Uber Data Analyst Interview Questions

“I was asked whether a faster delivery promise would raise orders. The data said yes, customers ordered more. But I also looked at the delivery staff, which nobody had asked for. With the faster promise, they had more late deliveries counted against them and more of them stopped working with us. I put both findings on one page, side by side, with the same weight. I did not recommend for or against. I showed what each option would cost and who would pay. The team chose a middle promise for busy hours. The outcome was your result.”

“Our overall waiting time looked stable for months and the team was satisfied. I noticed complaints rising in a few smaller regions, so I split the number by region. Waiting time had improved in the two largest regions and become clearly worse in several small ones. Because the large regions had most of the volume, the average did not move. I showed the team a chart with one line per region instead of one total. We changed the weekly report to list the regions that got worse, next to the overall number. The result was your result.”

“For riders I would track the share of requests that end in a completed trip, and the time from request to pickup. For drivers I would track how much of their online time is spent with a passenger, since waiting and driving to pickups bring them nothing. Then a balance measure: requests against available drivers, by area and hour. These have to be read together. Very short waits may simply mean too many idle drivers. Busy drivers may mean riders cannot get a car. Health is when both sides are reasonably served, so I would show them on one page.”

“I would start from the requests table, with one row per request and its final outcome. If outcomes are in a status history table, I first take the last status per request. The request time is converted to the local time of its city, because hour of day only makes sense locally. Then I group by city and local hour. Inside each group I count requests per outcome with conditional sums, and divide each by the total. I would check that the four shares add up to one. I would also decide how to treat a rider who requests again right after no driver was found.”

“In a normal test, the two groups do not affect each other. Here they share one pool of drivers. If the new matching gets drivers to its riders faster, those drivers are taken from riders in the old group, whose wait grows. The difference then looks bigger than it would be if everyone had the new method. I would split by something that keeps the pool separate: whole cities, or blocks of time within a city that switch between old and new in random order. That needs more time for a clear answer, and I must account for differences between hours and days.”

“The status table has one row per change, so first I need periods. For each driver, ordered by time, I use a window function to take the time of the next row. A row with status online, paired with the next change, gives one online period with start and end. Periods still open at the end of the day are cut at midnight. Then I left join trips on driver and on the trip start falling inside the period, and count trips per period. I keep periods longer than the chosen length with a count of zero. I would check a few drivers by hand.”

“First I would compare riders in that city on the new app version with those still on the old one. If both cancel more, the update is probably not the cause. Then I look at when riders cancel: right after seeing the price, while waiting for a match, or while the driver is on the way. Each stage suggests a different cause. I would also check what else changed in that city that week: fewer drivers, longer waits, weather, a local event. If only the new version is affected and only in that city, I would look for something local, such as a language or payment option.”

“What I like is that the data describes a balance, not a single customer. Every number has two readings, one for the person asking and one for the person providing, and a decision that ignores either will fail. It is also very local. A city is its own small economy with its own rush hour and weather. In your project I learned how much an average can hide, and I became careful about breaking numbers down. I want to learn how to design tests when groups affect each other, which I have only read about so far.”