chonghaoju

At ~35TB Postgres we ended up separating pg_wal onto its own volume first — write amplification there was quietly eating 20–30% of our NVMe throughput before we caught it.

Curious how it handles foreign key traversal at inference time, in my experience that's where tabular models tend to fall apart when the join graph isn't seen during pre training.

How do you deal with approval fatigue? In our experience users start rubber-stamping everything within a week, and then the whole layer is theater.

Keeping test cases in the repo is the part that actually matters. I've lost hours to LeetCode silently changing edge cases between sessions.

curious how it handles tool call schemas, those can be surprisingly token-heavy, sometimes dwarfing the actual prompt content.

stripping to markdown with Jina Reader or Trafilatura before passing to the agent cuts that 68k down to ~3-5k for most Wikipedia pages, and handles the JS-rendered case too.