Two methods of performing cohort analysis in MySQL using joins - mysql

Question

I make a cohort analysis processor. Input parameters: time range and step, condition (initial event) to exctract cohorts, additional condition (retention event) to check after each N hours/days/months. Output parameters: cohort analysis grid, like this:
0h | 16h | 32h | 48h | 64h | 80h | 96h |
cohort #00 15 | 6 | 4 | 1 | 1 | 2 | 2 |
cohort #01 1 | 35 | 8 | 0 | 2 | 0 | 1 |
cohort #02 0 | 3 | 31 | 11 | 5 | 3 | 0 |
cohort #03 0 | 0 | 4 | 27 | 7 | 6 | 2 |
cohort #04 0 | 1 | 1 | 4 | 29 | 4 | 3 |
Basically:
fetch cohorts: unique users who did something 1 in every period from time_begin every time_step.
find how many of them (in each cohort) did something 2 after N seconds, N*2 seconds, N*3, and so on until now.
In short - I have 2 solutions. One works too slow and includes a heavy select with joins for each data step: 1 day, 2 day, 3 day, etc. I want to optimize it by joining result for every data step to cohorts - and it's the second solution. It looks like it works but I'm not sure it's the best way and that it will give the same result even if cohorts will intersect. Please check it out.
Here's the whole story.
I have a table of > 100,000 events, something like this:
#user-id, timestamp, event_name
events_view (uid varchar(64), tm int(11), e varchar(64))
example input row:
"user_sampleid1", 1423836540, "level_end:001:win"
To make a cohort analisys first I extract cohorts: for example, users, who send special event '1st_launch' in 10 hour periods starting from 2015-02-13 and ending with 2015-02-16. All code in this post is simplified and shortened to see the idea.
DROP TABLE IF EXISTS tmp_c;
create temporary table tmp_c (uid varchar(64), tm int(11), c int(11) );
set beg = UNIX_TIMESTAMP('2015-02-13 00:00:00');
set en = UNIX_TIMESTAMP('2015-02-16 00:00:00');
select min(tm) into t_start from events_view ;
select max(tm) into t_end from events_view ;
if beg < t_start then
set beg = t_start;
end if;
if en > t_end then
set en = t_end;
end if;
set period = 3600 * 10;
set cnt_c = ceil((en - beg) / period) ;
/*works quick enough*/
WHILE i < cnt_c DO
insert into tmp_c (
select uid, min(tm), i from events_view where
locate("1st_launch", e) > 0 and tm > (beg + period * i)
AND tm <= (beg + period * (i+1)) group by uid );
SET i = i+1;
END WHILE;
Cohorts may consist the same user ids, though usually one user is exist only in one cohort. And in each cohort users are unique.
Now I have temp table like this:
user_id | 1st timestamp | cohort_no
uid1 1423836540 0
uid2 1423839540 0
uid3 1423841160 1
uid4 1423841460 2
...
uidN 1423843080 M
Then I need to again divide time range on periods and calculate for each period how many users from each cohort have sent event "level_end:001:win".
For each small period I select all unique users who have sent "level_end:001:win" event and left join them to tmp_c cohorts table. So I have something like this:
user_id | 1st timestamp | cohort_no | user_id | other fields...
uid1 1423836540 0 uid1
uid2 1423839540 0 null
uid3 1423841160 1 null
uid4 1423841460 2 uid4
...
uidN 1423843080 M null
This way I see how many users from my cohorts are in those who have sent "level_end:001:win", exclude not found by where clause: where t2.uid is not null.
Finally I perform grouping and have counts of users in each cohort, who have sent "level_end:001:win" in this particluar period.
