Turn off Git smugding in fetchGit

[fricklerhandwerk]

Describe the bug

As discussed with @nixos/nix-team:

While it's been around for a long time, smudging being enabled in fetchGit can be considered a bug.

More detailed discussion: #4635 (comment)

Steps To Reproduce

$ nix build git+ssh://[email protected]/tomberek/export-magic -L
text> /nix/store/8fdx2h960yf292czd4ik2f6fza63wlqh-.version
text> HEAD -> main
text>  (HEAD -> main)
text> f0fe26485056d92e78d84de3daf14a8875548783
text> Tue, 21 Feb 2023 19:51:28 -0500
text> 2023-02-21 19:51:28 -0500
text> 2023-02-21T19:51:28-05:00
text> HEAD -> main
text>  (HEAD -> main)
text> f0fe26485056d92e78d84de3daf14a8875548783
text> Tue, 21 Feb 2023 19:51:28 -0500
text> 2023-02-21 19:51:28 -0500
text> 2023-02-21T19:51:28-05:00

Expected behavior

It should output the unchanged contents of https://github.com/tomberek/export-magic

Proposal

• resolve commit hashes to tree hashes, this will incur a ~one-time overhead to fetch metadata; a trade-off towards correctness
  • GitHub tarballs are pre-smudged, but fetching those is an optimisation anyway
  • the new semantics will be much simpler

Priorities

Add 👍 to issues you find important.

[nixos-discourse]

This issue has been mentioned on NixOS Discourse. There might be relevant details there:

https://discourse.nixos.org/t/2023-11-06-nix-team-meeting-minutes-101/35247/1

[roberth]

We've recently used the term smudging to cover all behaviors that make file system objects diverge from the simple, well defined, and therefore well reproducible git tree data model.

The example in this issue is technically about export-ignore and not the git smudge filter feature. Submodules should also be considered, as they are a layer on top of git trees.

A more extensive explanation and discussion centered around the smudge filter feature can be found here #4635 (comment)

Essentially what we're trying to achieve is NAR hash-levels of reproducibility by other means, specifically a definitive translation from git objects to Nix file system objects.

Definitive is the keyword here. We won't have the luxury of changing this translation after stabilization, except for the addition of parameters that tweak the translation in ways such that their default value is aforementioned "definitive" behavior.

[fricklerhandwerk]

Essentially what we're trying to achieve is NAR hash-levels of reproducibility by other means, specifically a definitive translation from git objects to Nix file system objects.

Should we open a tracking issue from your comment then?

[thufschmitt]

This should hopefully be fixed by #9240

[roberth]

• See fetchGit/fetchTree reproducibility problems #5313
  • start of recent messages

[roberth]

Duplicate