Activate Python 3.9

[joshmoore]

In an attempt to track down #858
Closes #629

TODO:

• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

[jakirkham]

Thanks for starting this Josh 😄

Probably need to bump or relax the numcodecs requirement on CI

[joshmoore]

Do you know the reasoning behind the 0.6.4 pins, @jakirkham ?

[joshmoore]

Ok. So the codecov failure is the same as #811 I assume if you have any ideas, @jakirkham.

I'll fix the py39/numpy1.17 filtration problem (though maybe I just bump to numpy 1.18 ...)

[jakirkham]

Do you know the reasoning behind the 0.6.4 pins, @jakirkham ?

My guess is Matthias had some issues on CI and pinned it to fix them. Not really sure. PR ( #674 ) shows they have been here since adding GH Actions. Dropping them makes sense.

Ok. So the codecov failure is the same as #811 I assume if you have any ideas

No new thoughts.

[joshmoore]

No new thoughts.

Looking a little closer, it looks like self.nblocks always equals 1 for PartialReadBuffer. cc @andrewfulton9
in case that raises any flags.

[joshmoore]

Ok. Now down to just the codecov issue. The next clue in debugging is that:

        nbytes, self.cbytes, blocksize = cbuffer_sizes(header)

appears to now return matching values for nbytes and blocksize with the numcodecs update. That feels significant. Not sure if related to zarr-developers/numcodecs#283

[andrewfulton9]

@joshmoore, sorry I missed my mention earlier. It looks like something in has changed to cause nbytes to always be the same as blocksize. It definitely could be zarr-developers/numcodecs#283. I'll have to look closer to be sure though.

[jakirkham]

Oh interesting, meaning that Blosc may have changed something, which is affecting us here?

[andrewfulton9]

It looks like maybe Blosc has changed the default to just use 1 block if blocksize isn't defined so nbytes will equal blocksize by default, where previously it used to use some default (or tried to guess an optimum) blocksize. In TestArrayWithFSStoreNested I just used the default block object so that is probably why nbytes always equals blocksize now. I can make a PR defining a blocksize there to see if that fixes it.

[jakirkham]

Think that would be helpful. We'd like to keep the partial read functionality in good working order 🙂

[joshmoore]

With #867 merged, this is now green. Merging (but #858/debian-failure) is still a mystery.

[jakirkham]

Thanks Josh! 😀