How Uber Protects Against Retry Storms

Uber introduces a context-aware retry mechanism that limits retries to a configurable budget per hop, preventing amplification across deep dependency chains. By clamping retries to 10% of requests between specific nodes, the system maintains caller-perceived availability near 99% when callee availability drops up to 10%, while avoiding overload on downstream services. This approach replaces uniform retry policies with conditional, budget-controlled retries that reduce cascade failures.