Zero is not a small number

A recurring mistake, this time on revenue_usd in ab-test-checkout: treating zero as just another value on the low end.

Zero in a numeric column almost always means one of two things, and they need different handling. Either the thing was measured and came out zero, or the thing did not happen and zero is standing in for absence. 89.7% of this column is zero, which is far too much to be an accident of measurement.

Averaging across both meanings gives you a number that describes neither group. The mean of revenue_usd including zeros is 6.92. Excluding them it is roughly 67.33 — a different number about a different population.

select
  count(*) filter (where revenue_usd = 0)  as zero_rows,
  count(*) filter (where revenue_usd > 0)  as positive_rows,
  avg(revenue_usd)                          as mean_all,
  avg(revenue_usd) filter (where revenue_usd > 0) as mean_positive
from ab_test_checkout;

The reporting version is usually two numbers rather than one: the rate at which the thing happens, and the size when it does. Those move independently, and a single average hides which one changed. A drop in the combined mean could be fewer events or smaller events, and you cannot tell them apart after the fact.

Same shape as the null problem, except nothing warns you, because zero is a perfectly valid number and every aggregate happily includes it.