Finding the gaps in ab-test-checkout

Finding the gaps in ab-test-checkout comes up often enough to be worth a note. ab-test-checkout is a convenient thing to try it on — 6,000 rows, one row per exposed user.

with numbered as (
  select variant, exposed_on,
         row_number() over (partition by variant order by exposed_on) as rn
  from ab_test_checkout
)
select variant,
       min(exposed_on) as island_start,
       max(exposed_on) as island_end,
       count(*)   as readings
from (
  select *, date_trunc('day', exposed_on) - (rn * interval '1 day') as grp_key
  from numbered
) t
group by variant, grp_key
order by variant, island_start;

The trick is that subtracting a dense row number from a dense date gives a constant inside a run and a different constant across a gap. Once you have seen it the query is obvious; before that it looks like a magic trick.

Run it against the real file in the SQL playground — the engine there is enough of a SQL implementation to execute this as written, and the dataset is already loaded. The pattern page has the version with the failure modes spelled out.

Window functions are the difference between SQL that describes rows and SQL that describes sequences. Almost every "we exported it to pandas to do this bit" turns out to be one of these five.