Above The Noise Floor

A team tracks per-product sales with a count-min sketch created as CMS.INITBYPROB sales 0.001 0.002, which sets width $w = 2/\varepsilon = 2{,}000$ and a depth from the probability. CMS.INFO reports count (the sum of all increments) as 2,400,000. A product whose true count is 900 reads back as 3,100.

What is the correct reading of that 3,100?

Tighten the probability from 0.002 to 0.0002 so the extra depth lets the per-row minimum discard the colliding rows

The absolute error can reach $\varepsilon N = 0.001 \times 2{,}400{,}000 = 2{,}400$ , so anything under about 2,400 is indistinguishable from collision noise; trusting a true count of 900 needs $\varepsilon \le 0.0001$ , which means $w = 20{,}000$

The sketch under-counted somewhere else and redistributed the difference; merging a second sketch with CMS.MERGE averages the error away

3,100 is above the 2,400 threshold, so it is a trustworthy estimate and the 900 figure from the source system is the one to doubt