One Pass, Bounded Memory
Streaming means, variances, quantile sketches, and count-min counting.
Stream controls
Mean shifts upward as the stream advances.
How many rows arrive per Step or per animation tick.
Processed 500 of 20000 rows, one pass, no lookback.
Sketch sizes
Narrower width means more hash collisions and more overestimation.
Each extra row is an independent hash whose minimum we take.
Seed
Every result on this page is a deterministic function of the seed and the controls.
Running mean: Welford vs. exact batch
Recomputed batch statistics on the left axis are the ground truth; the streaming estimator should track them exactly.
Median: reservoir & bucket sketch vs. exact
Heavy hitters: count-min vs. exact
Memory vs. accuracy
Bytes retained by each estimator against its relative error on the median, at the current cursor.
Every streaming estimator here reads each row exactly once and never revisits it. Welford's update keeps the running mean and variance in four numbers total, immune to the cancellation that wrecks the naive sum-of-squares formula once values sit far from zero. The reservoir, bucket sketch, and count-min sketch each trade a fixed, small memory budget for an approximate — but bounded and provable — answer, which is the only option once n exceeds RAM.