Add the probabilistic_sampler processor to the Collector config under processors:
Set sampling_percentage to the desired percentage (e.g., 10 for 10%) and optionally set hash_seed for deterministic consistent sampling across Collector replicas
Place probabilistic_sampler after the memory_limiter and before exporters in the pipeline definition to avoid sampling before backpressure kicks in
Pair with the SDK-side parent-based sampler so spans whose parent was sampled always propagate as sampled regardless of the Collector rule
Validate with the debug exporter on a test pipeline to confirm the drop ratio before deploying to production
Known gotchas
Collector-side head sampling is stateless; it cannot guarantee complete traces if spans for a single trace arrive at different Collector replicas without a loadbalancing exporter upstream
A sampling_percentage below 100 combined with client-side AlwaysOn sampling causes the Collector to re-sample already-sampled data, which is correct but surprising
Changing hash_seed between replicas breaks consistent sampling, meaning a single trace may be partially kept and partially dropped
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp