How DoorDash Moved to a Service Mesh to Handle 80M Requests/Second

DoorDash experienced a platform-wide outage in 2021 caused by cascading failures in their microservices architecture. The payment service's high latency triggered retry storms that overwhelmed dependent services. This incident exposed inconsistent reliability patterns across their 1,000+ microservices. The engineering team responded by implementing a custom service mesh using Envoy as the data plane, rejecting both Istio (too complex) and Linkerd2 (insufficient features). They built a minimal control plane focused on adaptive concurrency, outlier detection, and traffic metrics. Starting with an MVP using file-based configuration and canary deployments, they gradually evolved to include zone-aware routing, header-based routing, and distributed tracing. The system now handles 80M requests/second across 2,000 Kubernetes nodes, with automated onboarding reducing migration time from days to under an hour.