Code style

OpenPerf rules for code and tests, on top of the shared rules in AGENTS.md.

Code

Write code for cases that occur. No fallbacks, duplicate checks or values without a consumer.
Make each decision in one place: the layer that owns it. Others rely on it.
An arrow function whose body doesn’t fit on one line gets a block with return. Components returning JSX are the exception.
Document what a published package exports from its entry points, such as @openperf/sdk/node, only where the name and type don’t tell the whole story: defaults, limits, value formats, errors and when it does nothing. Skip a JSDoc that restates the name. Write one or two plain sentences, state each fact once where it applies, put defaults in @default and errors in @throws. Types, interfaces and classes get no JSDoc; document their properties instead. Internal code stays without comments.

Test logic that can break on its own: branching, parsing, edge cases, security checks. Don’t test code that only restates itself, such as templates, lookup tables or config.
A test that needs live stores is *.integration.test.ts. Everything else is *.test.ts with mocks.

One statement per file, and every statement can run twice (IF NOT EXISTS, no unguarded backfills): dbmate may run a file again after a partial failure.
Retention is a TTL in the table DDL; change it with ALTER TABLE … MODIFY TTL in a new migration.

Lists, summaries and charts read hourly rollups through buildRollupSource: full hours come from the rollup, the edges of the window from the raw table. Raw tables answer drill-downs by a selective key: a trace, a fingerprint, a log page, the hours a rollup picked for samples.
Each rollup in telemetry/libs/rollups mirrors its materialized view. rollups.integration.test.ts fails when the two drift.
Keep aggregate states bounded: quantilesTDigest for durations, uniqCombined(12) for unique counts.
Measure a read change with bun run bench --project <id> --from <iso> --to <iso> on a dense dataset. It reports time, rows, bytes and peak memory per scenario with caches off.