Waymo Data Scientist Interview Loop
Waymo's data scientist interview opens with a one-hour online test of SQL and statistics, with fewer than ten questions on two tables. Then there's a 30-minute recruiter call and a 45-to-60-minute technical round of live SQL or Python on simulated fleet data. The virtual onsite is four to five interviews: A/B test design, metrics, a product or autonomous driving case, and behavioral. There's no LeetCode. SQL is central, and Python is allowed but not required.
What decides it
Statistical judgment on messy operational data. That means picking the right unit, the right test and the right denominator. Candidates who treat it as a product analytics loop miss how rigorous it is.
Practice by round
SQL basics in the shape of the online test, like completed trips per vehicle per day, disengagements per 1,000 miles, and cancellations by city and cause. Then quick statistics coding, such as Bayes' rule for a rare-event detector. Watch out for duplicates and labels that don't match.
Harder SQL, like trip duration percentiles by route type, miles between interventions, and cancellation rate by weather. Then statistics coding on fleet questions: comparing two Poisson rates, a switchback experiment, and simulated versus real distributions.
The cases Waymo is known for: whether new driving software is safer, testing a suggested pickup spot, and launching in a new city. Then mini-analyses on fleet data: On-Time Pickup Test, Near-Miss Event Rates and Remote Assistance Rate. Riders and vehicles share one fleet, so they affect each other. Say how you'd randomize before you say what you'd measure.
| Case study | Difficulty |
|---|---|
| Measuring Robotaxi Ride Comfort | Medium |
| New-City Robotaxi Launch | Medium |
| Wait Times After Fleet Growth | Medium |
| Is New Driving Software Safer? | Hard |
| Testing a Suggested Pickup Spot | Hard |
| Data project | Difficulty |
|---|---|
| Remote Assistance Rate | Easy |
| On-Time Pickup Test | Medium |
| Near-Miss Event Rates | Hard |
Rare events and their traps: zero collisions in two million miles, incidents bunched on a few routes, a release that's better in every city but worse overall, and how often a pedestrian alert is real.