sql: recent regression on planning comparisons #95633
Assignees
Labels
A-sql-execution
Relating to SQL execution.
A-sql-optimizer
SQL logical planning and optimizations.
C-bug
Code not up to spec/doc, specs & docs deemed correct. Solution expected to change code/behavior.
regression
Regression from a release.
T-sql-queries
SQL Queries Team
Comments
knz
added
C-bug
15 tasks
|
This is blocking #88061 from merging. |
|
I have also just bisected this to #94797. I'm not sure that PR is causing the issue but it revealed it (probably because pg_get_userbyid is a UDF now) |
|
I'm going to take a looks at this now since Marcus is OOO and it's blocking another project. |
DrewKimball
added a commit
to DrewKimball/cockroach
that referenced
this issue
Feb 23, 2023
During planning for a render node, the physical planning logic checks whether the last created processor rendered its own expressions, adds its render expressions to the processor if not. The check is necessary to handle the case when expressions in a parent render reference those from a child render. However, render nodes also handle projecting input columns: both removing them and reordering them. Correct planning of the processors depends on the columns at each step being in a particular order. Since physical planning didn't check whether the existing processor projected its columns in some way, the parent render could be combined with its child. This could lead to panics in the parent render expressions when the columns weren't projected as expected. This patch fixes the issue by adding the projection check before allowing the render planning code to add to the previous processor. Fixes cockroachdb#95633 Release note (bug fix): Fixed a rare bug existing since before 22.1 that could cause a projected expression to replace column references with the wrong values.
craig bot
pushed a commit
that referenced
this issue
Feb 23, 2023
97554: physicalplan: don't combine render nodes into one processor r=DrewKimball a=DrewKimball During planning for a render node, the physical planning logic checks whether the last created processor rendered its own expressions, adds its render expressions to the processor if not. The check is necessary to handle the case when expressions in a parent render reference those from a child render. However, render nodes also handle projecting input columns: both removing them and reordering them. Correct planning of the processors depends on the columns at each step being in a particular order. Since physical planning didn't check whether the existing processor projected its columns in some way, the parent render could be combined with its child. This could lead to panics in the parent render expressions when the columns weren't projected as expected. This patch fixes the issue by adding the projection check before allowing the render planning code to add to the previous processor. Fixes #95633 Release note (bug fix): Fixed a rare bug existing since before 22.1 that could cause a projected expression to replace column references with the wrong values. Co-authored-by: Drew Kimball <[email protected]>
blathers-crl bot
pushed a commit
that referenced
this issue
Feb 23, 2023
During planning for a render node, the physical planning logic checks whether the last created processor rendered its own expressions, adds its render expressions to the processor if not. The check is necessary to handle the case when expressions in a parent render reference those from a child render. However, render nodes also handle projecting input columns: both removing them and reordering them. Correct planning of the processors depends on the columns at each step being in a particular order. Since physical planning didn't check whether the existing processor projected its columns in some way, the parent render could be combined with its child. This could lead to panics in the parent render expressions when the columns weren't projected as expected. This patch fixes the issue by adding the projection check before allowing the render planning code to add to the previous processor. Fixes #95633 Release note (bug fix): Fixed a rare bug existing since before 22.1 that could cause a projected expression to replace column references with the wrong values.
