DoorDash · Primly Community

DoorDash machine learning engineer interview, what the loop actually looked like for recsys and ETA

ml_mike · 5 replies

Did the DoorDash MLE interview in Q1 2026 for a role on one of their prediction teams (I was targeting the ETA/logistics ML side, not ads). Writing this up because most MLE interview content online is generic and DoorDash's flavor is specific.

First: the role maturity is real. Their ML infrastructure is genuinely solid. If you've only seen early-stage ML at startups you'll feel a gap in the system design round.

The loop (senior MLE, IC5-equivalent):

Coding (45 min): Leetcode-style, medium difficulty. Not ML-specific at all. Mine was a graph problem. They want to know you can code, full stop. Python preferred but not required.

ML fundamentals (45 min): This is not a generic "explain gradient descent" session. They go deep. I got questions on: how would you detect and handle distribution shift in a production model, what's your approach to feature importance for a sparse feature set, how would you set up an A/B test for a ranking model change when the metric you care about (total orders) lags your proxy metric. Know your causal inference basics.

ML system design (60 min): Design a restaurant ranking system for the DoorDash home feed. This is the centerpiece round. They want: problem framing, feature engineering (what signals exist, what you'd collect), model choice and tradeoffs, offline vs online evaluation, how the model gets updated in production, and how you'd catch silent failures. I drew out a full pipeline and we spent 20 minutes on the offline-online metric gap alone.

Behavioral (45 min): DoorDash behavioral format. Impact-focused. "Tell me about the last model you shipped and what happened six months later" type questions. They care about post-launch ownership.

Cross-functional (30 min): How you work with ops and product when the model does something unexpected. Real scenario questions.

I came in with recsys background (recommendation systems, embedding models) and that was a good fit. The ETA side is more time-series and survival analysis flavored.

Comp offer I received (SF, 2026): base $210k, 25% bonus target, RSU valued at $280k over 4 years. Total TC around $280k first year.

5 replies

ds_dmitri

The offline-online metric gap question is one of the best MLE interview topics I've seen. Did they want you to propose a specific solution or just show that you understand the problem exists?

ml_mike

Both. Naming the problem and not hand-waving it is table stakes. They wanted to see that I'd actually dealt with it, meaning I had a concrete story of a model that looked great offline and underperformed in prod and what we diagnosed. The more specific the better.

hardware_hugo

TC of $280k first year for IC5 MLE in SF checks out with what I've seen. Their RSU refreshes have been decent in recent cycles. The bonus is actually paid out, which is not always the case at growth-stage companies.

analyst_ana

Do they distinguish between MLE roles and data science roles in their leveling? Or is it one track? I'm a DS with some ML eng experience wondering which track to target.

ml_mike

Separate tracks but they blur in practice. MLE at DoorDash is expected to own production systems (serving, monitoring, retraining). DS tends to be more analytical and experimental. If you're comfortable with both, apply to the MLE track and say so in the interview.