- The $2K/month cache: a high-usage Redis storyAn 800-key cardinality observation on the “company” dimension turned an in-memory LRU into the cheapest, fastest cache we ever shipped — and cut Redis traffic by 80%.9 min read→
- Load balancing gRPC vs REST: what actually changesL4 round-robin quietly breaks HTTP/2 multiplexing — one TCP connection pins every RPC to a single backend. A field guide to balancing long-lived gRPC streams without melting one pod.11 min read→
- Audit-proof MCP systems: designing for provenanceWhen an agent can call tools, every action needs a tamper-evident trail. How to design Model Context Protocol systems where each tool call is signed, replayable, and audit-ready by default.13 min read→
- OTel & distributed tracing without the noiseTrace IDs, spans, and golden signals that shorten time-to-root-cause instead of adding dashboard clutter. What to instrument, what to sample, and what to delete.10 min read→
- The boring checklist that prevents 80% of outagesTimeouts, retries with jitter, circuit breakers, and a deploy you can roll back in 30 seconds. None of it is clever — that's exactly why it works.7 min read→
- Idempotency keys: the cheapest insurance in distributed systemsA retry you didn't design for is a duplicate you'll page about at 3am. How idempotency keys make “at least once” safe to build on.8 min read→
- Reading p99 like a detectiveAverages lie; tails tell the truth. A walkthrough of chasing a p99 regression from a Grafana panel down to a single mis-tuned connection pool.9 min read→
- RFCs aren't overhead — they're the productDesign reviews and RFCs feel like friction until you realize they're how a team thinks together. Notes on running them so they speed delivery up, not down.6 min read→
No posts under this topic yet — check back soon.