Here's the code:
DROP TABLE IF EXISTS tmp_res;
create temporary table tmp_res (uid varchar(64) CHARACTER SET cp1251 NOT NULL, c int(11), cnt int(11) );
set i = 0;
set cnt_c = ceil((t_end - beg) / period) ;
WHILE i < cnt_c DO
insert into tmp_res
select concat(beg + period * i, "_", beg + period * (i+1)), c, count(distinct(uid)) from
(select t1.uid, t1.c from tmp_c t1 left join
(select uid, min(tm) from events_view where
locate("level_end:001:win", e) > 0 and
tm > (beg + period * i) AND tm <= (beg + period * (i+1)) group by uid ) t2
on t1.uid = t2.uid where t2.uid is not null) t3
group by c;
SET i = i+1;
END WHILE;
/*getting result of the first method: tooo slooooow!*/
select * from tmp_res;
The result I've got (it's ok that some cohorts are not appear on some periods):
"1423832400_1423890000","1","35"
"1423832400_1423890000","2","3"
"1423832400_1423890000","3","1"
"1423832400_1423890000","4","1"
"1423890000_1423947600","1","21"
"1423890000_1423947600","2","50"
"1423890000_1423947600","3","2"
"1423947600_1424005200","1","9"
"1423947600_1424005200","2","24"
"1423947600_1424005200","3","70"
"1423947600_1424005200","4","6"
"1424005200_1424062800","1","7"
"1424005200_1424062800","2","15"
"1424005200_1424062800","3","21"
"1424005200_1424062800","4","32"
"1424062800_1424120400","1","7"
"1424062800_1424120400","2","13"
"1424062800_1424120400","3","24"
"1424062800_1424120400","4","18"
"1424120400_1424178000","1","10"
"1424120400_1424178000","2","12"
"1424120400_1424178000","3","18"
"1424120400_1424178000","4","14"
"1424178000_1424235600","1","6"
"1424178000_1424235600","2","7"
"1424178000_1424235600","3","9"
"1424178000_1424235600","4","12"
"1424235600_1424293200","1","6"
"1424235600_1424293200","2","8"
"1424235600_1424293200","3","9"
"1424235600_1424293200","4","5"
"1424293200_1424350800","1","5"
"1424293200_1424350800","2","3"
"1424293200_1424350800","3","11"
"1424293200_1424350800","4","10"
"1424350800_1424408400","1","8"
"1424350800_1424408400","2","5"
"1424350800_1424408400","3","7"
"1424350800_1424408400","4","7"
"1424408400_1424466000","2","6"
"1424408400_1424466000","3","7"
"1424408400_1424466000","4","3"
"1424466000_1424523600","1","3"
"1424466000_1424523600","2","4"
"1424466000_1424523600","3","8"
"1424466000_1424523600","4","2"
"1424523600_1424581200","2","3"
"1424523600_1424581200","3","3"
It works but it takes too much time to process because there are many queries here instead of one, so I need to rewrite it.
I think it can be rewritten with joins, but I'm still not sure how.
I decided to make a temporary table and write period boundaries in it:
DROP TABLE IF EXISTS tmp_times;
create temporary table tmp_times (tm_start int(11), tm_end int(11));
set cnt_c = ceil((t_end - beg) / period) ;
set i = 0;
WHILE i < cnt_c DO
insert into tmp_times values( beg + period * i, beg + period * (i+1));
SET i = i+1;
END WHILE;
Then I get periods-to-events mapping (user_id + timestamp represent particular event) to temp table and left join it to cohorts table and group the result:
SELECT Concat(tm_start, "_", tm_end) per,
t1.c coh,
Count(DISTINCT( t2.uid ))
FROM tmp_c t1
LEFT JOIN (SELECT *
FROM tmp_times t3
LEFT JOIN (SELECT uid,
tm
FROM events_view
WHERE Locate("level_end:101:win", e) > 0)
t4
ON ( t4.tm > t3.tm_start
AND t4.tm <= t3.tm_end )
WHERE t4.uid IS NOT NULL
ORDER BY t3.tm_start) t2
ON t1.uid = t2.uid
WHERE t2.uid IS NOT NULL
GROUP BY per,
coh
ORDER BY per,
coh;
In my tests this returns the same result as method #1. I can't check the result manually, but I understand how method #1 work more and as far I can see it gives what I want. Method #2 is faster, but I'm not sure it's the best way and it will give the same result even if cohorts will intersect.
Maybe there are well-known common methods to perform a cohort analysis in SQL? Is method #1 I use more reliable than method #2? I work with joins not that often, that's why still do not fully understand joins magic yet.
Method #2 looks like pure magic, and I used to not believe in what I don't understand :)
Thanks for answers!