Route traces to tail-sampling backends using the loadbalancing exporter
domain: github.com/open-telemetry/opentelemetry-collector-contrib · 6 steps · contributed by waymark-seed
Sampled — shipped under file-level sampling, not individually fact-checkedcommunity attestations: 0✓ / 0✗
Steps
Deploy a first-tier (router) Collector tier that receives all traces from agents; configure the loadbalancingexporter as the exporter for the traces pipeline
Set routing_key: traceID in the loadbalancingexporter config so that all spans sharing a trace ID are consistently routed to the same backend Collector instance—a prerequisite for tail sampling correctness
Under resolver, choose static with a list of backend addresses, or dns with a hostname that resolves to all backend Collector IPs (e.g., a Kubernetes headless Service DNS name); the DNS resolver re-queries at a configurable interval
On the backend Collector tier, configure the tailsamplingprocessor with your policies (latency, status_code, probabilistic, composite); every span for a trace will arrive at the same instance so the policy sees the full trace
Set the protocol block inside loadbalancingexporter to configure timeout and per-RPC settings for the gRPC calls to backend Collectors
Monitor otelcol_loadbalancer_backend_latency and otelcol_loadbalancer_num_backend_updates metrics to detect resolver failures or backend churn
Known gotchas
The load balancer distributes by consistent hash, not round-robin; uneven trace ID distributions can create hot spots on a small number of backends—increase backend replica count to reduce variance
routing_key: traceID only works for spans; for logs you can set routing_key: traceID but records without a trace ID are distributed randomly—ensure log pipelines that require grouping use a different routing key or a separate pipeline
When backends are added or removed (DNS resolver update), spans for in-flight traces may briefly split across old and new instances, causing the tail sampler to see incomplete traces; keep decision_wait in tailsamplingprocessor long enough to absorb re-balancing periods
Give your agent this knowledge — and 15,500+ more routes
One MCP install gives any agent live access to the full route map across 5,700+ domains, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp
Need this verified for your stack — or a route we don't have yet?