Changelog

All notable changes to this project will be documented in this file. See commit-and-tag-version for commit guidelines.

0.1.1 (2026-10-02)

Packaging-only release. 0.1.0 reached npm and crates.io but never PyPI; this is the first release published there. No grammar changes.

Fixes

• PyPI sdist builds again. The sdist job now installs brotli, which setup.py needs to inflate the committed parser.c.br blobs.
• The sdist is installable. setup.py falls back to the already-inflated parser.c when there is no .br blob, imports brotli only when it actually inflates something, and the sdist now includes the base src/scanner.c that every dialect scanner #includes.
• A failed sdist build blocks the PyPI publish instead of uploading wheels without a source distribution.

0.1.0 (2026-08-31)

0.3.11 shipped an ANSI base with four dialects on top. There are now 22, the base grammar is strictly ISO SQL, and coverage is measured against outside parsers rather than against our own test suite. This is also, in practice, the project's first real release: nothing under the previous 0.3.x line was ever published to npm, crates.io or PyPI, so the version starts fresh at 0.1.0 instead of continuing 0.3.x — don't read the number as a downgrade or expect SemVer continuity with anything before it.

Breaking changes

• Project renamed and moved. The fork was cut loose from tree-sitter-sql-extended into its own repo, RelativelyUnknown/tree-sitter-sql-polyglot (GitHub org renamed from RedPandaMC). The npm scope, the crates.io/PyPI package name and the Go module path all changed to match; anything importing under the old names or pointing at the old org needs to update.
• Base grammar is ANSI only. LIMIT and LIMIT ... OFFSET, CREATE INDEX, CREATE MATERIALIZED VIEW, EXPLAIN, SELECT DISTINCT ON, trailing GROUP BY ... WITH ROLLUP/WITH CUBE, PL/SQL declarations and block/WHILE statements no longer parse in the base grammar, which keeps the ANSI OFFSET ... FETCH form instead. Every dialect that really has the syntax re-adds it in its own override, so no dialect lost a feature.
• Node names lost their dialect prefixes. bq_, db2_, plsql_, pg_, sf_, tsql_, iceberg_, oracle_, mysql_ and flink_ prefixes are gone; the nodes now use the plain cross-dialect names (optimizer_hint, table_partition_by, with_properties, and so on). Queries and tools written against the old names need updating.
• Dialect grammars were tightened to match their real engines. Constructs a dialect inherited but never actually supported now fail to parse: LIMIT in Oracle and Db2, OFFSET ... FETCH in MySQL and BigQuery, materialized views in Spark, MySQL, SQLite and T-SQL, savepoints in ClickHouse and Trino, QUALIFY in Trino and Athena.

Dialects

17 dialects joined the existing Spark/Hive, Databricks, PostgreSQL and MySQL grammars:

Dialect	Extends	Dialect	Extends
snowflake	base	clickhouse	base
bigquery	base	flink	base
sqlite	base	redshift	base
oracle	base	trino	base
db2	base	athena	trino
tsql	base	mariadb	mysql
duckdb	base	cockroachdb	postgres
teradata	base	spanner	bigquery
hana	base

Hive was also split out of the combined Spark/Hive grammar into its own dialect extending the base directly, giving the chain databricks -> spark -> hive -> base and bringing the total to 22.

