According to Azure pipeline docs the Mac runners use 3 core CPUs. Looking at the threadpoolctl info, it looks like the docs could be wrong.

Both are using mkl, which may not use all cpus (by default it does not use hyperthreads for instance). I think the most reliable would be to print joblib.cpu_count(only_physical_cores=True/False) to have a good overview of what we are given.

It seems slightly faster with CPU_COUNT=2:

Here is a previous build:

However there is a lot of variability between macOS builds so the improvement might not be significant though.

According to Azure pipeline docs the Mac runners use 3 core CPUs. Looking at the threadpoolctl info, it looks like the docs could be wrong.

Maybe it has changed and they did not keep the doc in sync.

Number of cores: 3
Number of physical cores: 3

Hum actually the number of physical cores is probably incorrect since joblib will return the total number of cores anyway when the number of usable cores is limited. Better to just print cpu_count

Lot of variability indeed.

So indeed there are 3 cores according to joblib. But then the overall runtime does not seem to improve by using 3 pytest-xdist workers. Let's merge this PR as it is?