Tags
Hardcoded cron means a redeploy every time a schedule changes. Store schedules in the database and have a daemon react to edits with change streams. Here is the design and its failure handling.
GPUs are expensive and run one heavy job at a time, so round-robin routing stalls behind busy workers. Here is a busy-aware scheduler built from Redis leases and keyspace notifications.
A plain lock deadlocks forever when the holder dies. A lease expires. Here is how to build one with compare-and-swap acquisition and a heartbeat, and the failure modes that decide the design.
When a database forbids automatic write retries, every write must be idempotent on its own. A content-derived id makes that free. Here is the derivation and the one invariant that keeps it safe.
Firestore speaks the MongoDB wire protocol, but rejects retryable writes. If your driver defaults to on, writes fail in a way that looks random. Here is why, and the fix.
Rust gave us safety we rarely needed and a compile loop we fought every day. Here is the honest tradeoff, with numbers, that pushed a service to Go.