I'm getting a lot of benchmarking alerts for PRs that very clearly should not impact performance (e.g. documentation-only changes such as #1524). I think very likely running time for benchmarks can just vary naturally on different runs (different machines, different loads), so the current approach of caching results to compare later is probably flawed
I'm getting a lot of benchmarking alerts for PRs that very clearly should not impact performance (e.g. documentation-only changes such as #1524). I think very likely running time for benchmarks can just vary naturally on different runs (different machines, different loads), so the current approach of caching results to compare later is probably flawed