tree-sitter-snowflake

A tree-sitter grammar for Snowflake SQL.

Scope: Snowflake only. This is not a general or multi-dialect SQL grammar and has no dialect-detection fallback. T-SQL and PostgreSQL syntax are out of scope. A file the grammar cannot parse produces ERROR nodes for a consumer to measure and act on, never a permissive best-effort tree.

Status: Tier 1 and Tier 2 coverage complete, plus most of Tier 3. SELECT in full (CTEs incl. recursive, set operators, joins incl. LATERAL, QUALIFY, GROUP BY GROUPING SETS/CUBE/ ROLLUP, window functions, MATCH_RECOGNIZE, PIVOT/UNPIVOT), MERGE, CREATE/ALTER/DROP for supported object families (table, view, materialized and dynamic table, stage, file format, stream, pipe, task, warehouse, database, schema, role, user, sequence, tag, integration, streamlit, Cortex search service and agent, semantic view), GRANT/REVOKE incl. future grants, COPY INTO both directions, PUT/GET/LIST/REMOVE, CALL/USE/SET/UNSET/SHOW/ DESCRIBE/COMMENT ON, transactions, procedure and function definitions with Snowflake Scripting control flow complete through cursors, RESULTSET, EXECUTE IMMEDIATE, and EXCEPTION ... WHEN, and the semi-structured : path operator, LATERAL FLATTEN, IDENTIFIER(...), * EXCLUDE/ RENAME, session variables ($var), positional/named binds, ARRAY/OBJECT constructors, and => named arguments. See "Consumer integration" below for application guidance.

The repository includes synthetic fixtures and a parse-rate harness for evaluating a local Snowflake SQL corpus. See "Testing" below for validation commands.

This is a from-scratch grammar, not a fork — see NOTICE for what it adapts from third-party sources (a substantially reduced dollar-quoting scanner) and what it takes as style reference (repository layout and rule organization).

Foreign-language procedure and function bodies

A LANGUAGE JAVASCRIPT/PYTHON/JAVA/SCALA body is not Snowflake SQL — it is a different language carried inside the body's delimiters. This grammar does not parse that language, but it exposes the body so another grammar can:

• The declared language is a language field on the create_procedure/create_function/ alter_procedure/alter_function node.
• The body's content — the text between the delimiters, excluding the delimiters themselves — is its own node, distinct for the two delimiter forms Snowflake accepts: dollar_quoted_body/dollar_quoted_script/dollar_quoted_expression for $$ ... $$, and string_body for the single-quoted AS '...' form.
• string_body's doubled-apostrophe escape ('') and backslash escapes are their own doubled_quote/escape nodes inside the content, so a consumer can reconstruct the real text and map a position in it back to a position in the file without re-lexing the body by hand.
• queries/injections.scm ships with the grammar, one branch per language. It only reaches the dollar-quoted form — tree-sitter's injection mechanism parses a byte range of the original source, and the single-quoted form's '' doubling means its byte range is not the embedded language's source.

The companion snowflake-bodies crate implements extraction, source-position mapping, embedded parsing, and advisory restriction checks. JavaScript and Python are included; Java and Scala are optional Cargo features. The grammar package does not depend on the companion. See docs/consumer-integration.md for measurement and masking integration.

Building

npx [email protected] -- tree-sitter generate
cargo build

The generated parser (src/parser.c, src/grammar.json, src/node-types.json) is committed, so consumers build the C parser through cc during cargo build without needing the tree-sitter CLI themselves. CI regenerates from grammar.js on every change and fails if the committed output differs.

src/scanner.c lexes $$-delimited bodies — the content between the delimiters can contain anything, including the grammar's own operators and unbalanced quotes, which a regex cannot match. It must be compiled alongside src/parser.c; vendoring one without the other produces a link error or a parser that cannot lex a dollar-quoted body.

generate takes under a minute on this grammar. Run it with a time cap and read the exit code, and check that src/parser.c's modification time moved before trusting a test run — a failed generate leaves the previous parser in place and writes nothing.

Testing

tree-sitter test
cargo test --workspace
cargo test -p snowflake-bodies --all-features

On a machine without MSVC, point the CLI at another C compiler:

CC=gcc CXX=g++ tree-sitter test

test/corpus/ holds the per-construct tree assertions (the regression suite); test/fixtures/ holds whole synthetic Snowflake files, checked by tools/fixture-check.mjs for zero ERROR/ MISSING and zero zero-width nodes, and by tools/fixture-leak-grep.mjs for absence of any source-corpus identifier. tools/keyword-probe.mjs probes every keyword against the AS-less alias slot and the unqualified column slot, the two positions where a keyword can silently eat an identifier. tools/node-kinds.mjs --check diffs the exported node-kind set against the committed snapshot in test/node-kinds.txt — a rename or removal there is a major version bump. tools/parse-rate/parse-rate.mjs measures the parse-rate budget over a real, locally-provided Snowflake corpus supplied by the caller. The committed fixtures are synthetic.

Consumer integration

Editors, source-analysis tools, and linters can use the grammar's named nodes and query files to analyze Snowflake SQL. The companion snowflake-bodies crate provides embedded-language parsing and source mapping. Tools with their own SQL parser can also use this grammar for differential testing.

See Consumer integration for structural node mappings, explicit dialect selection, comment/string masking, body analysis, and version compatibility.

License

MIT — see LICENSE and NOTICE.