVec.Nulls()

nit: this deserves a comment to spell out the "contract" (similar to what we have about _overloadHelper above).

pkg/sql/colexec/proj_const_ops_tmpl.go, line 140 at r1 (raw file):

		}
	}
	// {{if _HAS_NULLS}}

nit: initially I thought that there might be a bug that we're not updating nulls vector when the argument vector didn't have nulls (i.e. _HAS_NULLS == false), but in that case _outNulls has already been updated by _ASSIGN function if need be. I think it's worth adding a comment to mention this observation.

pkg/sql/colexec/proj_const_ops_tmpl.go, line 141 at r1 (raw file):

	}
	// {{if _HAS_NULLS}}
	projVec.SetNulls(_outNulls.Or(colNulls))

~~I think here we want to be more efficient in non-datum cases - coldata.Nulls.Copy() is faster than coldata.Nulls.Or(). I'd do something like this:~~

// {{if eq .VecMethod "Datum"}}
projVec.SetNulls(_outNulls.Or(colNulls))
// {{else}}
colNullsCopy := colNulls.Copy()
projVec.SetNulls(&colNullsCopy)
// {{end}}

~~This will also make _outNulls variable unused in many code paths, so you'll need to go around the unused warning.~~

Update: I tried it out, and the microbenchmarks don't show much difference, so the current approach looks good. (If you're curious, the diff is here.)

pkg/sql/colexec/proj_non_const_ops_tmpl.go, line 112 at r1 (raw file):

	col1 := vec1._L_TYP()
	col2 := vec2._R_TYP()
	_outNulls := projVec.Nulls()

nit: ditto for comments

RaduBerinde

Nice work!

Reviewable status: complete! 3 of 0 LGTMs obtained (waiting on @mgartner)

pkg/sql/opt/idxconstraint/index_constraints.go, line 996 at r2 (raw file):

	}

	rightConst := rhs.(*memo.ConstExpr)

, ok missing. Also, I think we have an extractconstval somewhere since other operators can be constant (e.g. True, False, Null)

mgartner

Reviewable status: complete! 0 of 0 LGTMs obtained (and 3 stale) (waiting on @RaduBerinde, @rytaft, and @yuzefovich)

pkg/sql/colexec/proj_const_ops_tmpl.go, line 101 at r1 (raw file):

Previously, yuzefovich wrote…

nit: this deserves a comment to spell out the "contract" (similar to what we have about _overloadHelper above).

Done.

pkg/sql/colexec/proj_const_ops_tmpl.go, line 140 at r1 (raw file):

Previously, yuzefovich wrote…

nit: initially I thought that there might be a bug that we're not updating nulls vector when the argument vector didn't have nulls (i.e. _HAS_NULLS == false), but in that case _outNulls has already been updated by _ASSIGN function if need be. I think it's worth adding a comment to mention this observation.

Good idea, done.

pkg/sql/colexec/proj_const_ops_tmpl.go, line 141 at r1 (raw file):

Previously, yuzefovich wrote…

~~I think here we want to be more efficient in non-datum cases - coldata.Nulls.Copy() is faster than coldata.Nulls.Or(). I'd do something like this:~~
// {{if eq .VecMethod "Datum"}}
projVec.SetNulls(_outNulls.Or(colNulls))
// {{else}}
colNullsCopy := colNulls.Copy()
projVec.SetNulls(&colNullsCopy)
// {{end}}
~~This will also make _outNulls variable unused in many code paths, so you'll need to go around the unused warning.~~

Update: I tried it out, and the microbenchmarks don't show much difference, so the current approach looks good. (If you're curious, the diff is here.)

I tried the same approach and also ran into _outNulls being unused, and gave up at that point.

