How We Solved the K8s Cluster Communication Challenge
summary
Aerospike develops a custom proxy stack to enable cross-cluster replication for its low-latency DB: a stateless ESP batches XDR messages and sends them over HTTP/gRPC to an XDR-Proxy on the target cluster, which maintains the TCP sticky connections to local pods. This converts many persistent TCP client connections into scalable stateless streams, avoids external LoadBalancer limitations for stateful sets, and proves practical in production for high-throughput, low-latency replication.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:42Agenda
- 1:06Aerospike
- 2:05Smart Clients
- 3:14Kubernetes
- 4:48Load Balancers
- 5:57Stateful vs Stateless
- 6:51XDR
- 9:40Problem
- 10:48Solution
- 11:28What they do
- 12:59What we did
- 13:56Outro
related talks
16mBuilding a K8s Operator for a Distributed Database
21mBuilding a K8s Operator for a Distributed Database
30mNetwork Configurations for High-Performance Streaming Connectivity in K8s
17mBuilding a gRPC Proxyless World: How Reddit Scaled Resilience...
28mProxy Service: A New Network Traffic Abstraction in Kubernetes
35mIntelligent Load Balancing in Kubernetes
31mHow We Moved Spotify To a Proxyless gRPC Service Mesh
28mProxy Service: A New Network Traffic Abstraction in Kubernetes