Part 10 of 12 · 1 min read
Quick notes
- Durable execution means a run survives the process that was executing it. A journal records each completed step; on restart, completed steps are replayed from the journal and execution continues at the first unfinished one.
- Restarting from scratch costs more and repeats side effects; journalled resume costs little more than a crash-free run.
- The dangerous window is "effect done, journal not written." Give each external effect an idempotency key (run id plus step name) to make the repeat harmless.
- Workflow code must be deterministic on replay: record clock and random values on first execution, keep LLM calls and I/O inside recorded steps, and version workflows so old journals are not resumed by new code.
- Waits are states, not sleeping workers, and sagas run compensations in reverse order when a later step fails permanently. Journal what is expensive or non-repeatable.