Features

• Base (ANSI): GRANT/REVOKE; GROUP BY ROLLUP/CUBE/GROUPING SETS; FETCH {FIRST|NEXT} n {ROW|ROWS} {ONLY|WITH TIES}; WITHIN GROUP (ORDER BY ...); TRIM([{BOTH|LEADING|TRAILING} [char] FROM] str); interval qualifiers (INTERVAL '1' DAY); SAVEPOINT/RELEASE SAVEPOINT/ROLLBACK TO SAVEPOINT; START TRANSACTION with ISOLATION LEVEL and READ ONLY/READ WRITE; BEGIN/COMMIT/ROLLBACK WORK; typed temporal literals (DATE '...', TIME '...', TIMESTAMP '...'); datetime value functions (CURRENT_DATE, CURRENT_TIME, CURRENT_TIMESTAMP, LOCALTIME, LOCALTIMESTAMP); LIKE ... ESCAPE '...'; DECLARE ... CURSOR with OPEN/FETCH/CLOSE; named CONSTRAINT fk FOREIGN KEY (...) REFERENCES ... table constraints.
• COMMENT ON is wired into the base DDL dispatch and every dialect override. It was previously unreachable in 14 dialects.
• PostgreSQL: CREATE DOMAIN, CREATE CAST, CREATE AGGREGATE, CREATE SERVER and CREATE FOREIGN TABLE; PUBLICATION/SUBSCRIPTION; DO blocks; identity columns; recursive SEARCH/CYCLE; LISTEN/NOTIFY/UNLISTEN; FOR UPDATE/SHARE with NOWAIT/SKIP LOCKED and LOCK TABLE ... IN ... MODE; GRANT ... ON ALL TABLES/SEQUENCES/FUNCTIONS IN SCHEMA; SELECT DISTINCT ON (...); ILIKE; PREPARE/EXECUTE/DEALLOCATE; DELETE ... USING; CALL; EXPLAIN ( option ... ); table partitioning, PARTITION OF ... FOR VALUES, CREATE TABLE (LIKE ...) and INHERITS; complete VACUUM, ALTER INDEX and ALTER MATERIALIZED VIEW.
• MySQL: SHOW, DESCRIBE/DESC, LIMIT offset, count, @/@@ variables; stored procedure and trigger bodies; cursors, DECLARE HANDLER/CONDITION and procedural CASE; partitioning DDL and ALTER TABLE partition management; FOR UPDATE/SHARE with NOWAIT/SKIP LOCKED and LOCK IN SHARE MODE; SET GLOBAL/SESSION/PERSIST[_ONLY], SET NAMES, SET CHARACTER SET; CREATE/ALTER/DROP/RENAME USER with 'user'@'host' accounts; REPAIR/CHECK/ ANALYZE TABLE including histograms; STRAIGHT_JOIN; ALTER TABLE ... ALGORITHM=/LOCK=; multi-table DELETE; MATCH (...) AGAINST (...); SELECT ... INTO OUTFILE/DUMPFILE; GROUP BY ... WITH ROLLUP; complete ALTER VIEW, DROP INDEX and LOAD DATA.
• MariaDB: split from MySQL with temporal tables, system versioning, RETURNING, PACKAGE and CREATE OR REPLACE TABLE.
• Oracle: PL/SQL bodies (assignment, IF/ELSIF, WHILE, LOOP with EXIT [WHEN], numeric FOR ... IN 1..10, RETURN, CONTINUE, NULL), FORALL, BULK COLLECT INTO, RETURNING INTO, PRAGMA and PIPE ROW; cursors; CREATE TRIGGER with a PL/SQL block body; Oracle scalar types and optimizer hints; partitioning DDL and DATE literals; FLASHBACK AS OF, CREATE SYNONYM, CREATE DATABASE LINK; FOR UPDATE [OF ...] [NOWAIT|WAIT n|SKIP LOCKED]; ANALYZE TABLE/INDEX; the MODEL clause; ALTER SYSTEM/ALTER SESSION, CREATE/DROP DIRECTORY, GRANT READ/WRITE ON DIRECTORY; PIVOT/UNPIVOT; MATCH_RECOGNIZE; fn(...) KEEP (DENSE_RANK {FIRST|LAST} ORDER BY ...); ORGANIZATION {HEAP|INDEX|EXTERNAL}; JSON_TABLE(...) and XMLTABLE(...); materialized view logs, zonemaps and index rebuild.
• Db2: SQL PL bodies (BEGIN...END, DECLARE, SET, IF/ELSEIF, WHILE...DO, LOOP, LEAVE, ITERATE); cursors and the FOR ... AS ... DO ... END FOR loop; SELECT ... FROM FINAL/NEW/OLD TABLE (dml); LABEL ON {TABLE|COLUMN}; temporal queries FOR {SYSTEM_TIME|BUSINESS_TIME}; DECLARE GLOBAL TEMPORARY TABLE; CREATE TABLE ... ORGANIZE BY {ROW|COLUMN|DIMENSIONS (...)} for column-organized and multidimensional clustering tables; AUDIT, WHENEVER, GOTO, ALLOCATE CURSOR, ASSOCIATE LOCATORS; SET SCHEMA and the special registers it accepts.
