Birds of a Feather - xDS and Load Balancing
summary
Discussion covers practical xDS work: many teams use go-control-plane, some hit scale limits (one report: open-source implementations can’t serve millions of subscriptions), and there’s no ready production xDS caching proxy even though a scalable pattern (stateless caching proxies multiplexing subscriptions) drops load on the authoritative server. An autosharding (slicer) protocol is being prototyped to push shard assignments via xDS so clients can route pre-batched requests, and teams debate extracting an XDS client library (Java, Go, Rust, core C++ with Python bindings) for reuse.
AI-written summary · about
chapters (13)
- 0:00Introduction and audience questions
- 1:13Autobatching challenges with sharding
- 3:00Slicer and autosharding protocol
- 5:47Custom sharding logic and XDS
- 7:55Control plane scalability at LinkedIn
- 8:57XDS caching proxy design
- 10:45Open source XDS implementations
- 12:08Extracting XDS client library
- 14:34XDS language support and parity
- 16:43Multi-language XDS feature gaps
- 18:50User experience with XDS vs DNS
- 20:36Debugging and control plane challenges
- 21:19Server load with XDS connections
related talks
23mxDS Support in gRPC
16mxDS in gRPC for Service Mesh
20mProxyless Service Mesh with gRPC
25mAdapting Our XDS Control Plane for Proxyless gRPC
17mBuilding a gRPC Proxyless World: How Reddit Scaled Resilience...
23mA Simplified Service Mesh With gRPC
46mProxyless Service Mesh for gRPC Java Services
23mHow Proxyless gRPC Works in a Service Mesh