Happy November! I spent the weekend visiting friends in Cambridge, MA and am on an Amtrak train back to New York. We’re a bit past peak fall-foliage season, but I’m hoping for a good view over the water between Providence and New Haven.

Data Infrastructure

• TernFS — An Exabyte Scale, Multi-Region Distributed Filesystem (XTX Markets): It’s not often we get a comprehensive peek into distributed systems of this scale. My main takeaway: avoid single points of coordination as much as possible, and make sure any coordination across multiple hosts is only required for low-throughput operations.

• Small Data (Stanislav Kozlovski): Even as machine learning eats up more and more CPU cores, GPUs and hard drives, this post proves that most other data needs firmly do not fit into that bucket. Small data can get away without distributed systems, but they certainly still benefit from improvements and technologies developed to support big data.

Software Engineering

• Be Simple: Much ink has been spilled over premature optimization, but premature generalization can be just as bad of an instinct. I liked the narrative structure of this post and also definitely agree with the takeaways. Optimize code for readability over writeability, and make sure errors are legible and actionable to whoever’s on call.

• How to Fix Any Bug (Dan Abramov): A well-written post demonstrating how good software engineering practices — in this case, spending time automating a reproduction of a bug — still matter in the age of LLM assistance.

• Vibing a Non-Trivial Ghostty Feature (Mitchell Hashimoto): Vibecoding still requires a lot of manual work and understanding to make something truly production-ready. My main takeaway came from Mitchell’s description of the time he spent structuring the view model for this feature. I wholeheartedly second this advice: if you spend time defining your protocol types well it’s a lot harder for an LLM to go off the rails.

• Use the saw, fear the saw (Steph Ango): Dangerous tools are powerful tools, and powerful tools are dangerous tools, too. I’ve always appreciated the programming-as-woodworking analogy and Steph’s extension of to LLMs and power saws clicked for me. So many technologies are “dual use” and it’s important to think about both edges to the blade.