The overall average hides 4 different numbers

channel splits retail-orders into 4 groups. Here is what discount_pct looks like inside each.

• marketplace — mean 0.25, median 0.25 (1,151 rows)
• web — mean 0.07, median 0.00 (2,734 rows)
• android — mean 0.07, median 0.00 (1,386 rows)
• ios — mean 0.07, median 0.00 (1,729 rows)

Top to bottom that is 0.25 against 0.07, a spread of 279.9%. The pooled average is 0.10.

select channel,
  count(*)                                                  as rows,
  round(avg(discount_pct)::numeric, 2)                      as mean,
  percentile_cont(0.5) within group (order by discount_pct) as median
from retail_orders
group by 1
order by mean desc;

A spread that wide means the pooled number is not a summary, it is an artefact of the mix. Change the proportion of marketplace rows and the overall average moves without any individual group changing at all — which is how a metric goes up while every segment goes down.

Notice the mean and median columns disagree only slightly here. Always compute both in the group-by. The comparison between them per segment is free and tells you whether you are looking at a level difference or a tail difference.

This is the setup for Simpson's paradox — the case where every segment moves one way and the total moves the other.