Question

You run a 5-node distributed datastore across AZ-a, AZ-b, AZ-c. At 02:00 you see a confusing pattern: writes succeed but some reads return stale data, leader elections flap, and two nodes each believe they're a valid quorum member while a third logs 'cannot reach' for one peer but not the others. Dashboards: AZ-a↔AZ-b connectivity is fine, AZ-a↔AZ-c is fine, but AZ-b↔AZ-c packet loss is ~70% — a PARTIAL partition where only one link is degraded, not a clean split. No deploy; cloud provider reports a 'networking event' in one AZ. How do you triage and mitigate?