install azcopy step failing

Stream: general

[Bryan Gin-ge Chen]

It's not too uncommon to see mathlib CI fail due to something going wrong with the "install azcopy" step: https://github.com/leanprover-community/mathlib/runs/2070026978

Two questions:

1. Is there anything we can do to make this more robust?

2. Why does the "push release to azure" step still run and successfully push an empty file when "install azcopy" failed?

[Rob Lewis]

I imagine there's a way to run the actions on a Docker image that already has azcopy, elan, etc preinstalled. I've never looked into this and have no idea how much work it is.

[Bryan Gin-ge Chen]

Has this step ever failed consistently across multiple PRs for an extended period of time? Would it be a bad idea / pointless to have the script retry the wget a few times?

[Julian Berman]

you also can just cache the tarball "locally" without a whole container

[Julian Berman]

about to run out so can't write a full example but here's how you do it for e.g. python deps

[Julian Berman]

https://github.com/Julian/til/blob/77d256a6f400e0a9e938aca6694527ae720a6566/.github/workflows/publish.yml#L30-L36 you basically just specify a key (e.g. here a version of that tool) and then GH knows if the version is the same you want the cached version

[Bryan Gin-ge Chen]

Cool! I didn't know about that. Looks like the documentation is here.

[Julian Berman]

Yup that's it, it's quite useful.

[Bryan Gin-ge Chen]

As it turns out, the reason why "push release to azure" still succeeds even when "install azcopy" fails is that azcopy comes pre-installed. #6613 removes this step and also makes a few other changes to the upload logic.

[Scott Morrison]

LGTM. I hope we can merge this quickly, as CI is painful at the moment.

[Bryan Gin-ge Chen]

To be honest, I don't think it would have improved things too much yesterday. There was one staging build that broke due to this "install azcopy" thing, but there were also 2 PRs on the queue which didn't build against master or failed the linter.