Some of the integration tests and benchmarks can be quite slow (or slower than CI... Slow enough that we need to wait on them a little when merging, anyway). I believe part of the motivation for breaking tests into testsets in the SciML ecosystem (there are several reasons) is to enable greater parallelisation at a testset granularity, with more runners active.
I don't think the gain would be huge, given that a large portion of runner time is spent setting up and getting a working environment going, but there are still some interesting gains to be had.
Some of the integration tests and benchmarks can be quite slow (or slower than CI... Slow enough that we need to wait on them a little when merging, anyway). I believe part of the motivation for breaking tests into testsets in the SciML ecosystem (there are several reasons) is to enable greater parallelisation at a testset granularity, with more runners active.
I don't think the gain would be huge, given that a large portion of runner time is spent setting up and getting a working environment going, but there are still some interesting gains to be had.