December 12, 2023

  • 10 minutes

User ID Stitching in Databricks

István Mészáros

User identification in data is a recurring problem for warehouse-first B2C companies. A single user can be identified with multiple aliases in the datasets:

✅ user_id - ID after sign-in
✅ device_id - Mobile or desktop device's ID
✅ anonymous_id - Tracking cookie
✅ customer_id - Stripe
✅ contact_id - Hubspot
✅ email - Used for sign-in
... and many more.

Picking a single identifier for user identification is risky. You may end up:
❌ not being able to use that identifier for all your users
❌ multiple identifiers for the same user

ℹ️ The correct solution is to merge all these IDs into a single merged ID and create a user alias table where all aliases are associated with their merged IDs.

This approach is called retroactive user recognition or ID Stitching and it simplifies data modelling significantly.
You won't have to think about which ID to use in joins. You can safely use the merged ID once they are backfilled to the all tables.

ID Stitching in Databricks Example

It is the simplest to explain this challenge with an example.  
Imagine you have a dataset that contains app events of users.

It has four columns:

  • event_time
  • device_id
  • user_id
User App Events

The device_id is always filled from the device's internal memory. However, in the data warehouse, it may change as the user switches to another device.

The goal is to find the first device_id of the user.

Creating The User Aliases Table With Graphframes

The user_ids and device_ids are in pairs sometimes, as we can see. You can consider these pairs as edges of graphs.

If we would have to visualize it, it would look like the following:

Sub graphs visualized

The subgraphs in this case are called components.
Each component represents a single user of the example app. To find the first device_id, we need to see all the user graph components.

Finding the graph components requires very complex SQL queries with recursive CTEs. SparkSQL does not yet support this at the moment. However, a handy library called Graphframes for Databricks (available in PySpark) is capable of finding the graph components with just a few lines of Python code in virtually any dataset.

Installing Graphframes on Databricks

First, we must install the Graphframes library on a Python-enabled Databricks cluster.

Find the Maven library graphframes:graphframes:0.8.3-spark3.5-s_2.12. There are multiple versions of this library. The only things that matter are the Scala and Spark version numbers.

For the above Graphframe's version is working only withDBR 14.2 (includes Apache Spark 3.5.0, Scala 2.12)


Libraries in Databricks (AWS)
Finding the Maven Package

To test if your configuration works, you must execute the following code snippet without problem.

Note, that Graphframes library require cluster checkpoints.

from pyspark.sql import SparkSession
from graphframes import *

spark = SparkSession.builder.getOrCreate()
spark.sparkContext.setCheckpointDir('s3://my_bucket/checkpoints/')

Finding the Connected Components

First, let's set up our test data.
Note: you might already have your users' usage logs. Graphframes were tested on multiple billions of rows without problem.


Here is the setup SQL code in Spark SQL with Delta tables on AWS S3.


DROP TABLE test_events;
CREATE TABLE test_events (
    event_time TIMESTAMP,
    device_id STRING,
    user_id STRING,
    merged_id STRING
)USING DELTA
LOCATION 's3://my_bucket/test/default/test_events/';

-- User1 aliases: d_1, d_2, u_1, u_2
-- User2 aliases: d_3, u_3
-- User3 aliases: d_4
INSERT INTO test_events (event_time, device_id, user_id) VALUES
('2023-11-01', 'd_1', null),
('2023-11-02', 'd_1', 'u_1'),
('2023-11-03', 'd_1', 'u_1'),
('2023-11-04', 'd_2', null),
('2023-11-06', 'd_2', 'u_1'),
('2023-11-07', 'd_2', null),
('2023-11-08', 'd_2', 'u_2'),
('2023-11-09', 'd_3', null),
('2023-11-10', 'd_3', 'u_3'),
('2023-11-09', 'd_3', null),
('2023-11-10', 'd_4', null);

select * from test_events;
Test Data Created by SparkSQL (Delta, AWS S3)

Note: the merged_id is empty. We will need to backfill that later once we have the user_aliases table.

Running the Connected Components algorithm is as simple as the following code snippet.

from pyspark.sql import SparkSession
from graphframes import *

spark = SparkSession.builder.getOrCreate()
spark.sparkContext.setCheckpointDir('s3://my_bucket/checkpoints/')

# Finding the vertices with their first occurence
vertices = spark.sql(
    """
    select          
        device_id as id,
        min(event_time) as first_event_time -- selecting first event time for the device_id
        from test_events 
        where device_id is not null
        group by 1
    union all
    select         
        user_id as id,
        min(event_time) as first_event_time -- selecting first event time for the user_id
        from test_events 
        where user is not null
        group by 1
    """
)
vertices.write.option("path", "s3://my_bucket/test/default/vertices/").saveAsTable("vertices", mode="overwrite")

# Finding the edges of the graph
edges = spark.sql(
    """
    with directed as (
    select 
        distinct 
        device_id as src, -- device_id is always present
        coalesce(user_id, device_id) as dst -- if edge pointing then pointing itself
    from test_events
    )
    select src, dst from directed
    union 
    select dst, src from directed
    """
)

g = GraphFrame(e=edges, v=vertices)
res = g.connectedComponents()

res.write.option("path", "s3://my_bucket/test/default/user_aliases/").saveAsTable("user_aliases", mode="overwrite")
Graph Components IDs per Alias

We could even stop now. The component column could be the user's merged_id. The only problem is that it may be a random number every time we run this algorithm.
To overcome this, we can create a better version of it simply by creating a derived table in the following way:

select
  id as alias,
  component, 
  first(id) over (partition by component order by first_event_time) as merged_id
from user_aliases

This way, the merged_id of the users won't be their component ID but the first device_id detected in their subgraph.

You can find the entire notebook here.

Finalisations

Once you have the user_aliases table, it is a good practice to merge the merged_ids back to the original event tables. 
You can do this by using the Spark SQL Delta Merge command.

Here is how to do it:

MERGE INTO test_events te
USING user_aliases ua
ON (
    ua.alias = te.device_id
    AND (
        te.merged_id IS NULL
        OR te.merged_id != ua.merged_id
    )    
)
WHEN MATCHED THEN UPDATE
SET te.merged_id = ua.merged_id;

It is essential to write only some merged_ids on some runs. This way, the Databricks cluster doesn't have to rewrite every file.

The last thing you should do is VACUUM, as the merge operation will create many new files.

Summary

In this blog post, we discussed how to do ID stitching with the Connected Components algorithm in Databricks.
We used the Graphframes spark library installed on a Databricks cluster. The algorithm created the user_aliases table to merge the merged_ids to the original event table.
This solution offers a robust and scalable way to find unique users in your event datasets. It is an essential tool for B2C companies that want to measure user conversions or calculate daily active users.

This simple code snippet is also extendable, feel free to include the user emails, customer_ids, contact_ids (from CRM and Payment provider solutions) to the user_aliases table. With that you can create a robust and complete
view about your users' aliases.

Link to the notebook code: Here

Link to Databricks Merge command: Here

Explore warehouse native product analytics

See how you can benefit from warehouse native product analytics

Blogs for your growth