Cache warm-up strategies for edge functions
April 8 · 6 min read
Been experimenting with different pre-warm patterns for cold starts. The naive approach of pinging every 30 seconds burned quota fast; a smarter version that keys off traffic patterns cut it by nearly 70% while keeping p99 latency flat.
The interesting bit was that "predictive" warming worked best not when it modeled traffic well, but when it modeled the gaps. Absence of requests is a stronger signal than presence.
Notes on TLS 1.3 session resumption
March 22 · 4 min read
Session tickets vs pre-shared keys — the tradeoff is subtle when your client base is heavily mobile. Migrating to 0-RTT gave us the biggest single latency win of the quarter, but the replay-safe request-vetting logic took longer to get right than the migration itself.
Reading list, Q1
March 2 · 2 min read
A short pile: Designing Data-Intensive Applications (revisiting for the third time, still finding new things), a few of Aphyr's Jepsen writeups, and Amos's essays on Rust ownership. Not much fiction this quarter.