yanacuti1121 OBEY Release It! by Michael T. Nygard
When to use
Use when operational survivability matters and context is tight.
Primary bias to correct
Production failure semantics, overload behavior, isolation, recovery, and diagnosis must be designed, not discovered after release.
Decision rules
- Assume dependencies, queues, caches, callers, timeouts, bad data, and degraded states fail slowly, partially, and for longer than expected.
- Prefer visible failure, blast-radius limits, load shedding, preserved core service, and diagnosis over happy-path elegance.
- Bound outbound calls, waits, retries, queues, pools, caches, logs, result sets, payloads, and scarce resources where finite response matters.
- Retry only when safe, bounded, and backed off or jittered; never retry permanent failures or stack retries into storms.
- Isolate failure with circuit breakers, bulkheads, fast failure, separate pools, degraded modes, and deterministic cleanup.
- Treat deployment, startup, automation, health, observability, configuration, rollback, security, and runtime state as production design.
- Validate external input and responses before trusting them; prevent bad data from poisoning caches, queues, state, or downstream systems.
- Make APIs, interconnects, caches, jobs, and operational controls explicit about failure, capacity, recovery, authorization, and stop behavior.
Trigger rules
- When adding a remote call, wait, queue, pool, cache, job, retry, or large result path, define bounds, failure behavior, validation, and saturation signals.
- When changing startup, deployment, migration, configuration, script, or control tooling, make recovery, auditability, observability, and restartability explicit.
- When traffic can concentrate through routing, scheduling, retries, fan-out, or hostile use, add back pressure, shedding, pacing, or isolation before expensive work.
- When using production tests, game days, chaos, or disaster drills, require hypothesis, blast-radius limit, observability, stop condition, and recovery path.
Final checklist
- Bounded?
- Retry disciplined?
- Failure isolated?
- Load controlled?
- Data validated?
- Observable?
- Recoverable?
- Stoppable?