About

I’m a platform engineer. Most of my working hours go into the parts of a system that nobody notices until they’re slow: the kernel’s network stack, the path a packet takes through a datacenter fabric, and the storage engine underneath a database. I like problems where the answer starts with a back-of-the-envelope number and ends with a flame graph.

On the side, I do mechanistic interpretability research. I want to know what a language model is actually computing, in the same way I’d want to know what a kernel is actually doing when a p99 spikes. The research log is where I keep work in progress, mistakes included.

Background

• Now: TODO, role and team, and the one system you’re proudest of keeping fast.
• Before: TODO, previous roles. Name a concrete thing you built and a number that describes it.
• School: TODO.

Things I keep coming back to

• Latency is not bandwidth, and p50 is not p99.
• fsync is usually the expensive part.
• If a design can’t survive a Fermi estimate, it won’t survive production.

Ways to follow along

• RSS: everything, or just writing or the research log. Feeds carry the full text.
• GitHub: I’m @thareqyusuf. Comments on posts are GitHub Discussions, so replying there works too.
• Email: I read everything and answer slowly. The addresses are below; decoding them is the spam filter.

The colophon explains how this site is built and served, including the numbers at the bottom of every page.