LinkedIn Professional Network · View 08 of 30 · 3 · Structure
Decisions
- Every colo runs the full stack, so a member is served end to end inside one colo
- Mid-tier services are Java on Rest.li, with D2 balancing load on the client side. No central load balancer sits between services
- Messaging delivery runs on Play and Akka, holding persistent SSE connections, because holding connections scales differently from serving requests
Published numbers
- Couchbase: 10M+ queries a second across 300+ clusters (LinkedIn, 2018)
- Brooklin: 2T+ messages a day (LinkedIn, 2019)
- LIquid: 2M queries a second at 99.99% availability (LinkedIn, 2023)
Deliberately omitted
- Hundreds of smaller services; the edges drawn are representative
- Offline grid internals: view 11
- Failure domains and colo count: view 22