I hope everyone’s been having a good February! This week’s Weekly is more of a Monthly, but I hope you enjoy the articles all the same. I had a lot to say about these, so it’s quality of commentary over quantity of posts this time around.

• My LLM codegen workflow atm (Harper Reed): I’m still trying to figure out where LLMs fit in my workflow. The post emphasizes how fast-moving this stuff is: its guidance “probably will not work in 2 weeks, or it will work twice as well. ¯\_(ツ)_/¯“

  Call me boring, but I’d bet that we end up somewhere between the doomers, haters and hypebeasts. On one hand, these models are clearly a step change in software development similar to how online search engines and StackOverflow changed development back at the turn of the millennium. Reasoning models like o1 and o3 are able to be more independent, but on the other hand I’m still skeptical of how generative models that are best at pushing out code will fit into the design and development of more complex systems. I spend less than 40% of my engineering time heads-down writing code and tests. The rest of my day-to-day is reacting to how that code reacts to different load patterns from users and changing conditions of our underlying system dependencies.

  On the other, other hand, I’m thinking about the shape of arguments that those more skeptical of AI systems are making feel similar to Gell-Mann Amnesia — it’s easy for me to say that while LLMs are excelling at full-stack development they’ll never be able to build distributed systems, since I feel like my own experience is important in my day-to-day work. But who knows? Maybe systems engineers will be the next subdiscipline to be blindsided by some new model.

• Querying Parquet with Millisecond Latency (Raphael Taylor-Davies and Andrew Lamb): The Apache Parquet format is used through a lot of OSS “big data” infrastructure since it compresses extremely well. Once data is in a place, though, people often have the gall to want to query it out in a reasonable amount of time. This post was a great introduction to the nitty-gritty details of the Parquet format along with a walk-through of read-time optimizations, some of which are still not implemented in the mainline Parquet reader implementations like Arrow three years later.

• Are Efficiency and Horizontal Scalability at odds? (Hillel Wayne): As always, a thought-provoking post from Hillel. From my own experience trawling through some large OSS codebases for distributed systems, it’s often surprising how obviously inneffecient some parts of the system are that are considered to be off of the hot path for the modal use-case.

  I agree with Hillel’s point at the end of this newsletter that this is probably cultural to an extent. Similarly to what I wrote about Parquet above, systems built for scale often explicitly trade off interactive latencies as a matter of allocating limited development resources even if there isn’t a strict technical tradeoff in every case.

  This is maybe an interesting place for LLM assisted development! Working with a model to optimize some local hot loop in a distributed system reduces the development overhead for parts of the system that are more easily unit-testable.