- Run comparison benchmark in an isolated git worktree to prevent workspace pollution and local branch conflicts.
- Harden security by passing benchmark target and SHA as environment variables to prevent shell injection, and adding '--' to bazel query and git rev-parse.
- Optimize workflow by removing pnpm caching to mitigate cache poisoning risks.
- Improve robustness of benchmark log parsing, supporting both ZIP outputs and raw directories, and safely checking for JSON reports.
- Centralize git command execution on the dev-infra GitClient for consistency.
- Add tslib to benchpress dependencies to prevent module resolution failures.
This commit creates a new script that solves the following use-cases:
- Running benchmarks. It's not trivial to figure out the benchmark
target names, and it's also easy to mess up the right Bazel flags.
- Performing comparisons. When e.g. working on a runtime senstive
change, it should be trivial to run benchmarks between the current
working stage, and a base revision (e.g. `main`).
The script takes care of both these use-cases and comes with a
prompt-based command line tool experience.
The script will also be used by a future GitHub action that can run
comparisons triggered via GitHub PR comment (by trusted team members).
PR Close#50745
Matthew Beck