When Failure is Not an Option: Processing Real Money at Monzo with Kubernetes and Linkerd
summary
Monzo runs ~200 microservices and thousands of instances on Kubernetes and adds Linkerd as an out‑of‑process mesh to protect real money flows. Kubernetes cuts ops and packaging costs; moving to it reduces infra spend ~60% immediately and ~80% per user since. Linkerd runs one daemon per node, uses the Endpoints API for discovery, adds request‑level retries with retry budgets, latency‑aware EWMA + power‑of‑two‑choices load balancing (test improved ~95% → ~99% success), deadline propagation, tracing, ingress and gRPC support, and a Rust TCP balancer for non‑HTTP protocols.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 1:29Vision
- 2:05Technology
- 2:32Requirements
- 5:24Cost Savings
- 6:15Communication
- 6:57RabbitMQ
- 7:34Linkerd
- 8:21Layer 5 load balancer
- 9:42Linker review
- 10:13Layer 3 history
- 10:46TCPIP history
- 11:01DNS history
- 11:27Kubernetes services
- 11:56Cloud native environment
- 12:46Logical naming
- 13:36Request overrides
- 14:17Dtrees
- 15:27Linkerd Deve
- 15:58Call Graph
- 16:47Timeouts
- 19:24Load shedding
- 20:04Backpressure
- 21:32Load Balancing
- 25:32Linkerd History
- 29:17Rust Load Balancer
- 30:25Questions
- 35:01Linkerd Routing
related talks
22mService Mesh at Scale: How Xbox Cloud Gaming Secures 22k Pods with Linkerd
18mHow Linkerd Helped bink Partner up with Barclays
28mOverview and State of Linkerd
52mService Mess to Service Mesh
34mService Mess to Service Mesh
33mService Meshes: Istio, Linkerd - or No Mesh at All?
28mService Meshes: At What Cost?
22mCompliance the Easy Way: Zero-conf mTLS for Dev and smooth Day-2 f...