Junk Drawer
For all those little papers scattered across your desk
Junk DrawerFor all those little papers scattered across your desk
I stand for the Constitution, for due process, and for community that takes care of each other.
Brief reflections on implementing Raft in Rust.
I hadn’t thought about consensus algorithms since my MS year in 2020/2021, and I’d certainly never thought deeply about them. Studying the paper and Ousterhout’s video presentations convinced me that implementing Raft ought to be possible, in that same way that hubris is necessary to take on certain challenges. I was demonstrably humbled more than once1, but Raft’s goal (as I understand it) is to be practically implementable correctly.
The details matter immensely for correctness. Each statement is load-bearing.
As I wrote in the public README, there’s little formal validation that my implementation is faithful to any spec or upholds the relevant properties. I have a few unit tests, but I haven’t taken the dive on fuzzing or property testing this code. It deserves it. (I probably won’t try any proof tools, though if memory serves creusot exists to fill that gap.2)
I especially enjoyed explicitly modelling an abstract Clock, which I did for one
of our warmup exercises during the course—it made some reasoning deliciously
simple. Its implementation is not perfect (a sleep loop doesn’t really get you
precise ticking frequency), but a real example could tie into hardware clocks.
This is my most extensive foray into Rust to-date, though I’d dabbled a bit before. I like to say I can read Rust fairly well, having done that for years now and especially due to the writings of Ralf Jung and others. I understand a bit of the ideas Rust builds on, so I didn’t find many ownership challenges while writing this project—I intentionally structured the driver loop and main algorithm so that who owned what data was clear, except for some fiddly bits about sending application results around which end up just cloning.
I still stumble over some parts of Rust’s module system, but that got easier with time.
Rust’s tools are great, of course.
I shied away from async Rust in this project; it seemed like too much to bite off at the time. I’m more comfortable with message-passing threads from some Go (years ago) and Racket (more recent), so I usually think in that model.
In a similar way, I didn’t go dependency hunting here (I don’t browse crates.io for fun). I have serde for serialization, which ended up being more code than I thought I would have to write (integrating with serde-lexpr to use S-expression formats), and rand for randomness. There are likely other things that would be nice to glue in, and for other projects dependencies are a necessity. Here I was mostly satisfied with the standard library3.
I suspect my use of dependencies would look different if Rust’s documentation were integrated in the same way that Racket’s is, but I realize that’s less practical for a bigger ecosystem.
I’m quite pleased I was able to implement such a fascinating system in a few days of hard work: the essence was written between August 26th and 29th, with less than 20 commits afterwards4:
λ g dategraph
2025-08-26 13 ███████████████████
2025-08-27 28 ████████████████████████████████████████
2025-08-28 45 █████████████████████████████████████████████████████████████████
2025-08-29 42 █████████████████████████████████████████████████████████████
2025-09-20 1 █
2026-01-03 1 █
2026-01-04 2 ███
2026-01-05 1 █
2026-01-06 3 ████
2026-01-10 1 █
2026-01-19 3 ████
2026-02-03 4 ██████
λ DATEGRAPH=additions g dategraph
2025-08-26 652 ████████████████████████████████████████████████████████████████
2025-08-27 530 ████████████████████████████████████████████████████
2025-08-28 436 ███████████████████████████████████████████
2025-08-29 409 ████████████████████████████████████████
2025-09-20 11 █
2026-01-03 3
2026-01-04 8 █
2026-01-05 52 █████
2026-01-06 12 █
2026-01-19 45 ████
2026-02-03 204 ████████████████████
I’m also pleased I was able to return months later and reconstruct my thinking without too much hassle. A few details were peculiar, but I’ve already forgotten which ones!
And I can’t resist5:
λ g ls-files | code-percent
RowNumber File PercentTotal CumulativePercentTotal
1 src/lib.rs 30.525 30.525
2 src/test_protocol.rs 26.2913 56.8163
3 src/main.rs 11.3463 68.1626
4 README.md 9.99153 78.1541
5 src/net.rs 5.7155 83.8696
6 src/net/bytes.rs 5.63082 89.5005
7 src/test_append_entries.rs 4.65707 94.1575
8 start-cluster 1.31245 95.47
9 src/test_save_restore.rs 1.10076 96.5707
10 src/bin/client.rs 1.98984 98.5606
11 watch-cluster 0.127011 98.6876
12 start-client 0.127011 98.8146
13 src/net/config.rs 0.550381 99.365
14 .gitignore 0.211685 99.5767
15 Cargo.toml 0.42337 100
Try git log --grep fix. ↩
I have an old incomplete proof of some supposed inverses I’d like to finish, too… alas. ↩
Though not with the un-hashability of mpsc::Senders. ↩