How to get a Cohort analysis using MySQL? - mysql

Question

Here is my current query:
SELECT DATEDIFF(created_at, '2020-07-01') DIV 6 period,
user_id FROM transactions
WHERE DATE(created_at) >= '2020-07-01'
GROUP BY user_id, DATEDIFF(created_at, '2020-07-01') DIV 6
ORDER BY period
It returns a list of users that have had at least one transaction per period (period === 6 days). Here is a simplified of the current output:
// res_table
+--------+---------+
| period | user_id |
+--------+---------+
| 0 | 1111 |
| 0 | 2222 |
| 0 | 3333 |
| 1 | 7777 |
| 1 | 1111 |
| 2 | 2222 |
| 2 | 1111 |
| 2 | 8888 |
| 2 | 3333 |
+--------+---------+
Now, I need to know, in which period, how many users have had at least one transaction again (in the term of marketing, I'm trying to picturing the retention rate by a Cohort chart). Therefore, the calculations must be done in the Cartesian algorithm; Like a self-join!
Here is the expected result:
+---------+---------+------------+
| periodX | periodY | percentage |
+---------+---------+------------+
| 0 | 0 | 100% | -- it means 3 users exist in period 0 and logically all of them exist in period 0. So 3/3=100%
| 0 | 1 | 33% | -- It means 3 users exist in period 0, and just 1 of them exist in period 1. So 1/3=33%
| 0 | 2 | 66% | -- It means 3 user exists in period 0, and just 2 of them exist in period 2. So 2/3=66%
| 1 | 1 | 100% | -- it means 1 user (only #777, actually #111 is ignored because it's duplicated in pervious periods) exists in period 1 and logically it exists in period 1. So 1/1=100%
| 1 | 2 | 0% |
| 2 | 2 | 100% |
+---------+---------+------------+
Is it possible to do this using MySQL purely?

Answer 1

You can use window functions:
SELECT first_period, period, COUNT(*),
COUNT(*) / SUM(COUNT(*)) OVER (PARTITION BY first_period) as ratio
FROM (SELECT DATEDIFF(created_at, '2020-07-01') DIV 6 period,
user_id,
MIN(MIN(DATEDIFF(created_at, '2020-07-01') DIV 6) OVER (PARTITION BY user_id)) as first_period
FROM transactions
WHERE DATE(created_at) >= '2020-07-01'
GROUP BY user_id, DATEDIFF(created_at, '2020-07-01') DIV 6
) u
GROUP BY first_period, period
ORDER BY first_period, period;
This does not include missing periods. That is a little trickers, because you need to enumerate all of them:
with periods as (
select 0 as period union all
select 1 as period union all
select 2 as period
)
select p1.period, p2.period, COUNT(u.user_id)
from periods p1 join
periods p2
on p1.period <= p2.period left join
(SELECT DATEDIFF(created_at, '2020-07-01') DIV 6 period,
user_id,
MIN(MIN(DATEDIFF(created_at, '2020-07-01') DIV 6) OVER (PARTITION BY user_id)) as first_period
FROM transactions
WHERE DATE(created_at) >= '2020-07-01'
GROUP BY user_id, DATEDIFF(created_at, '2020-07-01') DIV 6
) u
ON p1.period = u.first_period AND p2.period = u.period
GROUP BY p1.period, p2.period;