How LatencyStats solves coordinated omission
masterStandard latency tracking (measuring time immediately before and after an operation) suffers from coordinated omission. This happens in two ways:
- Pause during operation: A single long latency is recorded, but it fails to account for the fact that other pending requests were also stalled by the same pause.
- Pause outside operation: If a pause occurs between operations, no long latency is recorded at all, even though the system was effectively unavailable for those requests.
LatencyStats solves this by using pluggable pause detectors and interval estimators. These components work under the hood to transparently produce corrected histogram values that reflect the true latency behavior experienced by requests, including the impact of system pauses.