1. Can cache-downloads: true be set?

Yes we can do it. I'm not sure it's worth it given that the total amount of data that can be cached per repository is 10 GB (https://docs.github.com/en/actions/using-workflows/caching-dependencies-to-speed-up-workflows#usage-limits-and-eviction-policy) and some of the envs are 0.5 GB large.

2. Could you give an example before/after improvement in the setup?

Sure, here we have 2 min -> 0.5 min https://github.com/pandas-dev/pandas/runs/7000499676?check_suite_focus=true https://github.com/pandas-dev/pandas/runs/7007202415?check_suite_focus=true and here we have 4.5 min to 1.5 min https://github.com/pandas-dev/pandas/runs/6998916048?check_suite_focus=true https://github.com/pandas-dev/pandas/runs/7001171440?check_suite_focus=true

In my experience env setup can be considerably slower if there are network problems. For example here we have 8 min, although I'm not sure how much of it is related to using Mamba instead of Micromama https://github.com/pandas-dev/pandas/runs/7028218282?check_suite_focus=true.

1. Can cache-downloads: true be set?

Yes we can do it. I'm not sure it's worth it given that the total amount of data that can be cached per repository is 10 GB (https://docs.github.com/en/actions/using-workflows/caching-dependencies-to-speed-up-workflows#usage-limits-and-eviction-policy) and some of the envs are 0.5 GB large.

I see. I would still be partial in trying it as I've noticed we sometimes still hit connection errors from conda although I extended the timeouts in the condarc.yml. If environments are evicted from the cache during every build due to the size then yeah I can see it not being worth it

Thanks @jonashaag. These CI improvements are awesome!