• T-SQL: scripting, CROSS/OUTER APPLY, PIVOT/UNPIVOT, #temp/##global identifiers; cursors; EXEC/EXECUTE, RETURN, OPTION() hints, WAITFOR; USE database; CREATE/DROP SYNONYM; LOGIN and USER security DDL; IDENTITY[(seed, increment)] and computed columns col AS (expr) [PERSISTED]; CREATE TYPE ... AS TABLE (...) and CREATE TYPE ... FROM base_type; select-list variable assignment (SELECT @v = expr); CREATE/ALTER DATABASE file specs and option lists; GROUP BY ... WITH ROLLUP/WITH CUBE.
• BigQuery: native types (INT64, FLOAT64, BYTES, STRUCT<...>, ARRAY<...>), UNNEST, backtick identifiers, QUALIFY; PARTITION BY/CLUSTER BY DDL; CALL, RAISE [USING MESSAGE = ...] and RETURN; THEN RETURN on DML; PIVOT/UNPIVOT; FOR SYSTEM_TIME AS OF; UDF, table-function and procedure DDL; LOAD DATA {INTO|OVERWRITE} ... FROM FILES (...); CREATE SNAPSHOT TABLE ... CLONE; SET OPTIONS across ALTER, key constraints and collation.
• Snowflake: :: cast and @stage as a FROM source; SHOW and DESCRIBE; CLONE, UNDROP, CREATE/ALTER WAREHOUSE; CREATE FILE FORMAT and COPY INTO with an inline subquery; stage DDL and LIST @stage; CALL; INSERT OVERWRITE INTO; CREATE EXTERNAL TABLE; ALTER TABLE ... CLUSTER BY/DROP CLUSTERING KEY; USE DATABASE/SCHEMA/WAREHOUSE/ROLE; RETURNING on DML; START WITH ... CONNECT BY ... PRIOR; SELECT * [ILIKE] [EXCLUDE] [RENAME]; multi-table INSERT ALL/INSERT FIRST; property lists on database, schema and warehouse DDL.
• Redshift: DISTKEY/SORTKEY/DISTSTYLE/ENCODE, external schemas and tables, COPY/ UNLOAD; ALTER GROUP, diststyle and sortkey changes, session SET; PARTITIONED BY on external tables; GRANT ... ON ALL TABLES/FUNCTIONS/PROCEDURES IN SCHEMA; complete user-management DDL; ALTER TABLE ... APPEND FROM; CREATE EXTERNAL FUNCTION ... LAMBDA; CREATE MATERIALIZED VIEW ... [BACKUP|AUTO REFRESH]; CREATE/ALTER DATASHARE; CREATE MODEL (Redshift ML); ALTER MATERIALIZED VIEW and ALTER DATABASE actions.
• SQLite: INSERT OR REPLACE|IGNORE, UPSERT, AUTOINCREMENT, INDEXED BY/NOT INDEXED; generated columns; CREATE TRIGGER ... BEGIN stmt; ... END; GLOB and MATCH; RETURNING on UPDATE and DELETE; UPDATE OR <action>; BEGIN isolation modes.
• DuckDB: FROM-first SELECT, SELECT * EXCLUDE/REPLACE/RENAME, lambdas, struct/map/list literals, ASOF/POSITIONAL JOIN, ATTACH; COPY TO/FROM and file-reader table functions; CREATE MACRO, EXPORT/IMPORT DATABASE; a bare string literal as a FROM source; RETURNING on DML; UNION/EXCEPT/INTERSECT ... BY NAME; USING SAMPLE; open-ended list slicing (l[2:], l[:3]); DISTINCT ON; ATTACH IF NOT EXISTS/OR REPLACE and DETACH IF EXISTS.
• Trino: PREPARE/EXECUTE/DEALLOCATE, MATCH_RECOGNIZE, TABLESAMPLE, ARRAY/MAP/ROW types, lambdas; SHOW, DESCRIBE, ANALYZE; SELECT ... FOR UPDATE; CREATE VIEW ... SECURITY {DEFINER|INVOKER}; ALTER TABLE ... EXECUTE proc(...); SET ROLE; SET TIME ZONE; DENY; SET PATH; SET SESSION AUTHORIZATION; SET AUTHORIZATION, SET PROPERTIES and ALTER VIEW ... REFRESH.
• Athena: UNLOAD, MSCK REPAIR TABLE; CREATE EXTERNAL TABLE with Hive storage syntax; VACUUM; ALTER TABLE ADD/DROP PARTITION and SET LOCATION; OPTIMIZE ... REWRITE DATA USING BIN_PACK for Iceberg compaction; managed Iceberg CREATE TABLE ... TBLPROPERTIES ('table_type'='ICEBERG').
• ClickHouse: ENGINE = MergeTree(), column MATERIALIZED/ALIAS/EPHEMERAL/CODEC/TTL, PREWHERE, FINAL, ARRAY JOIN, LIMIT n BY, SAMPLE, WITH TOTALS, ORDER BY ... WITH FILL; mutations including ALTER UPDATE/DELETE ... IN PARTITION; KILL QUERY/MUTATION and access control DDL; BACKUP/RESTORE; EXCHANGE TABLES; ALTER TABLE FREEZE/UNFREEZE; INSERT ... RETURNING; GLOBAL JOIN and x GLOBAL [NOT] IN (...); SELECT * transformers (EXCEPT/APPLY/REPLACE); INSERT ... FORMAT fmt and INSERT INTO FUNCTION f(...); CHECK, DESCRIBE, EXISTS, MOVE, UNDROP and collection DDL; settings constraints on roles and profiles; OPTIMIZE options and PARTITION ID.
• Flink: connector DDL (WITH (...)), WATERMARK FOR, windowing TVFs, temporal joins, CREATE CATALOG, LOAD/UNLOAD MODULE, statement sets; LATERAL TABLE(func(...)) AS t(cols); MATCH_RECOGNIZE; REPLACE TABLE, DROP TEMPORARY and DROP DATABASE drop behaviour.
• Hive: LOAD DATA [LOCAL] INPATH, INSERT OVERWRITE [LOCAL] DIRECTORY, multi-table INSERT, CLUSTER/DISTRIBUTE/SORT BY; SELECT TRANSFORM(...) USING 'script' AS (...); SHOW and DESCRIBE [FORMATTED|EXTENDED]; ALTER TABLE ... EXCHANGE PARTITION; CONCATENATE; EXPORT/IMPORT TABLE; CREATE TEMPORARY MACRO; role DDL.
• Spark: QUALIFY, PIVOT/UNPIVOT, query-level CLUSTER/DISTRIBUTE/SORT BY, CREATE TABLE ... USING/OPTIONS; Spark 4.x VARIANT, MERGE ... BY SOURCE, collation and SET VAR; Iceberg DDL (CALL, partition transforms, WRITE ORDERED/DISTRIBUTED); GROUP BY ALL; INSERT {INTO|OVERWRITE} TABLE and INSERT INTO ... BY NAME; Delta time travel in FROM; CREATE [TEMPORARY] VIEW ... USING source [OPTIONS (...)]; database DDL options, DROP ... PURGE and DROP TEMPORARY FUNCTION; the inherited Hive role DDL.
• Databricks: CREATE STREAMING TABLE, LIVE TABLE and Unity Catalog objects; APPLY CHANGES INTO; Iceberg and Unity Catalog ALTER TABLE (branches, tags, column position); liquid clustering (CLUSTER BY (...), CLUSTER BY NONE); CREATE MATERIALIZED VIEW ... [SCHEDULE ...]; COPY INTO ... FILEFORMAT = ....
• CockroachDB: ALTER TABLE ... SPLIT AT/UNSPLIT AT/SCATTER; column families (FAMILY name (cols)); SET LOCALITY; CONFIGURE ZONE USING; multi-region DDL and ALTER INDEX actions.
• Spanner: generated columns col AS (expr) STORED; CREATE VIEW ... SQL SECURITY; CREATE SEQUENCE ... OPTIONS (...); ALTER DATABASE ... SET OPTIONS (...); ALTER TABLE ... {ADD|REPLACE|DROP} ROW DELETION POLICY; identity columns and synonym DDL.
• Teradata: [NOT] CASESPECIFIC, UPPERCASE and TITLE column attributes; CREATE {JOIN|HASH} INDEX ... AS SELECT; atomic UPSERT (UPDATE ... SET ... ELSE INSERT); CREATE TABLE ... AS ... WITH [NO] DATA [AND [NO] STATISTICS]; CALL, DUMP EXPLAIN, INSERT EXPLAIN, EXECUTE FUNCTION, SELECT AND CONSUME, USING request modifiers, and index and partition analysis.
• SAP HANA: SELECT TOP n; ALTER TABLE ... ADD (col type, ...); SQLScript control flow (IF/ELSEIF/ELSE, WHILE ... DO, FOR ... IN a..b DO, BREAK/CONTINUE); SQLScript table-variable assignment (tabvar = SELECT ...); EXPORT, IMPORT, UNSET and DO; index rebuild, role groups, table groups and schema synonyms.
• Upsert targets: ON CONFLICT (col) [WHERE ...] and ON CONSTRAINT name in PostgreSQL, SQLite, CockroachDB and DuckDB.
• GROUP BY ALL in Spark, Databricks, BigQuery and Snowflake.