craig bot
pushed a commit
that referenced
this issue
Feb 24, 2023
88061: clisqlshell: new infrastructure for describe commands r=rafiss,ZhouXing19 a=knz Fixes #95320. Epic: CRDB-23454 The SQL shell (`cockroach sql`, `demo`) now supports the client-side commands `\l`, `\dn`, `\d`, `\di`, `\dm`, `\ds`, `\dt`, `\dv`, `\dC`, `\dT`, `\dd`, `\dg`, `\du`, `\df` and `\dd` in a way similar to `psql`, including the modifier flags `S` and `+`, for convenience for users migrating from PostgreSQL. A notable difference is that when a pattern argument is specified, it should use the SQL "LIKE" syntax (with `%` representing the wildcard character) instead of PostgreSQL's glob-like syntax (with `*` representing wildcards). Issues discovered: - [x] join bug: #88096 - [x] semi-join exec error #91012 - [x] `pg_table_is_visible` should return true when given a valid index OID and the index is valid. #88097 - [x] missing pkey column in pg_index: #88106 - [x] missing stored columns in pg_index: #88107 - [x] pg_statistic_ext has problems #88108 - [x] missing view def on materialized views #88109 - [x] missing schema comments: #88098 - [x] missing pronamespace for functions #94952 - [x] broken pg_function_is_visible for UDFs #94953 - [x] generated columns #92545 - [x] indnullsnotdistinct #92583 - [x] missing prokind #95288 - [x] missing function comments in obj_description #95292 - [x] planning regression #95633 96397: builtins: mark some pg_.* builtins as strict r=DrewKimball a=mgartner Builtins defined using the UDF `Body` field will be wrapped in a `CASE` expression if they are strict, i.e., `CalledOnNullInput=false`. When the builtin is inlined, the `CASE` expression prevents decorrelation, leaving a slow apply-join in the query plan. This caused a significant regression of some ORM introspection queries. Some of these builtins have filters that cause the SQL body to return no rows if any of the arguments is NULL. In this case, the builtin will have the same behavior whether or not it is defined as being strict. We can safely optimize these builtins by setting `CalledOnNullInput=true`. The following conditions are sufficient to prove that `CalledOnNullInput` can be set for a builtin function with a SQL body: 1. The WHERE clause of the SQL query *null-rejects* every argument of the builtin. Operators like `=` and `<` *null-reject* their operands because they filter rows for which an operand is NULL. 2. The arguments are not used elsewhere in the query. This is not strictly necessary, but simplifies the proof because it ensures NULL arguments will not cause the builtin to error. Examples of SQL statements that would allow `CalledOnNullInput` to be set: ``` SELECT * FROM tab WHERE $1=1 AND $2='two'; SELECT * FROM tab WHERE $1 > 0; ``` Fixes #96218 Fixes #95569 Epic: None Release note: None 97585: cli: don't scope TLS client certs to a specific tenant by default r=stevendanna a=knz Epic: CRDB-23559 Fixes: #97584 This commit changes the default for `--tenant-scope` from "only the system tenant" to "cert valid for all tenants". Note that the scoping is generally useful for security, and it is used in CockroachCloud. However, CockroachCloud does not use our CLI code to generate certs and sets its cert tenant scopes on its own. Given that our CLI code is provided for convenience and developer productivity, and we don't expect certs generated here to be used in multi-tenant deployments where tenants are adversarial to each other, defaulting to certs that are valid on every tenant is a good choice. Release note: None Co-authored-by: Raphael 'kena' Poss <[email protected]> Co-authored-by: Marcus Gartner <[email protected]>
DrewKimball
added a commit
that referenced
this issue
Apr 3, 2023
During planning for a render node, the physical planning logic checks whether the last created processor rendered its own expressions, adds its render expressions to the processor if not. The check is necessary to handle the case when expressions in a parent render reference those from a child render. However, render nodes also handle projecting input columns: both removing them and reordering them. Correct planning of the processors depends on the columns at each step being in a particular order. Since physical planning didn't check whether the existing processor projected its columns in some way, the parent render could be combined with its child. This could lead to panics in the parent render expressions when the columns weren't projected as expected. This patch fixes the issue by adding the projection check before allowing the render planning code to add to the previous processor. Fixes #95633 Release note (bug fix): Fixed a rare bug existing since before 22.1 that could cause a projected expression to replace column references with the wrong values.
DrewKimball
added a commit
to DrewKimball/cockroach
that referenced
this issue
Apr 3, 2023
During planning for a render node, the physical planning logic checks whether the last created processor rendered its own expressions, adds its render expressions to the processor if not. The check is necessary to handle the case when expressions in a parent render reference those from a child render. However, render nodes also handle projecting input columns: both removing them and reordering them. Correct planning of the processors depends on the columns at each step being in a particular order. Since physical planning didn't check whether the existing processor projected its columns in some way, the parent render could be combined with its child. This could lead to panics in the parent render expressions when the columns weren't projected as expected. This patch fixes the issue by adding the projection check before allowing the render planning code to add to the previous processor. Fixes cockroachdb#95633 Release note (bug fix): Fixed a rare bug existing since before 22.1 that could cause a projected expression to replace column references with the wrong values.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Some time in the past 5 days, the planning / execution engine started to fail on the following query:
stack trace:
cc @yuzefovich for triage
Jira issue: CRDB-23642
Epic: CRDB-23454
The text was updated successfully, but these errors were encountered: