Month-over-month change in plays

Constraints

1. 1play_month is the YYYY-MM prefix of played_at.
2. 2plays counts rows. listeners counts distinct user_id within that month — someone active in two months counts in both.
3. 3plays_change is this month's plays minus the previous month's. The first month has no previous month, so it is NULL. Do not report 0.
4. 4Months come from the data. All three months here have plays, so there is no gap to fill.
5. 5Order by play_month.

Worked example

January has 14 plays, February 11 and March 11, so plays_change reads NULL, -3, 0. Reporting 0 for January instead of NULL would state that January was flat against a month that does not exist in the data. NULL says 'unknown', which is the truth, and it is also what makes the genuine 0 in March meaningful.

What this tests

Reading the previous row's value with LAG over an aggregate, and the discipline of leaving the first period NULL rather than implying it was flat.