Remove subdir, platform, arch from "packages", "packages.conda"

[dholth]

Checklist

• I added a descriptive title
• I searched open requests and couldn't find a duplicate

What is the idea?

Some packages contain subdir='linux-64', platform='linux', arch='x86_64' in info/index.json

It would be reasonable to remove these from the output repodata.json, at least if repodata.json["info"]["subdir"] matches.

Why is this needed?

These make repodata.json larger for no reason.

[dholth]

See also https://hackmd.io/Tt09IQJ6SKu28kV3Qw6Otg

Everyone would be happy to see the redundant data go away.

Suggests trying with different clients including older clients.

I noticed a significant fraction of the index packages do not these redundant fields already.

[jaimergp]

From the top of my head:

• conda-build aggregates subdirs together in a single dictionary. We need to make sure that the PackageRecords have subdir properly annotated by then.
• In libmamba 1.x, libsolv parses the repodata items on its own, so we need to make sure that the absence of this field doesn't break anything
• In libmamba 2.x, libmamba does the same, so the same thing as above, but for libmamba
• I hope there's no channel where they are serving everything in a single subdir.

at least if repodata.json["info"]["subdir"] matches

I would argue that either:

• We allow each package to override the global subdir, and leave the repodata_version unchanged (assuming Things Just Work)
• We forbid it altogether, force people to publish per subdir, and bump the repodata_version

[dholth]

@jaimergp I remember we added a check for libmamba or conda if subdirs were aggregated in a single directory. How did we manage that together with conda-build's behavior as well as the single package cache ~/miniconda3/pkgs/?
I do notice that ~/miniconda3/conda-bld/ has per-subdir folders.

[jaimergp]

conda-build's aggregated index is only used by classic. libmamba ignores that altogether and loads directly from the on-disk JSONs. We do analyse the in-memory index to find local channels that should be reloaded after each solver invocation in case there are new packages built.