Measuring, analyzing, and optimizing loops using Linux perf, Top-Down Microarchitectural Analysis, and the CPU’s micro-op cache

.title {text-wrap:balance;} #content > p:first-child {text-wrap:balance;} If Git had a nemesis, it’d be large files. Large files bloat Git’s storage, slow down git clone, and wreak havoc on Git forges. In 2015, GitHub released Git LFS—a Git extension that hacked around problems with large files. But Git LFS added new complications and storage costs. Meanwhile, the Git project has been quietly working on large files. And while LFS ain’t dead yet, the latest Git release shows the path towards a future where LFS is, finally, obsolete. What you can do today: replace Git LFS with Git partial clone Git LFS works by storing large files outside your repo. When you clone a project via LFS, you get the repo’s history and small files, but skip large files. Instead, Git LFS downloads only the large files you need for your working copy. In 2017, the Git project introduced partial clones that provide the same benefits as Git LFS: Partial clone allows us to avoid downloading [large binary assets] in advance during clone and fetch operations and thereby reduce download times and disk usage. – Partial Clone Design Notes, git-scm.com Git’s partial clone and LFS both make for: Small checkouts – On clone, you get the latest copy of big files instead of every copy. Fast clones – Because you avoid downloading large files, each clone is fast. Quick setup – Unlike shallow clones, you get the entire history of the project—you can get to work right away. What is a partial clone? A Git partial clone is a clone with a --filter. For example, to avoid downloading files bigger than 100KB, you’d use: git clone --filter='blobs:size=100k' <repo> Later, Git will lazily download any files over 100KB you need for your checkout. By default, if I git clone a repo with many revisions of a noisome 25 MB PNG file, then cloning is slow and the checkout is obnoxiously large: $ time git clone https://github.com/thcipriani/noise-over-git Cloning into '/tmp/noise-over-git'... ... Receiving objects: 100% (153/153), 1.19 GiB real 3m49.052s Almost four minutes to check out a single 25MB file! $ du --max-depth=0 --human-readable noise-over-git/. 1.3G noise-over-git/. $ ^ 🤬 And 50 revisions of that single 25MB file eat 1.3GB of space. But a partial clone side-steps these problems: $ git config --global alias.pclone 'clone --filter=blob:limit=100k' $ time git pclone https://github.com/thcipriani/noise-over-git Cloning into '/tmp/noise-over-git'... ... Receiving objects: 100% (1/1), 24.03 MiB real 0m6.132s $ du --max-depth=0 --human-readable noise-over-git/. 49M noise-over-git/ $ ^ 😻 (the same size as a git lfs checkout) My filter made cloning 97% faster (3m 49s → 6s), and it reduced my checkout size by 96% (1.3GB → 49M)! But there are still some caveats here. If you run a command that needs data you filtered out, Git will need to make a trip to the server to get it. So, commands like git diff, git blame, and git checkout will require a trip to your Git host to run. But, for large files, this is the same behavior as Git LFS. Plus, I can’t remember the last time I ran git blame on a PNG 🙃. Why go to the trouble? What’s wrong with Git LFS? Git LFS foists Git’s problems with large files onto users. And the problems are significant: 🖕 High vendor lock-in – When GitHub wrote Git LFS, the other large file systems—Git Fat, Git Annex, and Git Media—were agnostic about the server-side. But GitHub locked users to their proprietary server implementation and charged folks to use it.1 💸 Costly – GitHub won because it let users host repositories for free. But Git LFS started as a paid product. Nowadays, there’s a free tier, but you’re dependent on the whims of GitHub to set pricing. Today, a 50GB repo on GitHub will cost $40/year for storage. In contrast, storing 50GB on Amazon’s S3 standard storage is $13/year. 😰 Hard to undo – Once you’ve moved to Git LFS, it’s impossible to undo the move without rewriting history. 🌀 Ongoing set-up costs – All your collaborators need to install Git LFS. Without Git LFS installed, your collaborators will get confusing, metadata-filled text files instead of the large files they expect. The future: Git large object promisors Large files create problems for Git forges, too. GitHub and GitLab put limits on file size2 because big files cost more money to host. Git LFS keeps server-side costs low by offloading large files to CDNs. But the Git project has a new solution. Earlier this year, Git merged a new feature: large object promisers. Large object promisors aim to provide the same server-side benefits as LFS, minus the hassle to users. This effort aims to especially improve things on the server side, and especially for large blobs that are already compressed in a binary format. This effort aims to provide an alternative to Git LFS – Large Object Promisors, git-scm.com What is a large object promisor? Large object promisors are special Git remotes that only house large files. In the bright, shiny future, large object promisors will work like this: You push a large file to your Git host. In the background, your Git host offloads that large file to a large object promisor. When you clone, the Git host tells your Git client about the promisor. Your client will clone from the Git host, and automagically nab large files from the promisor remote. But we’re still a ways off from that bright, shiny future. Git large object promisors are still a work in progress. Pieces of large object promisors merged to Git in March of 2025. But there’s more to do and open questions yet to answer. And so, for today, you’re stuck with Git LFS for giant files. But once large object promisors see broad adoption, maybe GitHub will let you push files bigger than 100MB. The future of large files in Git is Git. The Git project is thinking hard about large files, so you don’t have to. Today, we’re stuck with Git LFS. But soon, the only obstacle for large files in Git will be your half-remembered, ominous hunch that it’s a bad idea to stow your MP3 library in Git. Edited by Refactoring English Later, other Git forges made their own LFS servers. Today, you can push to multiple Git forges or use an LFS transfer agent, but all this makes set up harder for contributors. You’re pretty much locked-in unless you put in extra effort to get unlocked.↩︎ File size limits: 100MB for GitHub, 100MB for GitLab.com↩︎

