Incident Replay · Resilience
When an Availability Zone Goes Dark, Your Stateful Pods Hang: How to Design Around It
Pick what fails, what on-call does and the design, and watch payments-0 recover, hang or risk its data, then replay the outage.
A dark AZ in 60 seconds
- EBS (Elastic Block Store) is durable but zonal: a volume attaches only in its AZ. When that AZ goesdark, data is safe but stuck.
- The wait is a guard. A Pod on an unreachable node stays Terminating. Force-delete it and the replacement hangs about 6 minutes on the disk.
- Fence before you force. Fence = confirm the old node is stopped, then mark it out of service. payments is back a minute after that.
- Skip the check and you risk the data. A node that's only cut off still writes; forcing the disk away loses writes.
- Survive a dark AZ with a replica and a volume in each AZ. Creating each volume where its Pod lands (WaitForFirstConsumer) can't move it later.