ParserSQL

This directory vendors the ParserSQL SQL parser library that ProxySQL links into libsqlparser.a and uses for mysql-set_parser_algorithm=3 / pgsql-set_parser_algorithm=3 (the ParserSQL-backed SET parser) and related dialect-aware parsing paths.

Layout

• parsersql-<VERSION>.tar.gz — vendored upstream release tarball, expected to be byte-identical to git archive --prefix=ParserSQL-<VERSION>/ v<VERSION> against the upstream repo.
• parsersql — symlink to the extracted source directory (parsersql-<VERSION>/), created by the top-level build at extract time. The symlink is what every other Makefile rule and -I / -L flag references, so the rest of the build is version-agnostic.
• README.md — this file.

deps/Makefile's parsersql/parsersql/libsqlparser.a target unpacks the tarball (renaming ParserSQL-* → parsersql-* for case-insensitive filesystem friendliness) and runs make lib inside it.

Upstream

• Repository: https://github.com/ProxySQL/ParserSQL
• License: see parsersql/LICENSE after extraction.
• Maintainers: same team as ProxySQL (sysown). It is a sibling project, not a third-party dependency — bug reports and PRs are welcome there.

Bumping the version

# In the ParserSQL repo, after tagging vX.Y.Z:
git archive --format=tar.gz --prefix=ParserSQL-X.Y.Z/ vX.Y.Z \
    -o /tmp/parsersql-X.Y.Z.tar.gz

# In ProxySQL:
cd deps/parsersql
rm parsersql-*.tar.gz
cp /tmp/parsersql-X.Y.Z.tar.gz .
rm parsersql && ln -s parsersql-X.Y.Z parsersql
# Then git add the new tarball + symlink and commit.

The Makefile auto-detects the version via parsersql-*.tar.gz and ParserSQL-*/ glob patterns, so no Makefile change is required for a routine bump.

Working on the parser: fix it upstream, don't work around it

When set_parser_algorithm=3 (or any other ParserSQL-backed code path) misbehaves, the preferred fix is in ParserSQL itself, not a workaround in lib/Query_Processor_ParserSQL.cpp or in the PgSQL_Session / MySQL_Session handlers.

Reasons:

1. One source of truth. ParserSQL exists to be the SQL parsing layer. If ProxySQL silently compensates for a parser bug, the next consumer (or a future ProxySQL refactor) will hit the same bug and re-discover the workaround.
2. Tests live where the code lives. Every parser fix should ship with regression coverage in tests/test_*.cpp upstream so the bug can't come back. Workarounds in ProxySQL only get tested by the downstream TAP suite, which is heavier and slower to run.
3. Single version bump. The expected workflow is: identify all related parser issues at once, fix them upstream in one PR, tag a new release, and bump here exactly once — rather than dribbling a ProxySQL workaround per bug followed by a parser bump that then needs the workarounds reverted.
4. AST richness. ParserSQL already carries node-type information (e.g. NODE_LITERAL_STRING vs NODE_COLUMN_REF for 'val' vs name, FLAG_IDENT_DELIMITED for "name" vs name). When ProxySQL collapses everything to text in the walker, that information is lost and downstream code can't make correct decisions. Push richer typing into the AST upstream rather than re-deriving it in the consumer.

When a ProxySQL-side change is the right call

• The fix is about what ProxySQL does with a correctly-parsed AST (e.g. which validator to apply, which session variable to track, whether to lock a hostgroup). The AST is fine; the policy is in ProxySQL.
• The fix is in the walker (lib/Query_Processor_ParserSQL.cpp) — for example, handling a new NODE_* type that ParserSQL emits but the walker hasn't been taught about yet.
• The bug is in ProxySQL's pre/post processing around the parse (digest, normalization, scope-prefix stripping, etc.).

When a ParserSQL fix is the right call

• A SQL form that PostgreSQL or MySQL documents as valid returns ParseResult::ERROR or PARTIAL, or produces an AST shape that loses information the standard preserves.
• Two semantically-distinct inputs produce identical AST nodes (the parser is throwing away information the consumer needs).
• A keyword or shorthand isn't recognized (SET SCHEMA, SET LOCAL, SET TRANSACTION ...).
• The lexer truncates or mis-classifies tokens (@@\var`losing the closing backtick,$word` falling through to a generic identifier instead of being rejected, etc.).

Audit history

Significant audits / bumps:

• v1.0.8 (2026-05-23) — Tokenizer: unclosed delimited identifier ("name in PG, `name in MySQL) now emits TK_ERROR instead of silently consuming everything to EOF as one giant identifier. Closes the case-#172 cascade in ProxySQL's pgsql-set_parameter_validation_test-t under set_parser_algorithm=3, where SET search_path = "unclosed_quote, public was being parsed as identifier unclosed_quote, public and then accepted by the search_path validator — corrupting the stored value with what PG itself would have rejected outright. (PR #44)
• v1.0.7 (2026-05-23) — Tokenizer: allow $ as PG identifier continuation char (per PG lexical-syntax docs). Before, SET search_path = schema$1 truncated to schema and $1 fell through as a placeholder, breaking ProxySQL's set_parser_algorithm=3 path in pgsql-set_parameter_validation_test-t (case 169). First-char constraint preserved ($<word> at token start still emits TK_ERROR); MySQL behaviour unchanged. (PR #43)
• v1.0.6 (2026-05-23) — Hot-fix on top of v1.0.5: parse_set()'s PARTIAL → ERROR downgrade was too aggressive for multi-assignment SETs where one element is malformed alongside well-formed ones (SET sql_mode='X', whatever=, autocommit=1). v1.0.6 only downgrades when the parse produced no children at all, so the successful elements remain in the AST. (PR #42)
• v1.0.5 (2026-05-23) — Post-1.0.4 audit follow-up after a wider SET-form sweep. Adds three new node types (NODE_SET_ROLE, NODE_SET_SESSION_AUTHORIZATION, NODE_SET_CONSTRAINTS) so PG non-GUC SET forms are no longer misclassified as unknown GUC assignments; recognizes schema-qualified GUC names (SET pg_catalog.search_path, SET myapp.setting); preserves the full SET TIME ZONE INTERVAL '1' HOUR literal (was dropping value + unit); and surfaces clearly-malformed SET inputs (SET = 1, SET x = ;, SET x =, SET x = ,foo, SET search_path TO, bare SET, SET GLOBAL) as ParseResult::ERROR instead of PARTIAL. (PR #40)
• v1.0.4 (2026-05-23) — Six fixes from the initial set_parser_algorithm=3 audit: backtick-delimited @@\var`identifiers, PGSET SCHEMA/SET SEED shorthands, identifier quote-style flag (FLAG_IDENT_DELIMITED), bare PG $word→ERROR`. (PR #39)
• v1.0.3 — PG SET TIME ZONE alias + PG multi-value list (SET search_path TO 'a', 'b'). (PR #38)