docs: Add a Jujutsu from first principles doc.

This document is WIP. 

This came up recently in Discord as a explanation was requested without referencing Git, which
all our other documents do.
This commit is contained in:
Philip Metzger 2024-02-18 17:48:46 +01:00
parent 205334098a
commit 042d76eb7a
2 changed files with 79 additions and 0 deletions

View File

@ -0,0 +1,77 @@
# Jujutsu from first principles (without Git)
This document describes what Jujutsu's core principles are and explores some
possibilities which a native backend could encompass.
## Preface
Why does Jujutsu exist and which problems does it solve? This document tries to
answer both of these questions while expanding on the design in a user-friendly
way.
At its core Jujutsu is [Version Control System][vcs] which scales to huge
repositories at [Google scale][billion-lines]. Many design choices are
influenced by the concurrent commits happening in Googles Monorepo, as there
are always multiple people working on the same file(s) at the same time.
## Core Tenets
Jujutsu's core tenets are:
* User-friendliness: Making the working copy a commit is simpler. This is
how the project started.
* The "repository", so the commit graph is the source of truth. The working
copy is just one way of editing commits.
* All operations must be able to scale to Google-scale repos (lots of commits
, lots of files): Laziness is important, must avoid accessing data
unnecessarily.
* Having as few states as possible.
* Make it incredibily hard to lose work in your repository.
* Allow concurrent edits on any commit, pending or finished.
* Make a "stacked diffs" workflow as easy as possible.
* Git-interop: Git is everywhere. We need to have good interop to be adapted.
* Pluggable storage: Must be easy to integrate with different commit storage,
virtual file systems and more.
## Base design
The initial base design is to be a conceptually simpler Mercurial, as
automatically snapshotting the working copy simplifies the UX of the
command-line interface by a huge amount and avoids many bad states.
By also choosing to operate by default on the history of the repository (
just called the "the Graph" from now on) instead of files, all history
modifying commands can be done at any point. This is a major improvement on
other version control systems as they need to re-apply a single patch on each
new ancestor before finishing the Graph rewrite. Since the Graph can be changed
at any point, the working copy cannot contain any state depending on it, thus
we have the working-copy commit, which just is another commit from the Graph's
point of view.
### Change-IDs and Changes
Since Jujutsu is oriented around a "stacked diffs" kind of workflow, which
primarily work on individually versioned patch sets, some kind of container is
needed, this is what a Change is. They are provided with a unique id to address
them easily. This mechanism is also customizable so a custom backend could add
a new scheme, which is a major win for tool integrations such as codereview.
And since each change can be addressed individually it simplifies the
commandline.
### Commit Evolution
### Operation store
The operation store is a abstraction for synchronizing multiple clients to a
common state which allows Jujutsu to seamlessly work across multiple
workstations and laptops. And since this part is also customizable, a custom
backend can stream them to all known client devices for a user, which enables
a transparent multi-machine rollback mechanism.
[billion-lines]: https://www.youtube.com/watch?v=W7*TkUbdqE&t=327s
[vcs]: https://en.wikipedia.org/wiki/Version_control

View File

@ -95,6 +95,8 @@ nav:
- FAQ: 'FAQ.md'
- 'Jujutsu from first principles': 'jj_first_principles.md'
- "CLI Reference": 'cli-reference.md'
- Testimonials: 'testimonials.md'