stat.change

Drift detectors. Each member tracks a running statistic, applies a configurable threshold, and exposes an alarm flag on its result. The three differ in what they assume about the in-control state and how they decide a shift has occurred.

Picking a detector

How they compare

CusumStat is the strictest and the cheapest if you already know the in-control target. It tracks two cumulative deviations from the target and alarms when either crosses a configured threshold; the allowance absorbs in-control noise. Suitable for monitoring against a known SLO or a calibrated baseline.

PageHinkleyStat is the next step up: same one-sided cumulative-deviation idea but the reference is the running mean rather than a fixed target. Add a tolerance to absorb in-control fluctuation. Suitable for streams whose baseline drifts over time but where you still want to alarm on sudden shifts.

AdwinStat is the heaviest and the most flexible. Maintains a window of recent observations as an exponential histogram of buckets; on every update it checks whether the window can be split into a recent portion and an older portion that differ by more than the Hoeffding bound at confidence delta. When such a split exists, the older portion is dropped and the running statistic reflects only the post-shift regime. Suitable when the stream may have multiple regimes and you want the detector to track the current one without explicit reset.

Output

All three produce a result with the running running statistic plus an alarm: Boolean field. The alarm doesn't auto-reset; caller decides what to do (page, reset the stat, write to an event store). Repeated alarm reads return the same value until the underlying statistic falls back below threshold.

For drift detectors on classification accuracy specifically, feed the binary correctness signal 1[predicted == truth] through one of these stats. The same pattern works for any binary signal.

Merge

All three drift detectors merge only approximately; the state is an order-dependent recurrence with no exact parallel combine. CusumStat and PageHinkleyStat average their cumulative-deviation cells (and PageHinkleyStat weight-averages the running mean); AdwinStat carries over the change counter without reconstructing the windowed histogram. Treat merge as a roll-up convenience; for distributed drift detection, run a detector per stream rather than merging partials.

Concurrency

All three keep coupled state (running statistic + previous reference) that has to stay consistent across reads. Locked under com.eignex.kumulant.core.Concurrency.Strict / com.eignex.kumulant.core.Concurrency.HighWrite. Under com.eignex.kumulant.core.Concurrency.Relaxed the cells race and the alarm signal may flicker briefly under contention.