When creating a cluster from a software environment that's missing distributed, you just get an error like cluster unexpectedly STOPPED. Going to the logs on the Coiled dashboard, I'm able to find
/opt/conda/envs/coiled/bin/python: Error while finding module specification for 'distributed.cli.dask_spec' (ModuleNotFoundError: No module named 'distributed')
But it would be nice to try-except that and display a more informative error on the client side.
I suppose it would be even nicer to catch at environment creation time—but perhaps there's a use-case for creating an environment that doesn't include distributed, only for local use?
Also, it's probably worth noting in the docs (and, better, via a warning at environment creation time) that bokeh is not included automatically either, nor is it a dependency of distributed (or dask[distributed], as recommended), so it should also be included explicitly. Seeing this and having to recreate everything was disappointing:
Dask needs bokeh >= 0.13.0 for the dashboard.
Install with conda: conda install bokeh>=0.13.0
Install with pip: pip install bokeh>=0.13.0
When creating a cluster from a software environment that's missing
distributed
, you just get an error likecluster unexpectedly STOPPED
. Going to the logs on the Coiled dashboard, I'm able to findBut it would be nice to try-except that and display a more informative error on the client side.
I suppose it would be even nicer to catch at environment creation time—but perhaps there's a use-case for creating an environment that doesn't include
distributed
, only for local use?Also, it's probably worth noting in the docs (and, better, via a warning at environment creation time) that
bokeh
is not included automatically either, nor is it a dependency ofdistributed
(ordask[distributed]
, as recommended), so it should also be included explicitly. Seeing this and having to recreate everything was disappointing: