Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
These live in the repository and map email addresses and names to canonical ones, as described in [`gitmailmap(5)`]. [`gitmailmap(5)`]: https://git-scm.com/docs/gitmailmap We introduce the notion of raw signatures, representing the name and email baked in to an object’s metadata, and canonical signatures, representing the up‐to‐date identity information obtained by looking up a raw signature in a `.mailmap` file. When there are no matching entries, the raw and canonical signatures are the same. Canonical signatures should be used for the majority of purposes, such as display and querying in user interfaces and automated batch processing (e.g., to collate statistics by commit author, or email committers in batch). Generally speaking, whenever you care about *who made the commit* rather than *what data happens to be encoded in the commit itself*, they are the appropriate thing to work with. Raw signatures should usually not be surfaced to users by default unless explicitly asked for. Valid reasons to work with them include low‐level processing of commits where a `.mailmap` is not accessible or would be inappropriate to use (e.g., rewriting or duplication of commits without intent to alter metadata), automated testing, forensics (examining the raw object data stored in the backend that is used to compute a commit’s cryptographic hash), and analysis and debugging of `.mailmap` files themselves. The only signatures Jujutsu currently processes are commit authors and committers, which can be obtained in raw and canonical form with `Commit::{author,committer}_raw` and `Mailmap::{author,committer}` respectively. If Jujutsu starts to store or process immutable identity data in other contexts (e.g. support for additional metadata on commits like Git’s `Co-authored-by`/`Signed-off-by`/`Reviewed-by` trailers, or detached metadata that nonetheless must remain immutable), then the notion of raw and canonical signatures will carry over to those and the same guidelines about preferring to work with and display canonical signatures whenever reasonable will apply. This is not meant to be a comprehensive solution to identity management or obsolete the discussion in jj-vcs#2957. There are many possible designs of forward‐thinking author and committer identity systems that would be a lot better than `.mailmap` files, but I don’t really want to get lost in the weeds trying to solve an open research problem here. Instead, this is just an acknowledgement that any system that treats user names and emails as immutable (as Jujutsu currently does) is going to need a mapping layer to keep them updated, and both Git and Mercurial adopted `.mailmap` files, meaning they are already in wide use to address this problem. All sufficiently large open source repositories tend to grow a substantial `.mailmap` file, e.g. [Linux], [Rust], [curl], [Mesa], [Node.js], and [Git] itself. Currently, people working on these repositories with Jujutsu see and search outdated and inconsistent authorship information that contradicts what Git queries and outputs, which is at the very least somewhere between confusing and unhelpful. Even if we had a perfect orthogonal solution in the native backend, as long as we support working on Git repositories it’s a compatibility‐relevant feature. [Linux]: https://github.com/torvalds/linux/blob/f2661062f16b2de5d7b6a5c42a9a5c96326b8454/.mailmap [Rust]: https://github.com/rust-lang/rust/blob/2c243d957008f5909f7a4af19e486ea8a3814be7/.mailmap [curl]: https://github.com/curl/curl/blob/a7ec6a76abf5e29fb3f951a09d429ce5fbff250f/.mailmap [Mesa]: https://gitlab.freedesktop.org/mesa/mesa/-/blob/cdf3228f88361410175c338704908ea74dc7b8ae/.mailmap [Node.js]: https://github.com/nodejs/node/blob/4c730aed7f825af1691740663d599e9de5958f89/.mailmap [Git]: https://github.com/git/git/blob/9005149a4a77e2d3409c6127bf4fd1a0893c3495/.mailmap That said, this is not exclusive to the Git backend. The `.mailmap` name and format is perfectly generic, already shared between Git and Mercurial, and applies to all systems that bake names and emails into commits, including the current local backend. The code uses Gitoxide, but only as a convenient implementation of the file format; in a hypothetical world where the Git backend was removed without Jujutsu changing its notion of commit signatures, `gix-mailmap` could be used standalone, or replaced with a bespoke implementation. I discussed this on the Discord server and we seemed to arrive at a consensus that this would be a good feature to have for Git compatibility and as a pragmatic stop‐gap measure for the larger identity management problem, and that I should have a crack at implementing it to see how complex it would be. Happily, it turned out to be pretty simple! No major plumbing of state is required as the users of the template and revset engines already have the working copy commit close to hand to support displaying and matching `@`; I think this should be more lightweight (but admittedly less powerful) than the commit rewriting approach @arxanas floated on Discord. ## Notes on various design decisions * The `.mailmap` file is read from the working copy commit of the current workspace. This is roughly equivalent to Git reading from `$GIT_WORK_TREE/.mailmap`, or `HEAD:.mailmap` in bare repositories, and seems like the best fit for Jujutsu’s model. I briefly looked into reading it from the actual on‐disk working copy, but it seemed a lot more complicated and I’m not sure if there’s any point. I didn’t add support for Git’s `mailmap.file` and `mailmap.blob` configuration options; unlike ignores, I don’t think I’ve ever seen this feature used other than directly in a repository, and `mailmap.blob` seems to mostly be there to keep it working in bare repositories. I can imagine something like a managed corporate multi‐repo environment with a globally‐shared `mailmap.file` so if people feel like this is important to keep consistency with I can look into implementing it. But genuinely I’ve never personally seen anybody use this. * The `author`/`committer` DSL functions respect the `.mailmap`, with `*_raw` variants to ignore it. If there’s a `.mailmap` available, signatures should be mapped through it unless there’s a specific reason not to; this matches Git’s behaviour and is the main thing that makes this feature worthwhile. There is a corresponding breaking change of the external Rust API, but hopefully the new method name and documentation will nudge people towards doing the right thing. I was initially considering a keyword argument to the template and revset functions to specify whether to map or not (and even implemented keyword arguments for template functions), but I decided it was probably overkill and settled on the current separate functions. A suggestion from Discord was to add a method on signatures to the template language, e.g. `.canonical()` or `.mailmap()`. While this seems elegant to me, I still wanted the short, simple construction to be right by default, and I couldn’t think of any immediate uses outside of `.author()` and `.committer()`. If this is added later, we will still get the elegant property that `commit.{author,committer}()` is short for `commit.{author,committer}_raw().canonical()`. * The mapping to canonical signatures is one‐way, and queries only match on the canonical form. This is the same behaviour as Git. The alternative would be to consider the mapped signatures as an equivalence set and allow a query for any member to match all of them, but this would contradict what is actually displayed for the commits, require a more complicated implementation (to support patterns, we would basically have to map signatures back to every possible form and run the pattern against each of them), and the `*_raw` functions may be more useful in such a case anyway. * There’s currently no real caching or optimization here. The `.mailmap` file is materialized and parsed whenever a template or revset context is initialized (although it’s still O(1), not parsing it for every processed commit), and `gix-mailmap` does a binary search to resolve signatures. I couldn’t measure any kind of substantial performance hit here, maybe 1‐3% percent on some `jj log` microbenchmarks, but it could just be noise; a couple times it was actually faster.
- Loading branch information
