Cohort analysis

When it comes to retention, cohort analysis is a common way to measure how users like your products. I helped our company to build it with our log system Elasticsearch. I decided using a faster way to build it, that means I might be abandoning some components. Cutting out the storage part is a sensible consideration. Just calculate retention rates, and show the results to my partners in a convenient way, sending the visualized data by Slack.

Before I started coding, I read many posts, trying to figure out how to build it. Some articles were too general to explain the whole process. Fortunately, there was a post I can clearly understand what it wanted to tell us for cohort analysis. So I recommend you to read it if you wanted to build a cohort analysis from scratch. Here, I won’t say too many details on calculation process, just talk about three components during implement.

An unified docker image. For all scripts of cohort analysis, I used an unified docker image to run them. It’s really easy to manage and run the various scripts by a docker image with the same version of packages. I used Python virtualenv to build a virtual environment for developing the scripts, and then used the docker image to run the scripts in our production environment.

Cohort analysis. In the scripts, selecting time frame that make a little bit annoying for me. I have to be very careful to make sure each cohort within the correct time frame, avoiding the results that were affected by a leap year or an intercalary month.

Notification by Slack. After a result of cohort analysis is calculated by the scripts, the result will be converted into a visualized output, an ascii table, to send to the Slack channels. The output in the retention channel looks like:

weekly-reporter:

+---------------+-------+--------+-------+-------+-------+
| Week          | Users | 1      | 2     | 3     | 4     |
+---------------+-------+--------+-------+-------+-------+
| 03/26 - 04/01 | 556   | 100.00 |       |       |       |
| 03/19 - 03/25 | 665   | 100.00 | 66.55 |       |       |
| 03/12 - 03/18 | 287   | 100.00 | 44.66 | 32.73 |       |
| 03/05 - 03/11 | 148   | 100.00 | 55.66 | 42.12 | 33.56 |
+---------------+-------+--------+-------+-------+-------+

Using this simple way to receive your retention rate that have some advantages: First, your partners get latest information of retention immediately by Slack notification. Second, you can get another retention rate with other conditions by a search command on your log systems. For example, if I wanted to know how gender effect retention rate, I just add a search condition gender:female to our log system Elasticsearch. Third, your logs still are kept in your site.