Open julienguy opened 6 years ago
+1 for isolating matplotlib imports, but unfortunately this appears to be coming from healpy not our code.
We might be able to do something in our desiconda installs to pre-build the font cache so that user code doesn't end up having N>>1 parallel processes all trying to build it and running into each other. Just guessing there.
Another hack would be to have MPI rank 0 import matplotlib followed by an MPI barrier, before proceeding with importing healpy. Ugh, but I think that would at least isolate any font cache building stuff to a single process.
For docker images, I do pre-build the font cache (and also the astropy config stuff). I could easily do that for regular desiconda installs too.
I am getting today the following matplotlib error when running
desi_pipe_run_mpi --first redshift --last redshift
. We should avoid this, because we do not need to produce plots in the pipeline run. Importing matplotlib in the pipeline is introducing an unnecessary fragility. We need it for QA, but this should be independent jobs.