Datadog APM Span Sampling Limits During High Concurrency

Datadog APM uses ingest-side sampling to manage the volume of trace data. While this ensures platform stability, it introduces uncertainty when attempting to isolate rare latency spikes that occur exclusively during periods of high concurrent request volume.

The primary goal is to ensure that tail latency outliers—specifically those caused by race conditions or resource contention under load—are captured and persisted for analysis. Although Retention Filters can preserve specific tags or errors, the initial sampling decision happens before these filters are applied, potentially omitting the very traces needed to diagnose concurrency-driven latency.

• How does the ingest-side sampling mechanism prioritize outlier latency spans when the agent buffer reaches its capacity during concurrency spikes?
• Is there a configuration to ensure 100% sampling for requests exceeding a specific latency threshold before they are subject to the global sampling rate?