mrr_usd: mean 447, median 88

A number worth keeping: on saas-subscriptions, mrr_usd has a mean of 447 and a median of 88.

The mean is 447. The median is 88. That is a gap of 410.2%, and it is not noise — the p90 is 930 and the max is 9,751, so the top of the distribution is dragging the average somewhere a minority of rows actually live.

select
  avg(mrr_usd)                                                     as mean,
  percentile_cont(0.5) within group (order by mrr_usd)             as median,
  percentile_cont(0.9) within group (order by mrr_usd)             as p90,
  max(mrr_usd)                                                     as max
from saas_subscriptions;

The practical consequence is that "average mrr usd" answers a question nobody asked. If someone wants to know what a typical row looks like, the median answers it. If someone is forecasting a total, the mean is the right tool and the skew does not matter, because the mean times the count is the total.

So the rule is not "never use the mean". It is: name which question you are answering, then pick the statistic that answers it. A report that shows 447 with no median next to it has quietly decided you meant the first question.

There is a longer version of this in Describing a column without lying, and the pattern is the short one.