Conrad Irwin has an article on the Zed blog “Why LLMs Can't Really Build Software”. He says it boils down to: the distinguishing factor of effective engineers is their ability to build and maintain clear mental models We do this by: Building a mental model of what you want to do Building a mental model of what the code does Reducing the difference between the two It’s kind of an interesting observation about how we (as humans) problem solve vs. how we use LLMs to problem solve: With LLMs, you stuff more and more information into context until it (hopefully) has enough to generate a solution. With your brain, you tweak, revise, or simplify your mental model more and more until the solution presents itself. One adds information — complexity you might even say — to solve a problem. The other eliminates it. You know what that sort of makes me think of? NPM driven development. Solving problems with LLMs is like solving front-end problems with NPM: the “solution” comes through installing more and more things — adding more and more context, i.e. more and more packages. LLM: Problem? Add more context. NPM: Problem? There’s a package for that. Contrast that with a solution that comes through simplification. You don’t add more context. You simplify your mental model so you need less to solve a problem — if you solve it at all, perhaps you eliminate the problem entirely! Rather than install another package to fix what ails you, you simplify your mental model which often eliminates the problem you had in the first place; thus eliminating the need to solve any problem at all, or to add any additional context or complexity (or dependency). As I’m typing this, I’m thinking of that image of the evolution of the Raptor engine, where it evolved in simplicity: This stands in contrast to my working with LLMs, which often wants more and more context from me to get to a generative solution: I know, I know. There’s probably a false equivalence here. This entire post started as a note and I just kept going. This post itself needs further thought and simplification. But that’ll have to come in a subsequent post, otherwise this never gets published lol. Email · Mastodon · Bluesky

You can just change things! That's the power of open source. But for a lot of people, it might seem like a theoretical power. Can you really change, say, Chrome? Well, yes! We've made a micro fork of Chromium for Omarchy (our new 37signals Linux distribution). Just to add one feature needed for live theming. And now it's released as a package anyone can install on any flavor of Arch using the AUR (Arch User Repository). We got it all done in just four days. From idea, to solicitation, to successful patch, to release, to incorporation. And now it'll be part of the next release of Omarchy. There are no speed limits in open source. Nobody to ask for permission. You have the code, so you can make the change. All you need is skill and will (and maybe, if you need someone else to do it for you, a $5,000 incentive 😄).

Jan Miksovsky lays out his idea for website creation as content transformation. He starts by talking about tools that hide what’s happening “under the hood”: A framework’s marketing usually pretends it is unnecessary for you to understand how its core transformation works — but without that knowledge, you can’t achieve the beautiful range of results you see in the framework’s sample site gallery. This is a great callout. Tools will say, “You don’t have to worry about the details.” But the reality is, you end up worrying about the details — at least to some degree. Why? Because what you want to build is full of personalization. That’s how you differentiate yourself, which means you’re going to need a tool that’s expressive enough to help you. So the question becomes: how hard is it to understand the details that are being intentionally hidden away? A lot of the time those details are not exposed directly. Instead they’re exposed through configuration. But configuration doesn’t really help you learn how something works. I mean, how many of you have learned how typescript works under the hood by using tsconfig.json? As Jan says: Configuration can lead to as many problems as it solves Nailed it. He continues: Configuring software is itself a form of programming, in fact a rather difficult and often baroque form. It can take more data files or code to configure a framework’s transformation than to write a program that directly implements that transformation itself. I’m not a Devops person, but that sounds like Devops in a nutshell right there. (It also perfectly encapsulates my feelings on trying to setup configuration in GitHub Actions.) Jan moves beyond site creation to also discuss site hosting. He gives good reasons for keeping your website’s architecture simple and decoupled from your hosting provider (something I’ve been a long time proponent of): These site hosting platforms typically charge an ongoing subscription fee. (Some offer a free tier that may meet your needs.) The monthly fee may not be large, but it’s forever. Ten years from now you’ll probably still want your content to be publicly available, but will you still be happy paying that monthly fee? If you stop paying, your site disappears. In subscription pricing, any price (however small) is recurring. Stated differently: pricing is forever. Anyhow, it’s a good read from Jan and lays out his vision for why he’s building Web Origami: a tool for that encourages you to understand (and customize) how you transform content to a website. He just launched version 0.4.0 which has some exciting stuff I’m excited to try out further (I’ll have to write about all that later). Email · Mastodon · Bluesky