insights/rust/insightsql/parser

insightsql_parser_rs

Hand-rolled Rust InsightsQL parser. Pratt + recursive descent. Same JSON AST shape as the C++ ANTLR parser so the two can be cross-validated query-by-query. Ships as a Python extension via maturin and is selected as the rust-json backend in insights/insightsql/parser.py.

About 15× faster than the C++ parser on parse_expr and 50–55× on parse_select, against the same input, on the same machine. The numbers come from insights/insightsql/scripts/parser_bench.py; re-run locally before and after any non-trivial change.

The C++ parser is the source of truth

When grammar, AST shape, or any visible behaviour disagrees between the two, the C++ ANTLR parser is right and this one is wrong. The C++ parser is generated from insights/insightsql/grammar/InsightsQLLexer.*.g4

• InsightsQLParser.g4 via ANTLR4. The Rust parser does not consume those grammar files; it hand-implements the same recognition behaviour.

This means any grammar change is a two-step change:

1. Update the ANTLR grammar and rebuild the C++ parser. Get the new shape working end-to-end on cpp-json. Pin the new behaviour with regression tests (see "Tools" below).

2. Bring the Rust parser to parity. Run the diagnostics, find the new divergences, fix them. This is the part an LLM agent can drive in a long-running loop.

Skipping step 1 produces a Rust parser that "works" but on a shape the C++ parser rejects, which means Cloud's printer / planner will reject too, because they're built on top of cpp-json's output. Get the oracle right first, then the candidate.

What's in this crate

Path	What it does
src/lib.rs	PyO3 entry points (parse_expr_json, parse_select_json, parse_program_json, parse_order_expr_json, parse_full_template_string_json). Each returns a JSON string; on error the JSON is an {"error": true, ...} envelope insights/insightsql/json_ast.py decodes into InsightsQLSyntaxError / ExposedInsightsQLError.
src/lex.rs	Lexer. Hand-rolled state machine matching the ANTLR-generated C++ lexer's tokens + mode stack (default / template-string / InsightsQLX-tag / InsightsQLX-text). When you add a new keyword to the grammar, add it here too.
src/parse.rs	Parser core: Parser struct, public entry points, the Pratt expression parser (parse_expr_bp), positions (pos_obj, wrap_pos, wrap_pos_to), char-offset / line-col tables, checkpoint / restore for speculative branches.
src/parse/{expr,select,program,join,cte,insightsqlx,template}.rs	Per-rule parsing. Most grammar changes land in one of these.
src/parse/bp.rs	Binding-power table + build_infix / merge_and_or / merge_concat. The precedence ladder lives here; new operators usually need an entry in infix_bp and a build_infix arm.
src/emit.rs	AST-node builders + position helpers (with_pos is idempotent, replace_pos overrides, no_pos reserves null keys to opt out of the wrap). When you add a new AST node, add a helper here so callers don't hand-build the JSON object.
src/error.rs	ParseError + the JSON error envelope.

Building locally

# One-time: install the rust toolchain via flox / rustup (the workspace
# Cargo.toml is at `rust/Cargo.toml`).

# Build + install the wheel into the venv (editable). Re-run after each
# rust source change.
uv pip install -e rust/insightsql/parser

# Or via maturin directly (faster incremental):
maturin develop --release --manifest-path rust/insightsql/parser/Cargo.toml

# Sanity check
python -c "import insightsql_parser_rs; print(insightsql_parser_rs.parse_expr_json('1 + 2'))"

maturin builds a single cp312-abi3 wheel that works on Python 3.12+. CI builds wheels for Linux x86_64/aarch64 (manylinux 2_28 + musllinux 1_2) and macOS arm64/x86_64; see .github/workflows/build-insightsql-parser-rs.yml.

