Deep-dive developer posts on scaling, system design, OS internals, AI, and the edge cases that break production — every post ships with real, working code.
Start reading →CREATE TABLE leader_lease (
job_name TEXT PRIMARY KEY,
holder TEXT NOT NULL,
expires_at TIMESTAMPTZ NOT NULL
);from: Leader Election Without ZooKeeper: A Postgres Lease, Not a New Service
Running 3 instances of a cron worker means the nightly report runs 3 times. You don't need ZooKeeper or etcd to fix that — a lease row in the database you already have gets you correct leader election in about 30 lines.
The order list page loads fine with 5 orders in dev and falls over with 10,000 in production. Here's the ORM lazy-loading trap that causes it, how to see it in the query log, and the two real fixes: eager loading and batching.
Save the order to Postgres, then publish to Kafka. The write between those two lines can fail and you'll never notice — until support tickets show orders that charged but never shipped. The outbox pattern fixes it.
RedBite's API starts throwing EMFILE at 3am, only under load, never in dev. Here's what a file descriptor actually is, why ulimit kills you in production and not locally, and the epoll-based fix that stops leaking them.
One cache key expires. 2,000 requests land in the same millisecond, all miss, all hit Postgres at once. Here's the thundering herd problem, and the request-coalescing fix that turns 2,000 queries into one.
Transfer money from account A to B at the same moment someone transfers from B to A, and Postgres kills one of them outright. Here's how to reproduce a real deadlock, read the error it gives you, and fix it with lock ordering.
