Netflix Service Topology: mapping thousands of microservices in real time
InfoQ news piece (2026-06) on Service Topology, an internal Netflix system that builds and maintains a live dependency graph of thousands of microservices in near real-time, so engineers can see how services connect and troubleshoot distributed failures faster. It is a concrete, production-scale instance of service-topology as an observability capability.
The problem
Engineers fixing distributed systems need unified visibility into service dependencies, a failure’s blast radius, and whether an issue is local or upstream. Existing observability tools gave fragmented views.
The three-source approach
Service Topology fuses three telemetry sources, each compensating for the others’ blind spots — the core design idea:
- eBPF network flow logs — kernel-level traffic capture. Broad coverage but lacks application context.
- IPC metrics — application-level endpoint data from instrumented services. Rich context but misses uninstrumented services.
- Distributed traces — real request paths, but sampled, so incomplete.
No single source is sufficient; merging the three yields a fuller graph than any one.
Architecture
- Three sources ingested into separate graph partitions, then merged.
- A three-stage aggregation pipeline collapses multi-hop paths through intermediaries into direct application-to-application connections.
- Historical queries use time-window aggregation rather than per-instant snapshots — cuts storage cost while still supporting incident correlation.
Stack
Apache Pekko Streams (processing pipeline), Apache Kafka (multi-region consumption), gRPC (topology query API), and Netflix’s internal distributed key-value store (graph storage).
Why it routed here
Production observability / SRE tooling — the third tight ops piece that triggered the spin-out of this spoke. Pairs with google-sre-agentic-ai (which uses observability + topology for incident investigation) and kubernetes-integration-tax (observability tooling integration in production). See synthesis.