Coverage-instrumented build (for the parser-parity PBT)

A second build path exists for fuzz-driven parser-parity work: build the crate with --features coverage and SanitizerCoverage's trace-pc-guard pass, and the resulting wheel exposes two PyO3 functions (cov_snapshot(), cov_reset()) that the pytest PBT (insights/insightsql/test/test_parser_pbt.py) uses to feed a per-example rust_edges novelty count into Hypothesis target() on top of the existing ast_depth / novel_kpaths signals. See src/cov.rs for the callback + bitmap implementation. The production wheel is unchanged: without the feature, neither cov.rs nor the sancov pass are compiled, and the PBT detects the absence via hasattr(insightsql_parser_rs, "cov_snapshot") and silently skips the third steering label.

Cargo applies RUSTFLAGS globally to every crate it compiles, including dependency proc macros and build scripts, and those don't have our __sanitizer_cov_trace_pc_guard* callbacks linked in, so a naive RUSTFLAGS=… maturin build --features coverage segfaults the host build scripts on macOS (and fails to link on Linux). The fix is a small per-crate rustc wrapper that strips the sancov flags from every invocation except the one for insightsql_parser_rs. That wrapper lives at scripts/cov-rustc-wrapper.py:

WRAPPER=$PWD/rust/insightsql/parser/scripts/cov-rustc-wrapper.py

# Build the instrumented wheel into a temp dir.
RUSTC_WRAPPER="$WRAPPER" \
RUSTFLAGS="-C passes=sancov-module \
           -C llvm-args=-sanitizer-coverage-level=3 \
           -C llvm-args=-sanitizer-coverage-trace-pc-guard" \
maturin build --release --manifest-path rust/insightsql/parser/Cargo.toml \
              --features coverage --out /tmp/cov-wheel

# Install it into the active venv (in the same shell session — the shared
# flox venv prunes manually-installed wheels otherwise).
VIRTUAL_ENV=$PWD/.flox/cache/venv uv pip install --reinstall --no-deps \
    /tmp/cov-wheel/insightsql_parser_rs-*.whl

# Verify the instrumented wheel is loaded.
python -c "import insightsql_parser_rs; print(hasattr(insightsql_parser_rs, 'cov_snapshot'))"

# Run the parity PBT; the rust_edges target appears in --hypothesis-show-statistics.
RUN_PBT=1 insightscli test insights/insightsql/test/test_parser_pbt.py::TestParserBackendEquivalence \
    --hypothesis-show-statistics

rust_edges will show up in Hypothesis's per-test target stats alongside ast_depth. The wrapper has been verified to work locally on macOS aarch64. On a normal install (no coverage build) the PBT just doesn't emit the third label; nothing else changes.

If you want a production CI job to produce this wheel as a sidecar, mirror the above invocation in a separate workflow file and publish it under a separate package name (e.g. insightsql_parser_rs_cov) so production deploys can't accidentally pull the instrumented wheel.

Publishing

The crate is pinned via the insightsql-parser-rs==X.Y.Z line in the repo-root pyproject.toml. Bump the version in both:

• Cargo.toml
• pyproject.toml

(They must match. The PR check at .github/workflows/build-insightsql-parser-rs.yml enforces this.)

Version is intentionally locked in step with common/insightsql_parser (the C++ parser PyPI package) so a bump signals "both parsers move together." The publish workflow builds wheels, pushes to PyPI via trusted publishing, then opens a follow-up PR that updates the repo-root pin.

Adding a new grammar feature

Steps 1–4 are the one-time grammar-update process — done once, human-driven. Step 5 (running the parity loop below) is the long-running, agent-friendly part.

1. Update InsightsQLLexer.*.g4 and InsightsQLParser.g4. Run pnpm grammar:build to regenerate the C++ ANTLR artefacts:

   pnpm grammar:build
   

   That step requires the antlr 4.13.2 binary on PATH; instructions in insights/insightsql/grammar/README.md. The script rewrites common/insightsql_parser/InsightsQL{Lexer,Parser}.{cpp,h,interp,tokens}. The C++ parser now recognises the new shape (the Rust parser is hand-written, so update it to match in the steps below).