Bug Fixes

• Lexer precedence across dialects. A dialect keyword declared as token(prec(1, ...)) beat a longer base keyword sharing its prefix, so MATCHED lexed as MATCH + ED and broke MERGE ... WHEN MATCHED. Audited and fixed across every dialect.
• DuckDB: restored MERGE INTO, removed on the belief that DuckDB has no MERGE. Version 1.4.0 added it.
• T-SQL: restored CREATE MATERIALIZED VIEW. Azure Synapse dedicated SQL pools support it, and this dialect covers Synapse.
• Spark: restored CREATE MATERIALIZED VIEW, which Spark 4.x documents under Declarative Pipelines.
• Spanner: restored SELECT ... FOR UPDATE, shadowed by the inherited BigQuery FOR SYSTEM_TIME AS OF clause.
• MariaDB: restored the MySQL features its _column_constraint override had dropped (AUTO_INCREMENT, STORED/VIRTUAL generated columns, foreign key ON DELETE/UPDATE).
• Databricks: restored iceberg_write_order and DISTRIBUTED BY PARTITION, dropped by the _alter_specifications override.
• Spark: restored the inherited Hive SHOW/DESCRIBE/MSCK REPAIR/LOAD DATA statements, dropped by the _ddl_statement override.
• Base: _decimal_number no longer eats the first dot of Oracle's .. range operator.
• Base: a CREATE FUNCTION with a bare RETURN body no longer swallows its own terminator and breaks the following statement.
• MySQL: dropped LINES STARTING BY, which broke MariaDB's SELECT path.
• Db2: SET SCHEMA is reachable as a statement, and CREATE SECURITY LABEL's element list is unambiguous.
• Teradata: request modifiers are prefixes rather than wrappers, and KEEP is reserved in INITIATE INDEX ANALYSIS.
• Oracle: flattened ALTER INDEX REBUILD's option list.
• BigQuery: ADD's unnamed constraint form is restricted to the shapes that allow it.
• ClickHouse: the operator table lists ILIKE rather than SIMILAR TO, which ClickHouse does not have.
• SAP HANA: ALTER SYSTEM option values are constrained to a scalar, and the rules duplicating hana_object_statement were removed.
• Windows: generate.js uses fileURLToPath(), fixing parser generation on Windows paths.
• Highlights: removed queries referencing node types that no longer exist.

Performance

• Resolved self-conflicts with prec.right on interval, time/timestamp and T-SQL's output_clause instead of GLR conflicts.
• Merged T-SQL's bracket and temp-identifier tokens, and dropped MySQL's redundant backtick-identifier sequence.
• Dropped unnecessary conflict declarations across dialects, and SIMILAR TO from SQLite.
• The dollar-quote scanner keeps its tag size counter on the stack. It was one malloc/free per character of a dollar-quoted body.

