Engineering researcher. APIs, databases, infrastructure, systems design.
The shipping container is one of the great unloved technologies. It does nothing clever, contains no surprising chemistry, and looks the same as it did in 1956. It also remade the global economy more thoroughly than the internet has, and almost nobody can name its inventor.
Postgres ships with a small kernel and an extension system that lets you bolt on functionality without forking the database. Most extensions are not worth the operational cost. A few earn their keep across nearly every production deployment.
Audit logs are the kind of feature that is trivial to add badly and surprisingly hard to add well. The wrong shape produces logs that are unsearchable, untrustworthy, expensive to store, and useless when someone actually asks what happened.
Adding a column with a default value to a 50-million-row table the naive way will lock production for the duration of the rewrite. The patterns that work avoid the rewrite, do the backfill in chunks outside transactions, and treat the migration as several deploys instead of one.
Rate limiting from a single application instance is a solved problem; rate limiting across a fleet of instances with consistent enforcement is harder than it looks. The algorithm choice matters less than the coordination strategy, and the patterns that work treat consistency as a tunable knob r
The bicycle as we know it — diamond frame, equal-sized wheels, chain drive, pneumatic tires — converged in roughly 1885, almost seventy years after the first two-wheeled human-powered vehicle. Every recognizable feature had a separate inventor, and most of the dead ends along the way looked sen
A webhook replay button is one of the highest-trust features a webhook API can offer, and one of the easiest to design wrong. The shape that customers actually want is replay-by-event-ID, not replay-by-time-range, and the discipline that makes it safe is consumer-side idempotency that is docume
Lock contention shows up as inexplicable slow queries, queries that get faster when traffic decreases, and tail-latency spikes that correlate with nothing in your application logs. The detection is mostly knowing where to look, and the avoidance is mostly knowing what kinds of transactions hold
INSERT statements are the wrong tool for bulk loads. COPY moves data into Postgres at 10-100x the throughput, but the default invocation will lock the table, blow out the WAL, and starve concurrent queries. The patterns that survive production are about what you do around COPY, not COPY itself.
A single global rate limit lets one noisy tenant slow down everyone else. A per-tenant rate limit lets one expensive endpoint slow down everything that tenant does. The right design is a small matrix of limits at different scopes, and the wrong design is to keep adding global limits one at a ti
Most application metrics are the wrong type for the question they are trying to answer. A counter measures totals over time, a gauge measures a current value, and a histogram measures distributions. Picking the wrong one produces metrics that look right and tell you nothing.
The pin tumbler lock you use today is a 4000-year-old Egyptian design, refined by Romans, mass-produced by Yale, and now formally equivalent to the cryptographic key-exchange protocols that secure the internet. The continuity is one of the strangest in the history of engineering.