Lyft · Primly Community

Lyft machine learning engineer interview: the recsys and ETA prediction questions they actually ask

ml_mike · 4 replies

Went through the Lyft MLE loop in spring 2026, targeting an L5 role on one of their marketplace ML teams. Wanted to write something more specific than "they ask ML design and coding" because that describes 90% of MLE interviews and tells you nothing.

Lyft's ML org is heavily focused on a few real problem areas: driver-rider matching, surge pricing, ETA prediction, and supply-demand forecasting. Your loop will almost certainly touch one of these in the ML system design round. This isn't a coincidence, so prep accordingly.

Structure of the loop:

Recruiter screen, then a technical phone screen with an MLE or senior scientist. The phone screen was one coding problem (medium, Python, feature engineering and basic model eval -- write a function that takes a dataset and a set of features, trains a logistic regression, and evaluates it. not a framework exercise; they want to see if you understand what you're calling.) plus 15-20 minutes of ML concept discussion: bias-variance trade-off, regularization, how you'd approach class imbalance.

Onsite was four rounds:

ML system design (60 min): I got "design an ETA prediction system for Lyft rides." This is a classic Lyft question at this point. Break it down: what features matter (distance, time of day, historical segment-level times, real-time traffic), how do you handle the cold start problem for new route segments, how do you retrain and how often, how do you evaluate (MAE vs. percentile errors), how do you deploy and monitor for distribution shift. They wanted specifics, not just frameworks.

Coding round 1: Algorithm-focused. Medium-hard. I got a problem about efficiently updating a streaming dataset with sliding-window statistics. Python. Time and space complexity discussion after.

Coding round 2: More applied. Given a dataset (provided as a dict), write code to detect anomalous driver behavior based on trip patterns. Not a pure algorithm problem -- they wanted to see your thinking about what "anomalous" even means here and how you'd operationalize it.

Behavioral: Pretty standard. A time you disagreed with a product decision on features for a model, a time a model you shipped underperformed and what you did. Have both stories ready.

What matters most: Deep on the ML system design. Not just "I'd use XGBoost" but why, how you'd evaluate it in production, what the failure modes are, how you'd handle feature staleness. They hire people who've shipped models, not just trained them.

Comp at L5 in SF was in the $300-$350k TC range from what I could piece together from the recruiter conversation and Levels.

4 replies

ds_dmitri

the ETA prediction design is basically a canonical Lyft question at this point -- i've seen three separate data points from people who got that or a close variant. good to know they're still running it. the distribution shift monitoring angle is worth really knowing cold. the default answer of 'retrain periodically' is not enough.

sec_sasha

the anomalous driver behavior question is interesting from a fraud/trust-and-safety angle too. did they ask about the feature engineering side (what signals actually indicate anomaly) or mostly about the algorithmic approach?

ml_mike

both, really, but the feature engineering side was where they pushed harder. 'what signals would you look at and why' before we got to 'how would you model it.' they wanted to see domain sense, not just model knowledge.

de_derek

curious whether the data engineering side of feature pipelines came up at all in the system design round. like who owns the feature store, what's the latency expectation on real-time features. or is that usually out of scope for MLE interviews there.