Traces missing half their spans
Tail sampling was added so every failed request is kept and the rest are dropped. The collectors run behind an ordinary load balancer, several replicas, round-robin.
Traces now render with gaps and orphaned spans. Nothing errors, no alarm fires, resource usage looks normal — and it gets worse every time the collector tier scales out.
Boundaries, outermost first: Samplers · stateful: Sampler, Sampler, Sampler Outside every boundary: Services (emit spans), an empty slot for the how spans are distributed Connections: Services publishes to how spans are distributed — spans (step 1) how spans are distributed publishes to Sampler — never split (step 2) how spans are distributed publishes to Sampler how spans are distributed publishes to Sampler