Postgres: support WHERE index_predicate in the ON CONFLICT target

[zvonimir-dd]

Dialects: PostgreSQL

Closes #2625.

In PostgreSQL a column-list ON CONFLICT target may end with WHERE index_predicate, which lets a partial unique index be inferred as the arbiter:

INSERT INTO t (a, b) VALUES (1, 2) ON CONFLICT (a) WHERE b > 0 DO NOTHING

• ConflictTarget::Columns becomes a struct variant, Columns { columns: Vec<Ident>, predicate: Option<Expr> }. The predicate is parsed after the column list and before DO, and printed back after it.
• Accepted only after a column list, as in PostgreSQL's grammar. ON CONSTRAINT … WHERE and a WHERE with no target are still rejected.
• Separate from the action's DO UPDATE SET … WHERE condition (Postgres: [pg on conflict method] does not support 'where' condition. #733, update on conflict with condition #735). A statement can carry both; the tests check that.
• The predicate is a regular Expr field, so visitors see the columns it names, and the target's span covers it.
• The syntax is added to the fuzz seed corpus.

This changes the shape of ConflictTarget::Columns (it was Columns(Vec<Ident>)), so code matching on it needs updating. The struct variant also leaves room for the ( index_expression ) form requested in #1977.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 94.11765% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 81.25%. Comparing base (dacdbc9) to head (8936117).
⚠️ Report is 6 commits behind head on main.

Files with missing lines	Patch %	Lines
src/ast/mod.rs	66.66%	0 Missing and 2 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2627      +/-   ##
==========================================
+ Coverage   81.16%   81.25%   +0.09%     
==========================================
  Files          42       42              
  Lines       33769    33902     +133     
  Branches    33769    33902     +133     
==========================================
+ Hits        27408    27547     +139     
+ Misses       2797     2780      -17     
- Partials     3564     3575      +11     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[LucaCappelletti94]

Thanks for the PR. The parsing and the tests look right. I'd go with the struct variant you offered in #2625, ConflictTarget::Columns { columns: Vec<Ident>, predicate: Option<Expr> }, instead of the separate OnConflict::conflict_target_predicate field.

Also do add an example of this syntax for the fuzzer corpus.

[zvonimir-dd]

Thanks! Done in ff41dab: ConflictTarget::Columns is now Columns { columns, predicate }, and OnConflict::conflict_target_predicate is gone. I added fuzz/fuzz_seeds/postgres_on_conflict_target_predicate.sql, and checked that both of its statements run on PostgreSQL 17. The description is updated to match.

[LucaCappelletti94]

Other than this nit, the code looks okay. If you have time, I would appreciate if you could also do some reviews of the other PRs so as to speed up the whole queue.

[LucaCappelletti94]

You should say SQLite accepts this predicate too. Its upsert grammar has the same ( indexed-column, ... ) WHERE expr target, and the parser reads it in every dialect. SQLite 3.51.1 runs INSERT INTO t VALUES (1, 2, 3) ON CONFLICT (a) WHERE b > 0 DO NOTHING and rejects ON CONFLICT WHERE b > 0. With the doc naming only PostgreSQL, a later change could gate the parse to PostgreSQL and drop SQLite without anyone noticing.

Suggested change

	///
	/// See <https://www.postgresql.org/docs/current/sql-insert.html>
	/// Target specified as a list of columns, optionally followed by
	/// `WHERE index_predicate`, which lets a partial unique index be inferred as the arbiter:
	/// `ON CONFLICT (a) WHERE b > 0 DO ...`. PostgreSQL and SQLite both accept the predicate.
	///
	/// See <https://www.postgresql.org/docs/current/sql-insert.html>
	/// and <https://www.sqlite.org/lang_upsert.html>

[zvonimir-dd]

Thanks, applied your suggestion as written in 8936117. I also added parse_on_conflict_target_predicate to tests/sqlparser_sqlite.rs, so restricting the parse to PostgreSQL would fail a test rather than only contradict the doc.

I checked it against SQLite 3.37.2 as well. With a partial unique index on t (a) WHERE b > 0, INSERT INTO t VALUES (1, 2, 3) ON CONFLICT (a) WHERE b > 0 DO NOTHING runs, and the second insert is a no-op. ON CONFLICT WHERE b > 0 is a syntax error.