Architect a Multi-Touch Attribution System
This tests your grasp of modern data challenges like signal loss. A good answer discusses data ingestion, identity resolution, and model trade-offs. A red flag is focusing only on the algorithm and ignoring the data pipeline's fragility.
What's really being asked
This isn't a pure algorithm question. It tests your understanding of the entire data value chain under modern constraints. The interviewer wants to see if you recognize that the biggest challenge isn't the model's math, but the data pipeline's fragility due to signal loss (iOS 14.5, cookie deprecation), fragmented user identity, and walled-garden APIs. They're testing for a senior-level perspective that prioritizes data governance and robust architecture over a theoretically perfect but impractical model.
The full answer
A strong answer addresses four key areas. First, the data pipeline architecture: describe ingesting touchpoint data from disparate sources (APIs like Google/Meta, webhooks, event streams) into a data lake or warehouse. Second, identity resolution: explain the process of stitching together anonymous (cookies, device IDs) and known (email, user ID) identifiers into a single customer journey view. Third, model implementation contrast: Last-touch is a simple SQL query (e.g., ROW_NUMBER() OVER (PARTITION BY user_id ORDER BY timestamp DESC) where rank = 1), while time-decay requires more complex logic to calculate weights based on time differences from conversion. Fourth, acknowledge business context: mention that model choice depends on the sales cycle (e.g., short 30-day windows under-credit top-of-funnel channels in B2B) and that rule-based models are more auditable for stakeholders.
The mistakes people make
The most common red flag is jumping straight to complex models like Shapley values or Markov chains without first addressing the foundational data problems. Candidates who ignore the 30-60% signal loss since 2020 sound dated. Another mistake is describing a perfect, unified log of all events, which is unrealistic. A senior engineer acknowledges the messy reality of reconciling data from walled gardens like Meta, which have their own attribution and don't easily share user-level data. Finally, failing to mention the operational cost of the pipeline is a miss.
What usually comes next
Expect questions that dig into the trade-offs. "How would you handle a user who interacts on a work laptop and then converts on a personal mobile phone?" (Tests identity resolution depth). "What's the minimum number of conversions you'd need to trust a data-driven model, and why?" (Tests statistical literacy; the answer is thousands, e.g., 2,000+ per month). "How would you A/B test two different attribution models to see which one drives better business outcomes?" (Tests business acumen and experimental design).
A concrete example
For a time-decay model, the implementation challenge is in the data transformation layer (e.g., dbt). For each conversion, you must first join all touchpoints for that user within the attribution window. Then, you apply a function to calculate a weight for each touchpoint, often an exponential decay formula like weight = 2^(-(time_to_conversion_in_days / half_life_in_days)). A 7-day half-life is common. This is computationally more expensive than last-touch, as it requires a window function over the entire user history, not just picking the last event. A last-touch model, by contrast, only needs to find the single most recent touchpoint before the conversion event.
Interview question
According to the provided card, what is the most significant challenge when designing a multi-touch attribution system in the current data environment?
- a.Establishing a unified, real-time log of all user interactions across all platforms.
- b.Determining the optimal attribution window and decay rate for time-decay models.
- c.Overcoming data pipeline fragility, signal loss, and fragmented user identity.Correct
- d.Implementing advanced data-driven models like Shapley values or Markov chains.
Why? this is the answer
The card explicitly states that the biggest challenge is "the data pipeline's fragility due to signal loss... fragmented user identity, and walled-garden APIs." Option D is identified as a "common red flag" because it ignores these foundational data problems.
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