Tooling and CI

• tools/coverage.py replaces scorecard.py, corroborate.py, parse_rate.py and diff_parse.py. Every feature probe is parsed by this grammar and by independent reference parsers (SQLGlot, ANTLR grammars-v4, pglast, sqlfluff), so a feature counts as covered only when an outside parser agrees. The report surfaces suspect probes (we accept, no reference does) and confirmed gaps (we reject, a reference accepts), and scores the base grammar against ISO Core.
• tools/coverage-probes.yml carries a per-dialect statement inventory. Scoring one shared feature list against all 23 grammars had counted PostgreSQL-only statements as gaps in 20 engines that never had them.
• tools/glr_scan.py ranks grammars by parse-table size, state count, generation time and peak memory.
• tools/parse_bench.py measures parse throughput per dialect on a real workload and on shapes that provoke GLR stack splitting, reporting how time grows with input size.
• Generated parsers are no longer committed. Generation is hash-cached, self-healing against poisoned caches, parallel with a memory cap, and pinned to tree-sitter CLI 0.26.3.
• CI shards the corpus tests, cancels superseded runs, and reuses the matrix's parsers in the coverage job.
• The docs site moved to VitePress, with the coverage page generated on every run.
• main now requires a pull request, one code-owner approval, and a passing ci-passed check; force-pushes and branch deletion are blocked (repo admins may still bypass in an emergency).
• Added a main + release/x.y branch model, so releases and hotfixes route through a release branch rather than landing directly on main; CODEOWNERS, a PR template, and a release/** trigger on ci.yml back it.
• npm, crates.io and PyPI publishing switched to OIDC trusted publishing, dropping the long-lived NPM_TOKEN/CARGO_REGISTRY_TOKEN/PYPI_API_TOKEN secrets entirely; each publish job now runs under its own GitHub Environment for secret scoping. PyPI's publish step vendors its sdist/wheel build directly into publish.yml rather than calling a reusable workflow, since OIDC can't verify a trusted publisher through one.
• Per-language binding tests were restructured: Go keeps the one full 23-dialect suite (Ubuntu only), while Node, Go, Python, Rust and C/CMake each get a lightweight base+postgres check across macOS/Ubuntu/Windows, catching an OS- or compiler-specific native-build break without paying to build all 22 dialects on every OS. Windows runners were also unified on windows-2025 (test_go had drifted onto windows-latest).

Bindings

• Rust, Node and Python bindings are now per-dialect instead of one bundle carrying all 22. Rust gets one Cargo feature per dialect (default [], full for everything); build.rs only compiles a dialect's parser.c when its feature is on. Node and Python each register one native module per dialect and only load the one actually imported, instead of dlopening/importing all 22 up front. Each dialect's generated parser.c/node-types.json is brotli-compressed into a committed .br blob and inflated at build time — the raw multi-dialect crate blew past crates.io's 10MB limit (17.8MB gzip-compressed); brotli gets the published crate down to 7.9MB.
• Go and Swift bindings get the same per-dialect, lazily-loaded treatment, using each ecosystem's own native mechanism instead of a feature-flag system: one importable Go subpackage per dialect (bindings/go/postgres, ...) with its own #cgo directive, so importing one never compiles the other 21's C sources; one SwiftPM target and library product per dialect in Package.swift (TreeSitterSqlPostgres, ...). Both bindings previously exposed only the base grammar and had no CI coverage at all; Go now gets a full 23-dialect build+test matrix and Swift builds and tests on macOS in CI.
• C/CMake now supports every dialect, not just base: one CMake option per dialect (-DTREE_SITTER_SQL_<DIALECT>=ON, or -DTREE_SITTER_SQL_FULL=ON for all 22 at once), mirroring Cargo's feature-flag shape. CMakeLists.txt and bindings/c/ are now generated by scripts/generate-bindings.js alongside the other four bindings, rather than hand-maintained and scoped to base only.

Documentation

• README.md, AGENTS.md and CONTRIBUTING.md cover all 22 dialects and the current architecture. README.md also gained a plain, factual fork-history and attribution note explaining its lineage from tree-sitter-sql/tree-sitter-sequel through tree-sitter-sql-extended.
• docs/usage.md is a new page covering install/import for every registry this project publishes to (npm, crates.io, PyPI), later extended to document the per-dialect Cargo features and Node/Python lazy loading, and finally to cover all six binding targets: Rust/Node/Python/Go/Swift/CMake+C.

0.3.11 (2026-05-23)

Features

• Multi-dialect grammar architecture: an ANSI SQL base with independently compiled dialect extensions for Spark/Hive, Databricks/Unity Catalog, PostgreSQL and MySQL
• Hash-based grammar caching (scripts/generate.js), which skips tree-sitter generate when the sources are unchanged
• Per-dialect test corpora in <dialect>/test/corpus/
• scripts/bump-version.sh, which syncs the version across all 5 manifests
• agents.md, an architecture guide for grammar composition and dialect extension