2. Pick the AST emission. Decide what JSON the cpp visitor should return for the new shape. Either reuse an existing AST node or add a new one in insights/insightsql/ast.py. The Python AST is shared between backends, so any new node has to land there first, otherwise insights/insightsql/json_ast.py::deserialize_ast will crash on it.

3. Update the cpp visitor. Add the VISIT(YourNewRule) arm in common/insightsql_parser/parser_json.cpp. Mirror cpp's conventions: call addPositionInfo(json, ctx) per rule unless you specifically want a position-less node (see "Position parity" below). Rebuild the cpp wheel (pip install ./common/insightsql_parser).

4. Pin the new behaviour with a regression test. Add a test (and a rust-rejects-it negative test if the grammar tightens) into the parser_test_factory suite in insights/insightsql/test/_test_parser.py. The factory runs every test against cpp-json, rust-json, and rust-py; on a fresh grammar change the test passes on cpp and fails on rust. That fail is the starting state for the parity loop.

5. Run the parity loop. See the next section.

The parity loop

The agent loop that brings rust to behavioural parity with cpp. Long-running; the diagnostics produce concrete diffs the agent attacks one at a time.

Before each iteration, rebuild both parser wheels from local source. uv pip install / pip install will happily resolve to the published PyPI wheel and ignore the working tree, so a fresh maturin develop and a fresh pip install ./common/insightsql_parser are non-negotiable — the diagnostics test what's installed, not what's on disk.

maturin develop --release --manifest-path rust/insightsql/parser/Cargo.toml
pip install --force-reinstall --no-deps ./common/insightsql_parser

maturin develop writes the wheel into the active venv as editable; --force-reinstall --no-deps for the cpp wheel sidesteps pip's "already-satisfied" short-circuit when the in-tree version matches the PyPI pin. Skip these and the loop will silently chase divergences that have already been fixed in the working tree.

1. Generate a new divergence. In priority order:

   • existing failing regression tests (highest signal);
   • real production queries via log_corpus_diagnostic.py / hog_corpus_diagnostic.py (the script corpus has been at 100% for a while — usually skip);
   • PBT (pbt_diagnostic.py --rule expr|select|program);
   • thinking hard about edge cases the grammar surface invites — pipe each candidate through shrink_query.py to confirm it diverges and get the minimal form.

   For everything other than regression tests, start with a small budget (lower --n, less thinking time) and increase until at least one divergence surfaces.

2. Reduce + pin. Shrink each divergence to its minimal form via shrinkray — pipe one query through shrink_query.py, or use --shrink-failures on the corpus runners — and add it as a regression test in _test_parser.py's factory so it runs on all three backends.

3. Read before fixing. Read the grammar AND the cpp visitor for the rule. 100% identical behaviour means knowing exactly what cpp does — guessing leads to fixes that resurface on a deeper PBT run.

4. Fix the rust parser. Prefer general fixes that won't break on deeper nesting; a depth-0-only special case is a smell. Print a one-paragraph report for the human operator so progress is visible while the loop runs autonomously.

5. Re-run the regression suite. Anything below the previous baseline goes back to step 1.

Generating divergences is the slow step. Run discovery in parallel in the background:

• pbt_diagnostic.py --rule select
• pbt_diagnostic.py --rule expr
• pbt_diagnostic.py --rule program
• log_corpus_diagnostic.py (real query corpus)
• a research subagent grepping for cpp-vs-rust visitor differences
• a research subagent brainstorming adversarial edge cases, each confirmed + minimised through shrink_query.py

Most of these can stream divergences as they're found. Once at least one known divergence is in hand, start fixing it while the parallel runs keep mining the long tail.

Tools for parity work

Every script below has the same --oracle / --candidate flag pair and defaults to cpp-json vs rust-json. The diagnostics include per-node start / end positions in the comparison by default; set CLEAR_LOCATIONS=1 to strip positions when you want a structural-only read.

