The work runs itself. Somebody still has to answer for it.
Automated work is easy to start and hard to keep. A retry that heals one failure duplicates a charge in another; a schedule that fires on time says nothing about whether the backlog cleared; a workflow that runs for three weeks will outlive every process that hosted it. These essays are about the mechanisms underneath — idempotency, queues, compensation, human approval steps — and about the cases where the right answer is to leave the task alone.
All essays
- The Automation Not Worth Building
Most automation cases are argued on the time the task takes and lost on the maintenance nobody counted. On unstable processes, exception rates, the cost of supervision, and the tasks that should stay manual on purpose.
- Undo Is a Business Decision
Compensation is not rollback. On sagas across systems that never agreed to a transaction, steps with no inverse, ordering compensations by reversibility, and the semantic gap between undoing a change and apologising for it.
- Nobody Watches the Night Shift
Background work fails without anyone noticing, because there is no user to complain. On staleness as the primary signal, alerting on absence, tracing across queue boundaries, and why a green dashboard proves very little.
- Work That Outlives the Process
A workflow lasting weeks will survive deployments, restarts and code changes that a request never has to face. On durable state, versioning running instances, timers you cannot trust to memory, and why determinism becomes a constraint.
- The Step Where a Person Decides
Human approval is a step in the workflow, not an escape from it. On the difference between review and rubber-stamping, timeouts on human steps, context at the point of decision, and what happens when the approver is on leave.
- A Queue Is Not a Schedule
Schedulers answer to the clock, queues answer to demand, and confusing the two produces systems that run punctually while falling behind. On triggers, backlogs, lateness and the reconciliation loop that saves both.
- Retries Are Cheap. Side Effects Are Not.
A retry is only safe if the step it repeats is safe to repeat. On timeout ambiguity, idempotency keys that come from the caller, dedupe windows, and why retry budgets matter more than backoff curves.