typescript-eslint Performance Comparisons

Various performance baselines for typescript-eslint.

Usage

You'll need hyperfine installed locally, such as with brew install hyperfine or winget install hyperfine. See sharkdp/hyperfine#installation.

npm install
npm run generate
npm run measure

To measure a local typescript-eslint build rather than the published one, point TYPESCRIPT_ESLINT_PATH at the checkout.

TYPESCRIPT_ESLINT_PATH=$(realpath ../typescript-eslint) npm run generate
TYPESCRIPT_ESLINT_PATH=$(realpath ../typescript-eslint) npm run measure

Comparison Types

generate and measure take an optional comparison type, defined in comparisons in src/data.ts:

• default: parserOptions.project against parserOptions.projectService, plus the native backend when TYPESCRIPT_ESLINT_PATH is set
• native: the classic project service against the TypeScript 7.1 native backend, across file counts and rule sets

TYPESCRIPT_ESLINT_PATH=$(realpath ../typescript-eslint) npm run generate:native
TYPESCRIPT_ESLINT_PATH=$(realpath ../typescript-eslint) npm run measure:native

The native backend is unpublished, so the native comparison requires TYPESCRIPT_ESLINT_PATH.

You can manually measure individual cases by running hyperfine ../../node_modules/eslint/bin/eslint.js --ignore-failure --warmup 1.

Measured Attributes

Each comparison in src/data.ts can be modified to test:

• files: roughly how many generated files should be linted
  • Files repeat a pattern of two each: super simple, relatively simple, using generics, using result unions, and using all of those plus fancier types
• layout: what rough shape of imports those files exhibit:
  • "even": a single root-level index.ts importing from roughly an even triangle shape of files
  • "references": a single root-level tsconfig.json with project references to a few projects
  • "wide": one root-level index.ts importing from all files in the project
• rules: which rules are enabled:
  • "floating": only @typescript-eslint/no-floating-promises
  • "recommended": recommendedTypeChecked, which queries type information far more often
• singleRun: whether to enable single-run inference as a performance boost
• types: how type information is obtained for typed linting:
  • "project": parserOptions.project
  • "service": parserOptions.projectService
  • "native": parserOptions.projectService with the experimental TypeScript 7.1 native backend

Results

Right now, parserOptions.project outperforms parserOptions.projectService, by roughly 7%. This is a performance issue and we are investigating it as a bug.

┌───────┬───────────────┬───────────────────────┬───────────────────────┐
│ files │ rules         │ project (even layout) │ service (even layout) │
├───────┼───────────────┼───────────────────────┼───────────────────────┤
│ 1024  │ 'recommended' │ '6.895 s ± 0.038 s'   │ '7.356 s ± 0.045 s'   │
└───────┴───────────────┴───────────────────────┴───────────────────────┘

See typescript-eslint/typescript-eslint#9571 Performance: parserOptions.projectService no longer outperforms parserOptions.project in typescript-eslint.

Result Measurement Notes

• Example measurements taken on an M1 Max Mac Studio with Node.js 24.15.0, TypeScript 6.0.3, and typescript-eslint 8.70.0
• Single-run inference is enabled for both types

Comparisons

The comparisons/ directory contains details on more specific comparisons. See each comparisons/*.md file for details on what's being measured.

Traces

The traces/ directory contains more specific traces for investigations.

✨ You might consider using 0x for nice flamegraph visuals.

All comparisons were run on a common shape of linting: 1024 files with the "even" (triangle-shaped) imports layout.

Contributors

Jake Bailey 🤔

Josh Goldberg ✨ 🤔 🚇 🚧 📆 🔧