Skip solver at install time

[jaimergp]

Description

Comes from and closes #541 (read the issue for more details)

I checked using just --no-deps as an install flag, and while the docs hint that it skips the dependency checks... what it actually does is to perform a full solve, and then trim the dependencies from the graph.

So instead I am opting for a feature already present in conda: @EXPLICIT installs using the remote URLs. Since the repodata and channel cache is mocked already to make the previous method work, we get almost everything for free.

Basically, constructor invokes conda install --offline --file <file>, with this file:

tzdata==2022a=hda174b7_0
ca-certificates==2022.07.19=hca03da5_0
libffi==3.4.2=hc377ac9_4
openssl==1.1.1q=h1a28f6b_0
zlib==1.2.12=h5a0b063_2
setuptools==63.4.1=py310hca03da5_0
certifi==2022.6.15=py310hca03da5_0
sqlite==3.39.2=h1058600_0
pip==22.1.2=py310hca03da5_0
ncurses==6.3=h1a28f6b_3
readline==8.1.2=h1a28f6b_1
python==3.10.4=hbdb9e5c_0
xz==5.2.5=h1a28f6b_1
wheel==0.37.1=pyhd3eb1b0_0
tk==8.6.12=hb8d0fd4_0
libcxx==12.0.0=hf6beb65_1
bzip2==1.0.8=h620ffc9_4

What we are doing in this PR is changing the file syntax to:

# platform: osx-arm64
@EXPLICIT
https://conda.anaconda.org/pkgs/main/noarch/tzdata-2022a-hda174b7_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/ca-certificates-2022.07.19-hca03da5_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/libffi-3.4.2-hc377ac9_4.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/openssl-1.1.1q-h1a28f6b_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/zlib-1.2.12-h5a0b063_2.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/setuptools-63.4.1-py310hca03da5_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/certifi-2022.6.15-py310hca03da5_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/sqlite-3.39.2-h1058600_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/pip-22.1.2-py310hca03da5_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/ncurses-6.3-h1a28f6b_3.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/readline-8.1.2-h1a28f6b_1.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/python-3.10.4-hbdb9e5c_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/xz-5.2.5-h1a28f6b_1.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/noarch/wheel-0.37.1-pyhd3eb1b0_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/tk-8.6.12-hb8d0fd4_0.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/libcxx-12.0.0-hf6beb65_1.tar.bz2#<md5hash>
https://conda.anaconda.org/pkgs/main/osx-arm64/bzip2-1.0.8-h620ffc9_4.tar.bz2#<md5hash>

This still goes through the transaction API, and the records are sorted in topological order, so we should be good.

Checklist - did you ...

• Add a file to the news directory (using the template) for the next release's release notes?
• Add / update necessary tests?
• Add / update outdated documentation?

[hmaarrfk]

Does this allow users to update their environments in the future?

[jaimergp]

It's like installing a conda-locked environment from scratch.

With the exclude option enabled, the environment will be deemed "inconsistent" (same as when you do conda remove --force <pkg>) and a further operation on the conda environment might bring it back, though.

[hoechenberger]

LGTM! Very clever approach

[hmaarrfk]

Do you think this will solve issues with users having an existing .condarc file that specifies different channels? We've been having lots of complaints at Miniforge regarding old installations stopping new installs due to lingering .condarc files

[jaimergp]

Yes, it will alleviate solver-derived problems because this solution bypasses it entirely. It might still suffer from other configuration keys, but I can't think of any examples right now.

We do need better condarc isolation for conda-standalone in general, but that's separate issue (#542) we need to solve at the conda level first (my idea is to expose the CONDARC_SEARCH_PATH constant as a context key).

[jaimergp]

Any other comments @conda/constructor? I will merge later this week.

[hmaarrfk]

this is interesting as a use case.

[jaimergp]

Thanks everybody!

[EisW]

The generated pkgs/envs.txt does not honour info['channel_remap'], so the generated installer fails for locally built channels, by reclaiming the path of the package from the environment running constructor.

The problem seems in preconda.py.

Files pkgs/urls.txt and pkgs/urls are written using get_final_url(info, url) (within all_final_urls_md5s --> this honours channel_remap; but pkgs/envs_txt uses info["_urls"] directly.

You have the remap info already in final_urls_md5s, so the solution -hopefully- is as simple as:

Replace

    # base environment file used with conda install --file
    # (list of specs/dists to install)
    write_env_txt(info, dst_dir, info["_urls"])

with

    # base environment file used with conda install --file
    # (list of specs/dists to install)
    write_env_txt(info, dst_dir, final_urls_md5s)

Equal change is required for sub environment:

replace

        write_env_txt(info, env_dst_dir, env_info["_urls"])

with

        write_env_txt(info, env_dst_dir, env_urls_md5)

==> generated installer than works in my cases

[EisW]

Is it possible to ping #656?
Meanwhile I signed the CLA, but do not know the next step ...