But I did look at the implementation of Or, and if one of the Nulls has no nulls, then it performs a copy. So it makes sense that your microbenchmarks show little difference.

pkg/sql/colexec/proj_non_const_ops_tmpl.go, line 112 at r1 (raw file):

Previously, yuzefovich wrote…

nit: ditto for comments

Done.

pkg/sql/opt/idxconstraint/index_constraints.go, line 996 at r2 (raw file):

Previously, RaduBerinde wrote…

, ok missing. Also, I think we have an extractconstval somewhere since other operators can be constant (e.g. True, False, Null)

Great catch on the missing , ok! Fixed.

We only care about the case when the RHS is a constant DJSON, so neglecting True, False, and Null is not an issue. But I've updated to use memo.CanExtractConstDatum and memo.ExtractConstDatum anyway since it seems more robust.

yuzefovich

Thanks! You need to run make execgen on the first commit to regenerate the colexec code with comments.

Reviewed 3 of 7 files at r3.
Reviewable status: complete! 0 of 0 LGTMs obtained (and 3 stale) (waiting on @RaduBerinde, @rytaft, and @yuzefovich)

mgartner

You need to run make execgen on the first commit to regenerate the colexec code with comments.

Oops! Done.

Reviewable status: complete! 0 of 0 LGTMs obtained (and 3 stale) (waiting on @RaduBerinde, @rytaft, and @yuzefovich)

This commit fixes incorrect evaluation of the JSON fetch value operator, `->`. There were two bugs causing incorrect evaluation. The first issue was a result of optimizer normalization rules that produced unequivalent scalar expressions when converting `->` to `@>`. The rules would convert expressions like `j -> '"a"' = 1` to `j @> '{"a": 1}'`. This is invalid because `->` results in `NULL` when the LHS does not contain the RHS key, but the resulting `@>` expression is always either `true` or `false`, never `NULL`. These normalization rules have been removed. These two rules existed to provide inverted index-acceleration for queries with `->`, because the optimizer can only index-accerlate `@>` operators during exploration. As a result of their removal, queries with `->` operators are no longer index-accelerated. This will be remedied in a future commit. The second issue was a bug in the vectorized overload of the `->` operator. Previously, when the operator evaluated to `NULL` with two non-`NULL` inputs, the resulting `NULL` would not be tracked by the `Nulls` struct. Fixes cockroachdb#49143 Release note (bug fix): Previously, the JSON fetch value operator, `->`, would evaluate incorrectly in some cases. This has been fixed.

This commit allows inverted indexes to be scanned to satisfy query filters in the form: `j->'a' = '1'`. The optimizer had previously supported this by normalizing these expressions into expressions with JSON containment operators, `@>`, but it lost this support when these normalization rules were found to produce inequivalent expressions. Note that query filters of several forms, which were previously accelerated via the incorrect normalization rules, are not accelerated as part of this commit, such as: j->'a' @> '{"x": "y"} j->'a'->'b' = '"c"' Release note: None

mgartner · 2020-10-09T17:30:06Z

TFTRs!

bors r+

craig · 2020-10-09T18:35:33Z

Build succeeded:

GitHub CI (Cockroach)

rafiss · 2020-10-10T01:14:46Z

@mgartner are you planning to backport this? (just wondering since it affects JSON compatibility with Django 3.1)

RaduBerinde · 2020-10-10T01:23:17Z

I think we should backport for 20.2.1.

mgartner · 2020-10-12T16:19:45Z

Created a backport PR: #55447

Prior to cockroachdb#55316, the optimizer generated inverted index scans on indexed JSON columns when queries had filters with chained fetch value operators, for example `j->'a'->'b' = '1'`. The logic that made this possible was found to create query plans not equivalent to the query, so it was removed. This commit restores the ability to index accelerate chained -> operators. Fixes cockroachdb#55317 Release note (performance improvement): A bug fix included in 20.2.1 for for the JSON fetch value operator, `->`, resulted in chained `->` operators in query filters not being index accelerated, e.g., `j->'a'->'b' = '1'`. Chained `->` are now index accelerated.

