Archives

Deploying Workron to a local Kubernetes cluster as the first real integration test, and the JSON tag mismatch the cluster surfaced in five minutes.

Closing the stuck-gang gap: a coordinated drain path where one task's failure pulls its siblings down cleanly, with a retry-budget refund and opaque checkpoints that survive the round-trip.

Adding gang scheduling to Workron: a reserved state, an admission cycle that places N tasks simultaneously, and why running siblings are left untouched when one task fails.

Replacing log.Printf with slog, adding Prometheus metrics, and building request ID tracing. How observability turned Workron from a system that works into one that explains itself.

Why I went back on my own decision to stop at SQLite, what FOR UPDATE SKIP LOCKED actually does, and how context.Context propagation turned a one-file change into a twenty-file refactor.

Adding job dependencies to Workron — a new status, a cycle detection algorithm that guards against a problem that can't happen yet, and a SQL trick for querying into JSON arrays.

Swapping Workron's in-memory store for SQLite, and the interface design that made it a one-file change outside the store package.

Splitting Workron into separate binaries, then solving the harder problem: detecting when a worker dies mid-job without telling anyone.

How to build a job queue where multiple goroutines compete for work without stepping on each other, mutex-protected maps, atomic claiming, and retry logic.

Architectural decisions, trade-offs, and the blueprint behind building a job scheduler from scratch.