Coordinated Omission
A benchmarking error in which the load generator stops or delays issuing work during stalls, causing measured latency distributions to omit requests that would have waited.
Measurement and Bottleneck Boundary
Coordinated omission is a measurement artifact, not an application optimization problem; fixing it often makes reported latency look worse because previously omitted waiting time becomes visible.
Related Performance Concepts
- Tail Latency
- Load Testing
- Queueing Theory
- Little's Law
Why a Histogram Alone Cannot Repair the Experiment
The defect exists in the sampling process before histogram resolution becomes relevant. Suppose a load generator intends to issue one request every 10 ms but waits for the previous request to finish. If the system stalls for 1 s, many arrivals that would have existed in an independent arrival process are never generated. A histogram sees only issued requests, so the waiting time of those missing arrivals cannot appear in the measured distribution.
Tools such as HdrHistogram can model correction when a known expected interval is available, but that correction depends on the traffic model being known and appropriate. A cleaner experiment can use an open-loop or scheduled-arrival generator that produces the intended load directly. A Tail Latency chart is therefore incomplete benchmark evidence unless the load-generation model is also specified. Implementation reference: HdrHistogram.