add twine check to azure CI + 3.8 matrix

[Zethson]

Dear everyone,

This PR adds

1. Python 3.8 to the build matrix. We were currently only testing vs 3.6 and 3.7. 3.10 is already coming up in April, so it is in my opinion time to use the more latest Python versions. If you feel like this adds too much to the Azure/CI bill, then I would suggest to remove 3.6.
2. Solves Add build and twine check CI task #1585

Signed-off-by: Zethson lukas.heumos@posteo.net

[ivirshup]

👍 to twine check 👎 to the python versions

Ultimately, we do have a limited amount of CI, so I think it's important to be a bit cautious adding many jobs. Of the dependencies I'm worried about being an issue: generally not newer python versions. Minimum python versions are important for catching us using newer features.

I am up for swapping python 3.7 with 3.8. I don't think 3.9 is going to work for now. Last time I tried to use 3.9 (a month ago) numpy builds weren't working. I believe numba currently isn't working: numba/numba#6345

Higher priorities to me (roughly in order):

• Test on windows
• Test against lower bounds of our requirements
• Test on Mac
• Dev builds

[Zethson]

@ivirshup perfectly fine with me.
Removed 3.6

So we are now testing against 3.7 and 3.8.

[Zethson]

@ivirshup is there any reason why we are currently not additionally using Github Actions?

[ivirshup]

Removed 3.6

We should keep 3.6 as long as we support it. It's easy to accidentally add features which only work with 3.7+ otherwise. I'd be happy to drop 3.6 once numpy does (and in general roughly follow NEP 29 as soon as the ecosystem does).

is there any reason why we are currently not additionally using Github Actions?

Depends on the task. Also depends on the definition of github actions I think? We aren't using any of their runners for testing because we'd like the ability to integrate with hosted resources on azure. Also, azure seemed like much more of a standard for numeric python packages at the time we chose it.

I'd be happy to have github actions for other things, like precommit. twine check could be another one, but I haven't looked in to how "artifact" type things are handled with github actions to know if we'd be able to recover the built objects.

We'd talked about using codecov too, which I'd like to add a check for. I'm not totally clear on the distinction between checks and actions yet.

[ivirshup]

btw, it'd be great if this got a sibling PR in anndata

[Zethson]

We should keep 3.6 as long as we support it. It's easy to accidentally add features which only work with 3.7+ otherwise. I'd be happy to drop 3.6 once numpy does (and in general roughly follow NEP 29 as soon as the ecosystem does).

All right.
So do you want
3.6 + 3.7 + 3.8
Or
3.6 + 3.7
?

[Zethson]

Regarding Github Actions. It is in my opinion the fastest option out there and jobs spin up really fast.

We could certainly move some of the fast and easy checks to Github Actions and leave the more heavy jobs for Azure.
But let's keep this for a discussion in another issue or Slack.

I'll add a sibling PR to Anndata as soon as we merged this one.

[ivirshup]

I think 3.6 + 3.8 would be reasonable, covering the upper and lower bounds.

[Zethson]

I think 3.6 + 3.8 would be reasonable, covering the upper and lower bounds.

Feels weird to me honestly, but I see what you're trying to do.
Is there no possibility that our Azure credits cover 3.6, 3.7 and 3.8?

If not, then I would be fine with 3.6 and 3.8

[ivirshup]

10 simultaneous jobs is the supposed limit (though I'm not sure how much this is enforced). I think upper and lower bounds are generally a good way to go:

• lower bounds make sure we don't add incompatible features
• upper bounds give us deprecation warnings

[ivirshup]

Just a minor change, otherwise all good.

[ivirshup]

@meeseeksdev backport to 1.7.x