59494: opt: index accelerate chained fetch value operators r=rytaft a=mgartner #### opt: add test for JSON fetch val inverse Release note: None #### opt: index accelerate chained fetch value operators Prior to #55316, the optimizer generated inverted index scans on indexed JSON columns when queries had filters with chained fetch value operators, for example `j->'a'->'b' = '1'`. The logic that made this possible was found to create query plans not equivalent to the query, so it was removed. This commit restores the ability to index accelerate chained -> operators. Fixes #55317 Release note (performance improvement): A bug fix included in 20.2.1 for for the JSON fetch value operator, `->`, resulted in chained `->` operators in query filters not being index accelerated, e.g., `j->'a'->'b' = '1'`. Chained `->` are now index accelerated. Co-authored-by: Marcus Gartner <[email protected]>

This commit removes an invalid normalization from the NormalizeVisitor. It was previously discovered that transforming expressions in the form `j->'a' = '1'` to `j @> '{"a": 1}'` is invalid (see cockroachdb#49143). This transformation rule was removed from the optimizer in cockroachdb#55316. But the same transformation was not removed from the NormalizeVisitor. This visitor is only used to normalize scalar expressions in table descriptors (`DEFAULT` expressions, computed column expressions, and partial index predicates) during a backfill. Fixes cockroachdb#75097 Release note (bug fix): A bug has been fixed that caused incorrect values to be written to computed columns when their expressions were of the form `j->x = y`, where `j` is a `JSON` column and `x` and `y` are constants. This bug also caused corruption of partial indexes with `WHERE` clauses containing expressions of the same form. This bug was present since version 2.0.0.

75908: scripts: per-branch bump-pebble.sh script r=jbowens a=nicktrav Currently, the master branch, in addition to each release branch relies on the same `bump-pebbble.sh` from the master branch. There are subtle differences between master and the release branches (i.e. build system) that ends up breaking the script as the changes are introduced on the master branch but not backported to the release branches. One solution to this problem is to continue to maintain one script on the master branch, but include switching logic for each release branch to account for the differences. An alternative approach is to have a script per release branch. Rather than having switching logic, the script has the appropriate logic for that branch. When a new release branch is cut, the script inherits the most up-to-date logic from master, and all that needs to change is the name of the branch and the corresponding Pebble branch. Pin the `bump-pebble.sh` script to the master branch. The script will error out if it is run from a different branch. Release note: None 75914: tree: remove invalid normalization r=mgartner a=mgartner This commit removes an invalid normalization from the NormalizeVisitor. It was previously discovered that transforming expressions in the form `j->'a' = '1'` to `j @> '{"a": 1}'` is invalid (see #49143). This transformation rule was removed from the optimizer in #55316. But the same transformation was not removed from the NormalizeVisitor. This visitor is only used to normalize scalar expressions in table descriptors (`DEFAULT` expressions, computed column expressions, and partial index predicates) during a backfill. Fixes #75907 Release note (bug fix): A bug has been fixed that caused incorrect values to be written to computed columns when their expressions were of the form `j->x = y`, where `j` is a `JSON` column and `x` and `y` are constants. This bug also caused corruption of partial indexes with `WHERE` clauses containing expressions of the same form. This bug was present since version 2.0.0. 75963: batcheval: use same stats timestamp for `AddSSTable` assertions r=rhu713 a=erikgrinaker Resolves #75643. Resolves #75642. Release note: None Co-authored-by: Nick Travers <[email protected]> Co-authored-by: Marcus Gartner <[email protected]> Co-authored-by: Erik Grinaker <[email protected]>

mgartner requested a review from a team as a code owner October 8, 2020 02:31

mgartner force-pushed the containment-norm-fix branch 2 times, most recently from 6bcf389 to a5c02fe Compare October 8, 2020 02:34

mgartner mentioned this pull request Oct 8, 2020

opt: generate inverted index scans for chained JSON fetch value operators with equality expression #55317

Closed

mgartner commented Oct 8, 2020

View reviewed changes

This was referenced Oct 8, 2020

opt: generate inverted index scans for JSON fetch value and containment operators #55318