Regression tests in insights/insightsql/test/

insightscli test insights/insightsql/test/test_parser_cpp_json.py
insightscli test insights/insightsql/test/test_parser_rust_json.py
insightscli test insights/insightsql/test/test_parser_rust_py.py

The behaviour suite + regression pins live in _test_parser.py's parser_test_factory. The three files above are thin subclasses that spawn one runnable test entry per (backend, case) combination. When you find a new divergence, add a reduced regression to the factory — it picks up all three backends automatically.

Property-based testing via insights/insightsql/scripts/pbt_diagnostic.py

PYTHONPATH=. python insights/insightsql/scripts/pbt_diagnostic.py \
    --n 5000 --rule program

# Per rule:
--rule expr     # standalone column expressions
--rule select   # SELECT / SELECT-set statements
--rule program  # full Script programs (declarations + statements + exprs)

Generates ~5 000 random grammar surface examples per rule, parses with oracle and candidate, buckets divergences by AST shape, and prints reproducers. Use --shrink-failures to reduce each divergence to a minimal example via shrinkray (needs the optional insightsql-parser-parity group — uv sync --group insightsql-parser-parity).

Real-query corpora via log_corpus_diagnostic.py / hog_corpus_diagnostic.py

# SELECT queries from the last 7 days of production traffic
# (redacted, AI-data-processing-approved teams only):
PYTHONPATH=. python insights/insightsql/scripts/log_corpus_diagnostic.py

# Script programs from production (transformations, destinations, …):
PYTHONPATH=. python insights/insightsql/scripts/hog_corpus_diagnostic.py

Both auto-download via insightscli metabase:query and cache locally under insights/insightsql/scripts/.local/. Pass --skip-download to reuse the existing dump while iterating. Failures are written one block per divergence to a .sql / .script file the agent can chew through. Add --shrink-failures to reduce each failing query to a minimal repro via shrinkray before it's written.

Shrink one query via shrink_query.py

# Pipe in a query that diverges (or that you think might); get the
# minimal repro back on stdout:
echo '<query>' | PYTHONPATH=. python insights/insightsql/scripts/shrink_query.py --rule select

The "think hard about edge cases" arm of the loop, and the reducer for any single divergence. Reads one query on stdin, writes the smallest variant that still triggers the same divergence to stdout (chatter goes to stderr, so stdout is exactly the repro). Doubles as a divergence check: a query the backends agree on exits non-zero with the reason on stderr. An agent brainstorming edge cases pipes each candidate through it and keeps the ones that come back shrunk.

The diagnostics need the optional parity group

Shrinking is powered by shrinkray, which lives in the insightsql-parser-parity group rather than the default dev install (its transitive deps — textual, libcst, black — are heavy). The parity scripts in insights/insightsql/scripts/ import it at module level, so they require that group:

uv sync --group insightsql-parser-parity

Without it they fail fast at startup with a ModuleNotFoundError — these are parity-work-only tools, so they just require the parity dependency rather than degrading to a no-shrink mode.

Perf bench via insights/insightsql/scripts/parser_bench.py

CANDIDATE_BACKEND=rust-json PYTHONPATH=. \
    python insights/insightsql/scripts/parser_bench.py

Runs both parsers against a fixed corpus of representative queries (small / medium / nested / pathological) and prints an oracle / candidate ratio per row. Re-run before and after any non-trivial change. If parse_select mean drops noticeably (the parse_select speedup is the headline number), find out why before landing.

Shadow compare in TEST via cpp-with-rust-shadow

In TEST mode the default backend is cpp-with-rust-shadow: both backends parse, ASTs are compared including per-node start / end positions (matching the diagnostics' default), mismatches raise so the failing test points right at the offending query. In production this same mode runs at a 1% sample and only logs — a mismatch there is tagged insightsql_parser_position_only_mismatch so position-only divergences are triaged apart from structural ones. Useful when a regression slips past the PBT but shows up in the suite.

from insights.insightsql.constants import InsightsQLParserBackend
parse_expr(src, backend=InsightsQLParserBackend.CPP_WITH_RUST_SHADOW)

Rules of thumb for the parity loop

These aren't always obvious from the diagnostics alone:

• Prefer the generalising fix. When two implementations both pass the failing cases, pick the one that doesn't depend on the input shape. A wrap_pos call at a single emit site beats a depth-aware conditional. A change to the binding-power table beats an ad-hoc check in the consumer.

