From 4a72c2ea7b29e7ba87fd298a38428e38075b1884 Mon Sep 17 00:00:00 2001 From: Michael Zhang Date: Tue, 10 Sep 2024 17:59:54 -0500 Subject: [PATCH] pijul post --- .../index.md | 2 +- .../posts/2024-09-10-lambda-calculus/index.md | 2 +- .../2024-09-10-whats-cool-about-pijul.md | 64 +++++++++++++++++++ 3 files changed, 66 insertions(+), 2 deletions(-) create mode 100644 src/content/posts/2024-09-10-whats-cool-about-pijul.md diff --git a/src/content/posts/2024-06-21-coping-with-refactoring/index.md b/src/content/posts/2024-06-21-coping-with-refactoring/index.md index 07cfc41..ebed6d2 100644 --- a/src/content/posts/2024-06-21-coping-with-refactoring/index.md +++ b/src/content/posts/2024-06-21-coping-with-refactoring/index.md @@ -2,7 +2,7 @@ title: Coping with refactoring date: 2024-06-21T05:56:50.594Z tags: - - software-engineering + - engineering heroImage: ./ruinsHero.png heroAlt: ruins diff --git a/src/content/posts/2024-09-10-lambda-calculus/index.md b/src/content/posts/2024-09-10-lambda-calculus/index.md index 776c2f6..a518e2e 100644 --- a/src/content/posts/2024-09-10-lambda-calculus/index.md +++ b/src/content/posts/2024-09-10-lambda-calculus/index.md @@ -124,7 +124,7 @@ This says that there are only 3 possible terms: 3. $t(t)$ which is just function application, one term calling another All of the terms above, including $\mathsf{plus}$, can be represented as a -combination of these constructors. +combination of these constructors. Try it for yourself! Great, but we can't really do anything if all we know is how to write down some terms. We need some rules to tell us how to evaluate it. Something like this: diff --git a/src/content/posts/2024-09-10-whats-cool-about-pijul.md b/src/content/posts/2024-09-10-whats-cool-about-pijul.md new file mode 100644 index 0000000..93a59f4 --- /dev/null +++ b/src/content/posts/2024-09-10-whats-cool-about-pijul.md @@ -0,0 +1,64 @@ +--- +title: What's cool about Pijul? +date: 2024-09-10T22:37:03.684Z +tags: [engineering] +draft: true +--- + +Version control is one of those things that few of us hardly ever think about +until it causes problems. Lots of projects claim to be able to cover up the +internals and expose a glossy interface but once you really start using it you +notice all the pitfalls. + +In particular, there seems to be this never-ending battle between the merge camp +and the rebase camp. Every project I join seems to have different opinions about +the process. + +First of all, both have fundamental weaknesses. In the **merge** side, merge +commits lose information about the merge. Essentially a merge is two separate pieces of information: + +- The commits that came before it +- The new final snapshot of the code + +You can't commute changes made during a merge with different change, because the merge is simply just a new state with no origin information. Also, if some conflicts are missed during the resolution of the merge, they just go through. + +On the other hand, the problem with **rebase** is that it fundamentally changes +history, which is a _huge_ problem. On your own feature branches, this might be +ok, but if you are actively working with someone and need to rebase against +main, you're essentially changing their work. + +--- + +My understand is that the fundamental differences between [Pijul] (and Darcs, which it's based on) and git are: + +[Pijul]: https://pijul.org/ + +1. Conflicts are a known state +1. The patches are not ordered into a merkle tree + +At least these are the differences I care about. You'll note that the first one fixes the merge issue and the second one fixes the rebase issue. + +With conflicts being recorded as a known state in the tree, we can add another commit on top that "fixes" the conflict. Since conflicts are deterministic, and each file's conflicts are recorded, you never have the problem of glossing over conflicts by accident. + +Also, since patches don't depend on "previous" patches' hashes, this means +changing the order in which it lies in history doesn't need to modify the patch +itself, unlike in Git where moving a commit back or forth will modify the entire +future of the tree. Instead, dependencies are just tracked as metadata alongside the contents of the patch itself. + +--- + +With all this said, what's _not_ cool about Pijul? Despite being close to a 1.0, +the Pijul software itself is still heavily lacking in documentation. For +example, the on-disk format and the push protocol are not specified in the +documentation. Also, the only real available forge is the proprietary Nest, +operated by the Pijul developers themselves. Nest was also incredibly +unintuitive and hard to use. + +It's quite a shame because the fundamental structure of the underlying data does +impact the surface level operations of the tool quite a bit. Git has a lot of +clunky corners as we've discovered over the years, but it's so hard to work +around it because of the massive engineering effort involved. + +At some point I'd love to pick apart the tool and see how it works and try to +write up some tooling, but I'm still sinking from SIFT work and PhD applications +right now, so I'll just stick to the sidelines.