We used to track benchmarks with a GitHub action, but it became incredibly slow to use the web page, and also caused clones of the main Ibis repo to be slow due to hosting the page on GH (and thus using the gh-pages branch).
We are currently running the benchmarks in CI and upload data to GCS as compressed Parquet files, but we're not doing anything with that data.
While benchmarking in CI has its downsides, I think there's value in using it to detect very large performance changes, e.g., those above 5x, so I think we should bring back some kind of performance regression testing.
I've recently been playing with https://bencher.dev to see if it's a viable alternative, and have been in contact with the maintainer. It seems promising, at the very least as a way to visualize and store the data without us having to manage it. It also supports alerting based on a few different benchmarking methodologies, but I haven't yet played with that.
We used to track benchmarks with a GitHub action, but it became incredibly slow to use the web page, and also caused clones of the main Ibis repo to be slow due to hosting the page on GH (and thus using the
gh-pages
branch).We are currently running the benchmarks in CI and upload data to GCS as compressed Parquet files, but we're not doing anything with that data.
While benchmarking in CI has its downsides, I think there's value in using it to detect very large performance changes, e.g., those above 5x, so I think we should bring back some kind of performance regression testing.
I've recently been playing with https://bencher.dev to see if it's a viable alternative, and have been in contact with the maintainer. It seems promising, at the very least as a way to visualize and store the data without us having to manage it. It also supports alerting based on a few different benchmarking methodologies, but I haven't yet played with that.
Anyway, I plan to keep this simmering until I get all the existing benchmark data uploaded, which is currently blocked on https://github.com/bencherdev/bencher/issues/460.