Architect a Multi-Touch Attribution System
Tests your grasp of data pipeline trade-offs under real-world signal loss. A great answer outlines the pipeline (ingest, store, model), contrasts last-touch (simple state) vs.
What's really being asked
This question tests your ability to connect a data architecture to business reality. The interviewer isn't just looking for a textbook pipeline diagram. They are evaluating if you understand the severe practical limitations of modern marketing data, specifically signal loss and identity fragmentation. A senior answer demonstrates awareness that the data governance and identity resolution layers are far more challenging and impactful than the choice of attribution algorithm itself.
The full answer
First, the data pipeline architecture. This involves ingesting touchpoint data from disparate sources (ad platform APIs, CRM webhooks, pixel/client-side trackers) into a staging area. Second, a robust identity resolution process to stitch anonymous (cookie IDs) and known (email, user ID) identifiers into a single customer journey view. Third, loading this unified data into a warehouse (e.g., Snowflake, BigQuery) for modeling. Fourth, contrasting the models based on implementation. Last-touch is simple, requiring only the final touchpoint per user to be stored. Time-decay is stateful, requiring the storage and processing of the entire event stream for each user within the attribution window. Finally, a great answer must quantify the problem, noting that privacy changes like iOS 14.5 have reduced MTA signal coverage to just 30-60% of 2020 levels, making any model's output a heavily caveated estimate.
The mistakes people make
A common red flag is jumping straight to a complex, data-driven model (like Shapley values or Markov chains) without first addressing the foundational problems of data collection, normalization, and identity resolution. Another mistake is ignoring signal loss and speaking as if a complete, perfect view of the customer journey is achievable. Candidates also fail by describing models abstractly without detailing the concrete implementation differences in terms of data storage (storing one final event vs. a full event stream) and computational complexity (a simple lookup vs. a weighted calculation over a series).
What usually comes next
Expect questions digging into the details. For example: "How would you design the identity resolution schema?" or "How do you validate the model's output?" (A: Holdback tests, sensitivity analysis). They might also ask about business impact: "How does a 90-day vs. a 7-day attribution window change the results?" A strong answer would state that it can shift channel credit by over 20% and that B2B sales cycles of 60-180 days make short windows systematically under-credit top-of-funnel channels.
A concrete example
For a last-touch model, the pipeline's goal is to produce a table like (user_id, conversion_id, final_touchpoint_source). The logic is a simple GROUP BY user_id and MAX(timestamp) lookup before the conversion event. For a time-decay model, the pipeline must maintain a full event stream table like (user_id, event_timestamp, event_source, event_type). The attribution logic then becomes a complex function or script that pulls all events for a converted user, calculates a decaying weight for each based on its time proximity to the conversion, and assigns fractional credit. This is computationally and storage-intensive.
Interview question
When architecting a new multi-touch attribution system, what is the most critical foundational challenge that determines the system's ultimate accuracy and viability?
- a.Choosing the right data warehouse, such as Snowflake or BigQuery, for optimal query performance and scalability.
- b.Defining the appropriate attribution window (e.g., 7-day vs. 90-day) to match the business's sales cycle.
- c.Selecting the most statistically advanced model, like Shapley values or a Markov chain, to ensure fair credit allocation.
- d.Ingesting data from disparate sources and resolving user identities across devices and sessions.Correct
Why? this is the answer
The card emphasizes that the primary challenge is data foundation—ingesting and unifying data via identity resolution—not the model choice. Jumping to a complex model (C) is a common mistake without first solving the much harder problem of creating a reliable, unified user journey view.
Just read this? Test yourself on what you have been reading.
Read the original → improvado.io
You just looked this up. Could you explain it out loud?
That is the part interviews actually test. Tezvyn takes questions like this one and gives you what the interviewer is really checking, the answer that lands, and the mistake that ends the conversation, in the four minutes before your next meeting.
The iPhone app is on the way
We are building it. Until it lands, nothing here is held back from you: every interview card, your saved cards, streaks and the job board all work in Safari, plus hundreds of free practice quizzes of thirty questions each. Sign in and it all carries over to the app the day it arrives.
Want it as an icon? Tap Share at the bottom of Safari, then Add to Home Screen. It opens full screen and the cards you have read stay available offline.
We are hiring for this. Open roles that interview on data engineering — each one lists the topics its interview covers.
See open roles