git-lfs support

[b-camacho]

Hi! This PR works, but there's definitely a better way to support LFS. Looking for maintainer feedback.

Motivation

nix fetches git repos using libgit2, which does not run filters by default. This means LFS-enabled repos can be fetched, but LFS pointer files are not smudged.

This change adds a lfs attribute to fetcher URLs. With lfs=1, when fetching LFS-enabled repos, nix will smudge all the files.

Context

See #10079.
Git Large File Storage lets you track large files directly in git, using git filters. A clean filter runs on your LFS-enrolled files before push, replacing large files with small "pointer files". Upon checkout, a "smudge" filter replaces pointer files with full file contents. When this works correctly, it is not visible to users, which is nice.

Limitations

Right now, for repos with lfs=1, the LFS-tracked files are materialized during nix flake lock - this is bad, I'm looking for a way to avoid this.

Priorities and Process

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[b-camacho]

Small complication:
it seems that nix flake lock calls fetch which in turn calls Input::fetch -> InputScheme::fetch -> fetchToStore.

In other words, I don't currently see a way to:

• materialize LFS files when fetching -source store paths, but
• don't materialize LFS files during nix flake lock.

[L-as]

What use case do you have in mind? Isn't LFS typically for large files, that wouldn't usually affect evaluation anyway?

[b-camacho]

What use case do you have in mind? Isn't LFS typically for large files, that wouldn't usually affect evaluation anyway?

builtins.fetchGit populates the nix store with a <hash>-store path. This path is used as the source when building a derivation. Currently, the builder will see the unsmudged LFS pointer files, but I'd like the builder to optionally see smudged files. I agree that smudged files are usually not needed at eval time, but I don't see a good alternative of making them available at build time, besides a fixed-output derivation (but fixed-output derivations have their own problems)

[L-as]

A FOD seems optimal here, in general you shouldn't use builtins.fetchGit if you're only going to use it at build time.

[b-camacho]

In general I agree, but (afaik) other fetchers can't use git credentials.

[nixos-discourse]

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

https://discourse.nixos.org/t/2024-03-11-nix-team-meeting-132/42960/1

[khoitd1997]

@roberth have you had a chance to take a look at this issue? We have been staying at older versions of Nix as a workaround but newer versions now have fixes for critical issues so sticking to old ones would no longer be optimal.

[roberth]

Hi @b-camacho, thanks for the ping and sorry for the delay. This PR was assigned to me, but I hadn't prioritized it because it was a draft. Wrong assumption on my end, because I do think this is valuable, and I have some things to say :)

when fetching LFS-enabled repos, nix will smudge all the files.

That's a good start, but we need to make sure that the smudging happens in a controlled manner; otherwise we risk adding impurities.

Specifically, we should parse the attribute to check that they're supposed to be unsmudged by lfs; if not, ignore the smudge rule. It seems you were already investigating how this could be implemented.

Furthermore, we should validate the sha256 so that we don't increase the potential for silent errors by a whole external program. The hash should be easy to parse from the pointer file, and while reading other programs' inputs is a little ad hoc, I don't expect any serious issues from this, as we won't cause users to accidentally rely on a bug this way.

---

the LFS-tracked files are materialized during nix flake lock - this is bad

This won't happen unnecessarily either of these are implemented

• Lazy trees #6530
• Lazy fetchTree outPath path values #10252

If we need to backtrack on the removal of narHashes (#6530), we can also avoid re-locking transitive inputs whose lock has already been computed by the dependency's lock.

• Reuse input lock files #7730

So yes, this isn't efficient yet, but it will be.

A FOD seems optimal here, in general you shouldn't use builtins.fetchGit if you're only going to use it at build time.

A fixed output derivation works best when all you're using it for is as an input to another derivation (and it's publicly available, as mentioned).
However, if the "fetched" source is a flake (e.g. you have a flake.nix in a repo with LFS files), then you also need to evaluate files from the fetched source, which would constitute import from derivation, which is not optimal. Furthermore you'd need to produce fixed-output hashes for your local repo files, which is such horrible UX we don't need to consider it as a solution.

---

To summarize, this is worth implementing, I see no blocking issues, design or otherwise, and the following needs to be done:

• lfs attribute with default false, LGTM
• figure out which files are LFS
• invoke the Git LFS filter from $PATH; no need for a rigid dependency or makeWrapper
• check the sha256
• add a test, perhaps extending tests/nixos/fetch-git
• documentation for the lfs attribute (currently under fetchGit's entry in doc/manual/src/language/builtins.md); mention the runtime dependency on the LFS package.

[kip93]

What's the state on this PR? Seems to unfortunately be a bit stale given the delayed review. This issue has been plaguing us for a while, so I'm willing to pick up the torch here and try to get this out the door (was actually starting to see how to fix this myself back in March when I saw this PR and decided to see what came out of this).

[roberth]

@kip93 I think your question was directed towards @b-camacho, but I'd like to add that we would welcome and support anyone who'd like to work on this.

Feel free to ask questions here or in the meetings if you can make them. We generally have some agenda, but we also like to make time for contributors during or after, when we often hang out while we get some things done. Link to the video conference is in the scratchpad linked there. We also have a matrix room, although personally I'm guilty of neglecting that one sometimes.

[b-camacho]

Thanks for the thorough writeup @roberth !
I owe you all an update. To avoid shelling out and implement some features we need to merge, I need a subset of a git-lft C/C++ client. I reimplemented one from Python into C++ here https://github.com/b-camacho/git-lfs-fetch-cpp.

Once I add some tests and integrate git-lfs-fetch-cpp here, we should be ready for another review!

I'm still on vacation with not-great internet, but back in 6 days and will update you all on 7/31 regardless.

Thanks for the feedback and sorry for the wait!

[roberth]

Oh, I don't think shelling out was such a big deal because we can verify the correctness of the result, kind of like how fixed output derivations are allowed to do "grossly impure" things because we can verify the output.

I guess a library implementation of it is still nice for a consistent UX with a small closure size though.

[L-as]

Really we need to reimplement the subset of git we use too.

[roberth]

To a very large degree we've done so. Since #9240 we're using libgit2 so that we use significantly fewer git behaviors by default, such as the overly open ended smudge filters and export-* attributes. We do still use the git command for fetching, which is essentially risk-free as far as reproducibility problems go. It doesn't really matter how the blobs and trees and commits arrive because we can check them thanks to them being content addressed.

[L-as]

We however notably only want shallow fetching (not --depth, but --filter=tree:0). Git doesn't do that well.