• Position bugs hide behind structural bugs. Always run the PBT with positions on (the default); CLEAR_LOCATIONS=1 is for diagnosing structural regressions only. A 99% structural match can mask a 50% position-aware match.

• Look at the cpp visitor before guessing. Every per-node position decision in this parser has a cpp counterpart in common/insightsql_parser/parser_json.cpp. If the cpp visitor calls addPositionInfo(json, ctx) you need a wrap on the rust side; if it doesn't, you need emit::no_pos (or the helper for that node already does it).

• Watch the perf bench. Position emission isn't free. Cache O(N) computations on Parser rather than recomputing per emit; the is_ascii_src field is the canonical example.

• Don't fix one rule at a time at the expense of others. A one-line wrap in parse/expr.rs can move three PBT rules at once. Run all three PBTs after each change, not just the one you started with.

Position parity (the non-obvious part)

The C++ visitor decides per-node whether to emit positions via addPositionInfo(json, ctx). Some nodes are deliberately position-less (NamedArgument, ColumnsExpr in qualified-asterisk column slots, etc.) so the rust parser has to match that exactly.

Three position helpers in emit.rs cover the three cases:

Helper	When to use
with_pos	Default. Adds start / end if not already set. Used by Parser::wrap_pos and wrap_pos_to. Idempotent so the outer pratt-loop wrap doesn't trample inner spans.
replace_pos	Override existing start / end. Used by the bare-paren grammar alts ((* REPLACE(...))) where the inner wrap captured only the inner content but cpp's grammar ctx includes the outer parens.
no_pos	Pre-insert start: null, end: null so the outer wrap leaves the node bare. Used for nodes cpp explicitly doesn't position (NamedArgument, ColumnExprNamedArg).

Two more things to keep in mind:

• Offsets are character indices, not byte indices. cpp's getStartIndex() is char-based; rust's source slices are byte-based. Parser::pos_obj converts via byte_to_char_index for non-ASCII sources, short-circuits for ASCII. If you bypass pos_obj (e.g. hand-building a position object for a node-builder you control), you have to do the conversion yourself.

• Column is character-position-in-line, not byte-position. Same reason. The ASCII fast path in pos_obj handles this for free; the slow path counts chars between line-start and offset.

Known long-tail divergences

The PBT for expr and select exposes adversarial grammar surface that the production corpora never see: extreme WITHIN GROUP (ORDER BY …) shapes and multi-token-AND-merged operands. These take focused per-shape investigation; the PR description has the current numbers. (Deep BETWEEN low AND high chains with embedded aliases and ternaries used to be on this list; the two-tier grammar made BETWEEN bounds value-tier, so those shapes are now either unambiguous or rejected by both backends.)

The production corpora (log_corpus_diagnostic, hog_corpus_diagnostic) stay above 90%, so anything the PBT surfaces that doesn't appear there is technically grammar-parity work but not user-visible.

Selecting from Python

from insights.insightsql.parser import parse_expr, parse_select, parse_program

ast = parse_expr("1 + event.properties.$browser", backend="rust-json")

Backends live in insights/insightsql/constants.InsightsQLParserBackend:

Backend	Use case
cpp-json	Production default. ANTLR-based, oracle for everything below.
rust-json	This crate. ~15× / ~50× faster, behaviour identical (modulo the long tail).
rust-py	This crate, returning insights.insightsql.ast dataclasses directly via PyO3 (no JSON round-trip). Production primary.
cpp-with-rust-shadow	Parses with cpp, shadow-parses with rust, raises on mismatch (TEST) / logs at 1% sample (prod).