Notes from the teams building B3 — on the systems, decisions, and trade-offs behind the product. Deep dives on distributed infrastructure, performance, tooling, and the craft of shipping reliable software.
We moved our primary OLTP cluster to direct-attached NVMe. The gains were obvious, the foot-guns less so. A long-form retrospective with benchmarks.
A practical tour through our migration — remote cache strategy, query-level parallelism, the boring parts that actually mattered.
Our graphs said everything was green. Users disagreed. A walk through a three-hour degradation and the monitoring gaps we closed.
How the borrow checker nudged us toward a cleaner IR than we would have drawn on a whiteboard.
Flame graphs, eBPF, and a lot of humility. A case study in peeling off latency one nanosecond at a time.
We treat vector search as plumbing, not magic. Here's the architecture, the cost model, and the trade-offs.
Zero-trust sounds great on a slide. In practice it's a supply chain of certificate rotations, SPIFFE IDs, and on-call pages.
A story about idempotency keys, message de-duplication, and choosing boring correctness over clever guarantees.
A 400-line Go binary replaced a cross-team process nobody loved. A quiet lesson in shipping the tool your coworkers actually need.
Follow the blog via RSS — no email, no tracking. New articles appear in your reader the moment they publish.