Closed

opt: calculate statistics for the JSON fetch val and equality operator #55319

Closed

mgartner commented Oct 8, 2020

View reviewed changes

mgartner requested review from yuzefovich, RaduBerinde and rytaft October 8, 2020 02:48

rytaft approved these changes Oct 8, 2020

View reviewed changes

yuzefovich approved these changes Oct 8, 2020

View reviewed changes

RaduBerinde approved these changes Oct 8, 2020

View reviewed changes

mgartner force-pushed the containment-norm-fix branch from a5c02fe to 31a5d89 Compare October 8, 2020 20:36

mgartner commented Oct 8, 2020

View reviewed changes

yuzefovich reviewed Oct 9, 2020

View reviewed changes

mgartner force-pushed the containment-norm-fix branch from 31a5d89 to 104e5e2 Compare October 9, 2020 15:11

mgartner commented Oct 9, 2020

View reviewed changes

mgartner added 2 commits October 9, 2020 09:52

mgartner force-pushed the containment-norm-fix branch from 104e5e2 to 30f8072 Compare October 9, 2020 16:52

craig bot merged commit 9bfd067 into cockroachdb:master Oct 9, 2020

mgartner deleted the containment-norm-fix branch October 9, 2020 19:02

mgartner mentioned this pull request Jan 27, 2021

opt: index accelerate chained fetch value operators #59494

Merged

mgartner mentioned this pull request Feb 8, 2021

opt: inverted-index accelerate filters of the form j->'a' = '{"b": "c"} #59605

Closed

mgartner mentioned this pull request Feb 2, 2022

tree: backfill incorrectly normalized j->'a' = '1' to j @> '{"a": 1}' #75907

Closed

mgartner mentioned this pull request Feb 3, 2022

tree: remove invalid normalization #75914

Merged

DrewKimball mentioned this pull request Sep 22, 2022

colexec: make sure nulls are set for datum columns #88418

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

sql: fix JSON fetch value operator evaluation #55316

sql: fix JSON fetch value operator evaluation #55316

mgartner commented Oct 8, 2020 •

edited

Loading

cockroach-teamcity commented Oct 8, 2020

mgartner Oct 8, 2020

mgartner Oct 8, 2020

mgartner Oct 8, 2020

mgartner commented Oct 8, 2020

rytaft left a comment

yuzefovich left a comment

RaduBerinde left a comment

mgartner left a comment

yuzefovich left a comment

mgartner left a comment

mgartner commented Oct 9, 2020

craig bot commented Oct 9, 2020

rafiss commented Oct 10, 2020

RaduBerinde commented Oct 10, 2020

mgartner commented Oct 12, 2020

sql: fix JSON fetch value operator evaluation #55316

sql: fix JSON fetch value operator evaluation #55316

Conversation

mgartner commented Oct 8, 2020 • edited Loading

sql: fix JSON fetch value operator evaluation

opt: index-accelerate equalities with JSON fetch expressions

cockroach-teamcity commented Oct 8, 2020

mgartner Oct 8, 2020

Choose a reason for hiding this comment

mgartner Oct 8, 2020

Choose a reason for hiding this comment

mgartner Oct 8, 2020

Choose a reason for hiding this comment

mgartner commented Oct 8, 2020

rytaft left a comment

Choose a reason for hiding this comment

yuzefovich left a comment

Choose a reason for hiding this comment

RaduBerinde left a comment

Choose a reason for hiding this comment

mgartner left a comment

Choose a reason for hiding this comment

yuzefovich left a comment

Choose a reason for hiding this comment

mgartner left a comment

Choose a reason for hiding this comment

mgartner commented Oct 9, 2020

craig bot commented Oct 9, 2020

rafiss commented Oct 10, 2020

RaduBerinde commented Oct 10, 2020

mgartner commented Oct 12, 2020

mgartner commented Oct 8, 2020 •

edited

Loading