2026W28
A few interesting articles I read over the past few days
This is the output of an automated process. Every Sunday, a script retrieves articles I've saved and read, uses AI to expand my quick notes into something more coherent, then publishes them. This post is one of those articles.
- @clickhouse/rowbinary: when your library is also a parser compiler — The framing that stuck with me is that the compiler here is the LLM itself. Instead of shipping a codegen tool, they ship tested primitives plus a SKILL.md and let the agent fork and monomorphize a parser per schema. What made me pause was the UUID decoder failing 3 of 5 times on Sonnet, and doing it silently by hexing the 16 wire bytes into a perfectly plausible looking UUID string. That is exactly the failure mode I would never catch in review, which is why giving the model a reference implementation to copy instead of reasoning from memory feels like the actual lesson here. At roughly 0.20 dollars per generated parser it is cheap, but I keep wondering how you regression test something the model rewrites on every query.
- SlateDB: An Object-Native LSM for Online Systems — SlateDB — The thing I had not connected before is that an LSM and object storage want the same shape. Both batch mutations into immutable blobs, so rather than fighting S3 the way a RocksDB on S3 hack does, the data structure and the storage model finally agree. I appreciated that they stayed honest about the cost: reads lag, and p99 keeps them under RocksDB’s 100K QPS ceiling, which is the price for durability living in S3 instead of local NVMe. The part I keep turning over is O(1) branching through manifest references, basically git branch for a database with no data copied. Whether that read latency is acceptable is entirely workload dependent, and I am genuinely not sure where my own workloads would land.