Stream Processing of Website Click Data

Project

Consider we run an e-commerce website. An everyday use case with e-commerce is to identify, for every product purchased, the click that led to this purchase. Attribution is the joining of checkout(purchase) of a product to a click. There are multiple types of attribution; we will focus on First Click Attribution.

Our objectives are:

1. Enrich checkout data with the user name. The user data is in a transactional database.
2. Identify which click leads to a checkout (aka attribution). For every product checkout, we consider the earliest click a user made on that product in the previous hour to be the click that led to a checkout.
3. Log the checkouts and their corresponding attributed clicks (if any) into a table.

Prerequisites

To run the code, you'll need the following:

1. git
2. Docker with at least 4GB of RAM and Docker Compose v1.27.0 or later
3. psql

If you are using windows please setup WSL and a local Ubuntu Virtual machine following the instructions here. Install the above prerequisites on your ubuntu terminal, if you have trouble installing docker follow the steps here.

Architecture

Our streaming pipeline architecture is as follows (from left to right):

1. Application: Website generates clicks and checkout event data.
2. Queue: The clicks and checkout data are sent to their corresponding Kafka topics.
3. Stream processing:
   1. Flink reads data from the Kafka topics.
   2. The click data is stored in our cluster state. Note that we only store click information for the last hour, and we only store one click per user-product combination.
   3. The checkout data is enriched with user information by querying the user table in Postgres.
   4. The checkout data is left joined with the click data( in the cluster state) to see if the checkout can be attributed to a click.
   5. The enriched and attributed checkout data is logged into a Postgres sink table.
4. Monitoring & Alerting: Apache Flink metrics are pulled by Prometheus and visualized using Graphana.

Code design

We use Apache Table API to

1. Define Source systems: **[clicks, checkouts and users] generates fake click and checkout data.
2. Define how to process the data (enrich and attribute): **[Enriching with user data and attributing checkouts ]
3. Define Sink system: **[sink]

We store the SQL DDL and DML in the folders source, process, and sink corresponding to the above steps. We use *

Run streaming job

Clone and run the streaming job (via terminal) as shown below:

git clone the repo
cd streaming_click_data
make run # restart all containers, & start streaming job

1. Apache Flink UI: Open http://localhost:8081/ or run make ui and click on Jobs -> Running Jobs -> checkout-attribution-job to see our running job.
2. Graphana: Visualize system metrics with Graphana, use the make open command or go to http://localhost:3000 via your browser (username: admin, password:flink).

*## Check output

Once we start the job, it will run asynchronously. We can check the Flink UI (http://localhost:8081/ or make ui) and clicking on Jobs -> Running Jobs -> checkout-attribution-job to see our running job.

We can check the output of our job, by looking at the attributed checkouts.

Open a postgres terminal as shown below.

pgcli -h localhost -p 5432 -U postgres -d postgres 
# password: postgres

Use the below query to check that the output updates every few seconds.

SELECT checkout_id, click_id, checkout_time, click_time, user_name FROM commerce.attributed_checkouts order by checkout_time desc limit 5;

Tear down

Use make down to spin down the containers.

Contributing

Contributions are welcome. If you would like to contribute you can help by opening a github issue or putting up a PR.

References

1. Apache Flink docs
2. Flink Prometheus example project