CUSUM Change Detection: Average Run Length, False Alarms and Detection Delay
A noisy data stream suddenly shifts its mean. A one-sided CUSUM accumulator S = max(0, S + x - k) fills a Blender-built glass tube and rings an alarm when it crosses the threshold h. Tune the reference value k, threshold h and the size of the shift to trade false-alarm run length against detection delay, with the average run length solved from the Brook integral equation and checked by Siegmund approximation and Monte Carlo.
Big-data systems answer queries over data that never stops arriving, which forces a tradeoff between latency and exactness. This simulator renders that tradeoff as a live 3D pipeline: events fall down an append-only log, split into a fast speed lane that updates an approximate running estimate on every event, and a slow batch lane that accumulates in a bucket and periodically recomputes an exact aggregate from scratch. A serving-layer podium merges both into a single number using the real reconciliation formula distributed systems use — anchor to the last exact recompute, then add only the delta the fast layer has seen since — with a ghost marker showing the true value so you can watch the error shrink every time a batch fires. A one-click toggle swaps the whole pipeline from Lambda architecture (two separate engines) to Kappa architecture (one stream engine that "recomputes" by replaying the log through itself), showing exactly why Kappa is simpler but never quite as exact.
A noisy data stream suddenly shifts its mean. A one-sided CUSUM accumulator S = max(0, S + x - k) fills a Blender-built glass tube and rings an alarm when it crosses the threshold h. Tune the reference value k, threshold h and the size of the shift to trade false-alarm run length against detection delay, with the average run length solved from the Brook integral equation and checked by Siegmund approximation and Monte Carlo.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install