Open dswij opened 9 months ago
Yea, I think a problem with that benchmark is that it would most helpful when we can "control" the bandwidth. Currently, since it's a connection of localhost, it can blast all the bytes at full speed (fuller than real world), and then it's just based on the scheduler and other guests on the CI runner.
The benchmark for the http2 adaptive window is inconsistent across runs. This will cause CI to be marked as a failure occasionally whenever there is a drop in benchmark results.
See for example https://github.com/hyperium/hyper/actions/runs/7786695842/job/21232101297.