#engineering

Scheduling 25 million checks with one Go binary — and no next_run_at column

Most schedulers store the next run time in a database and fight write amplification forever. Ours stores no per-run scheduling state at all: every job's schedule is a pure function, recomputable anywhere. Here's the design — a 256-shard time wheel, an append-only change log — and the honest numbers from load-testing it to 25 million active jobs.