[ajitpratap0]

Summary

• PostgreSQL / Redshift / SQLite VACUUM and ANALYZE, and PostgreSQL LISTEN / UNLISTEN / NOTIFY, now use a small per-dialect grammar (new pkg/sql/parser/utility_maintenance.go) following the existing GRANT / COPY pattern (parseGrammarStatement + utilityTailError). Trailing same-line tokens such as VACUUM t junk or LISTEN c extra are now an E2002 "expected semicolon or end of statement" error. Before this change they were folded into the verbatim text.
• Generic mode accepts the union of the PostgreSQL, Redshift and SQLite forms and picks the longest complete match without mixing clauses. MySQL / MariaDB / Oracle ANALYZE TABLE ... and MariaDB ANALYZE <statement> keep their existing paths.
• Adds regression tests for the boundary items that origin/main had already fixed: CREATE TRIGGER / PREPARE folding, a same-line second statement after GRANT / COPY, SplitStatements with MySQL/MariaDB backslash escapes, and PostgreSQL CREATE RULE ... DO [ALSO|INSTEAD] (a; b).

Test Plan

• task check (fmt, vet, lint, test:race): passed
• task test:cbinding: passed
• New tests TestParser_MaintenanceBoundaries, TestParser_MaintenanceForms, TestParser_MaintenanceText, TestParser_MaintenanceIncomplete, TestParser_UtilitySelectBoundaryRegressions, TestSplitStatements_MySQLBackslashBoundaries, TestSplitStatements_PostgreSQLRuleCommandList
• Parser benchmarks BenchmarkParserStatementTypes and BenchmarkParserMixedWorkload, -benchmem -count=10, before and after runs interleaved: B/op and allocs/op are identical and ns/op shows no regression (for example SELECT_Complex 6089 to 5610 ns/op, MixedWorkload_Sequential 933 to 930 ns/op)

Breaking Changes

Input that origin/main accepted is now rejected when it puts invalid trailing tokens after these maintenance statements. One example is VACUUM t junk in PostgreSQL, SQLite or generic mode. Valid forms still parse, including every such statement in the testdata/realworld corpus.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DnbRnTFErhoMJdAjVEeCg2

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
gosqlx-website	Ready	Preview	Oct 28, 2026 5:31am UTC

[github-actions]

Claude finished @ajitpratap0's task in 39s —— View job

---

Review of PR #878: end VACUUM, ANALYZE and LISTEN at their grammar

I read the diff and the supporting code by hand. I could not run the test suite here (go test needs approval in this environment), so none of this has been executed.

• Read the diff
• Check the dialect gating
• Check tests and changelog

Potential bug: VACUUM under other dialects (needs verification)

maintenanceEnd in pkg/sql/parser/utility_maintenance.go:22-46 skips every form whose dialect isn't in the mask. For MySQL, MariaDB, Oracle, SQL Server, Snowflake, ClickHouse and BigQuery, no form runs. That leaves best == -1 and firstErr == nil, so it returns (0, nil).

The new case "VACUUM": in utility_stmt.go calls parseGrammarStatement(word, p.vacuumEnd) with no dialect check. The ANALYZE case does have one. parseGrammarStatement (utility_grammar.go:55-68) then treats 0 as a valid end index:

• utilityTailError(0) runs against the wrong token.
• renderTokenRange(start, 0) gets an inverted range.
• p.seek(0) rewinds the parser.

I expect that to produce garbled text, a spurious error, or a loop. On main, VACUUM in these dialects fell through to verbatim handling, so this would be a regression. The same applies to LISTEN, UNLISTEN and NOTIFY: listenEnd doesn't gate on dialect, so those become strict everywhere, which may be fine, but the PR text says PostgreSQL only.

Suggested fix, either of these:

• In maintenanceEnd, fall back to the old verbatim path when best < 0 && firstErr == nil. Or make the VACUUM case mirror the ANALYZE gate and call parseVerbatimStatement for dialects that have no grammar.
• Add a test for VACUUM t under mysql, snowflake, clickhouse and bigquery. The new tests only cover PostgreSQL, Redshift, SQLite and generic.

Other observations

• ANALYZE gate (utility_stmt.go, mask == dialectAll && next != "TABLE" ...): in generic mode ANALYZE LOCAL TABLE and ANALYZE NO_WRITE_TO_BINLOG TABLE keep the old path, but ANALYZE LOCAL t goes to the grammar and fails with "table name". That seems acceptable. A short comment on why those three words are excluded would help.
• postgresVacuum legacy options: a repeated option such as VACUUM FULL FULL t stops the option loop, and FULL is then read as a table name. That is permissive but harmless.
• sqliteAnalyze caps names at [schema.]object by truncating nameEnd. The leftover .extra then triggers the trailing-token error, which the test covers. Good.
• maintenanceNameAt excludes implicitStartWords. A table literally named after a statement starter, such as VACUUM analyze, would be mis-split. That is an edge case, but a quoted identifier is probably the only workaround, so it's worth a note.
• listenEnd accepts LISTEN a, b. The boundary test expects this to be rejected, which matches PostgreSQL.
• Performance: the grammar walk is a single pass over the tokens with no allocations beyond the stack cursor. Your benchmarks show no change, which is plausible.
• Security: nothing concerning. The change only tightens parsing and keeps no new state.
• Tests and changelog: coverage of the new forms and boundaries is good, and the extra regression tests (SplitStatements, CREATE RULE) are welcome. The changelog entry should list the breaking-change example, VACUUM t junk, because it is a behaviour change.

Summary

The design follows the existing GRANT/COPY pattern and reads cleanly. The main risk is the missing fallback for dialects with no grammar, so please check that before merging. The rest are minor nits.
· branch fix/wp-h