Oxidize Radicle-Git

[Byron]

Here is the repository: https://app.radicle.xyz/nodes/seed.radicle.xyz/rad:z6cFWeWpnZNHh9rUW8phgA3b5yGt

Required Features

• gix-protocol::fetch() without Delegates #1634
  • Remove code-duplication in heartwood (see this comment for all the details on where to look; also check the server side of things for reference)

More Analysis TBD

Preferences

• crate git-ref-format - typesafe/statically analysed ref handling
• crate radicle-git-ext for better tools to work with commits

Notes

The repository contains a lot of extra-functionality that also isn't present in git2. This should probably be available from gix as long as it's not Radicle specific.

• Re-implement some basic features that were lacking in git2 in gix
• Use gix where possible, slowly replacing git2
• Remove git2 once there is nothing left.

[FintanH]

This is exciting!

Something I'd like to point out is that replacing git2 functionality with gix wouldn't be my largest concern. While a lot of work, I imagine it would be straight forward.

The heaviest lifting and biggest concern for a first task is the fact that we're still using the Delegate approach for the fetch protocol. I did look at using the more modern approach that Gitoxide was using, however, iirc it didn't provide enough control for us to use the current staged approach to fetching from one remote to the other. If this is something that I could go into more detail about, please let me know.

[Byron]

Thank you for clarifying that. From my experience, using gix in the frontend has advantages in terms of compatibility at the very least, but I also hear that recently libgit2 really ramped up its contributions so these issues might even go away in the mid-term.

And I'd definitely love to finally come up with a fetch-API that is easy to use but not unnecessarily limiting, and can thus work for you as well. If that was the case, I think you could start tracking a more recent version which might ultimately pay off.

If you would link the latest code that uses the Delegate here, I should be able to see what can or can't be done with higher-level APIs and fix these. Sharing what it specifically was that prevented the adoption last time you tried would probably certainly be helpful to me as well.

Thanks again!

[FintanH]

Thank you for clarifying that. From my experience, using gix in the frontend has advantages in terms of compatibility at the very least, but I also hear that recently libgit2 really ramped up its contributions so these issues might even go away in the mid-term.

Compatibility in which sense? :) In my mind, the radicle-fetch code is quite isolated and the only conversion points are OIDs and refnames, which are already in place.

And I'd definitely love to finally come up with a fetch-API that is easy to use but not unnecessarily limiting, and can thus work for you as well. If that was the case, I think you could start tracking a more recent version which might ultimately pay off.

Interesting. I can have a look again because, admittedly, it was a while that I looked and then got swept away by other tasks.

If you would link the latest code that uses the Delegate here, I should be able to see what can or can't be done with higher-level APIs and fix these. Sharing what it specifically was that prevented the adoption last time you tried would probably certainly be helpful to me as well.

I should have taken notes when I explored trying to use the updated version, but unfortunately my foresight is not 20/20 😅

So the main entry point to using the gix fetch code are the following helpers:

• handshake
• ls_refs
• fetch

The ls_refs function calls into the Delegate found here -- a lot of the code is a modified version of the gix code. The same goes for fetch (here).

I think the best way to understand how we use the delegate approach is by describing the high-level flow. The important context is that there are two, special rad references:

• rad/id -- this reference points to a repository's identity which contains important information about the repository, in this case the delegates of the repository. A delegate being an entity that manages the repository and is a trusted peer for the context of the repository.
• rad/sigrefs -- each peer who contributes to the repository has a rad/sigrefs. It points to a signed payload of that peer's reference state, key/value pairs of reference and SHAs, e.g. refs/heads/main deadbeef, refs/cobs/xyz.radicle.patch/1234 beefdead, etc. The pairing of the signature and payload allows the protocol to cryptographically verify a peer's set of refs and gives us the ground truth for that peer.

With this in mind, the protocol is essentially a series of stages where we fetch a set of this data and ensure some data validity. The state is kept in a type called FetchState and each stage is executed via it's method FetchState::run_stage.

The stages are run in the following order:

1. Fetch rad/id -- we fetch the refs/rad/id so that the delegates can be identified and their references are included in the fetch.
2. Fetch rad/sigrefs -- we fetch this reference for each delegate and, depending on our configuration, each peer that we are following.
3. Fetch the contents of rad/sigrefs -- we fetch the references that are listed in each peer's rad/sigrefs. Note that we have the SHAs so we can effectively calculate the wants/haves. Also, in some cases we also have the rad/sigrefs SHA since it can be announced as part of gossip so we can ask for its SHA directly too.

All of this happens over a long-lasting connection, so we only perform the handshake step once. The handshake payload is passed down to each of the ls_refs and fetch calls.

The protocol finishes by signalling to the other end that it's done and it can cease sending upload-packs. The fetcher can then validate all the data it received checking signatures and ensuring that the data is in a consistent state, applying all the updates to the refdb if it's all consistent.

I hope that makes sense, but please let me know if I can elaborate on any points!

[Byron]

Thanks a lot for writing all this down! All this sounds familiar, particularly the multi-step process of downloading packs for different refs. If I remember correctly, the server-side is a plain git server over QUIC transport, which supports git protocol V2. That allows to use a single connection for multiple requests/commands, which are tuned according to the needs, with each stage informing what the next stage can or should do.

My goal here would be to see how I can transform the code away from the Delegate approach to the new command-oriented API, and I think I could consider that successful once the test-suite passes again. This also means that gix probably isn't the right abstraction for now, as it's way too high-level, and I don't know if it should be able to support such a specialised while plumbing exists.

Once successful, this should allow you to track the latest versions of these plumbing crates, and since I do it I would make the API adjustments necessary to support this case as well. This probably also means you don't have to recheck the gix level code of the fetch API, even I don't like to look at it, it's so much and quite complex, always troublesome to find anything 😅.

I probably won't get to it very soon, but it's on my list now and I will make it priority once the last stage of gix status is implemented (HEAD->index diff).

[FintanH]

I should have also linked the upload-pack side of things. Here's the file. It reads and writes from a set of channels, which act as the intermediary for the QUIC connection, afair (@cloudhead might be able to say more to that).

This also means that gix probably isn't the right abstraction for now, as it's way too high-level, and I don't know if it should be able to support such a specialised while plumbing exists.

Aye, that makes sense and resonates with my memory of trying to use gix way back :)

This probably also means you don't have to recheck the gix level code of the fetch API, even I don't like to look at it, it's so much and quite complex, always troublesome to find anything 😅.

Hahaha thank you for saving me that time up-front 😄

I probably won't get to it very soon, but it's on my list now and I will make it priority once the last stage of gix status is implemented (HEAD->index diff).

Sounds good! Ping me if there's anything I can help with when you get around to it.

[cloudhead]

We're no longer using QUIC, we're using a custom framed protocol over TCP, though I don't think it's so relevant since the fetch code just works with generic writers.

[Byron]

I see, thanks for the update!