Production Judgment: Avoiding Fallback, Constant Work, Queue Backlogs, Metastable Failures & Rollback-Safe Deploys
The senior-level instincts from the Amazon Builders' Library and metastable-failure research: why rarely-run code paths betray you, how constant-work designs remove modes, how queues fall into insurmountable backlogs, why outages outlive their triggers, and how to make every deploy reversible.
About 110 minutes · 17 exam questions · 10 interview questions · 2 code labs
What you will learn
- The production-judgment lens: modes are where systems die
- Avoiding fallback: why the backup plan is usually the outage
- What to do instead of fallback: reliable primaries, pushed data, always-on failover
- Constant work: the coffee urn, Route 53 health checks and Hyperplane
- Applying constant work in Node: level-triggered reconcilers that heal themselves
- Queue backlogs are bimodal: the recovery math nobody does up front
- Backlog-tolerant queue design: fresh-first, drop the superseded, isolate the noisy
- Consumers under backlog: leases, heartbeats, in-flight limits and Little's law
- Metastable failures: when the outage outlives its trigger
- Breaking metastable loops: characteristic metrics, hidden capacity and goodput-first policies
- Rollback safety: two-phase prepare/activate for formats and protocols
- Hands-off deployments: one-box, waves, bake time and alarm-driven rollback