Netflix Data Analyst Interview Questions

“A team had spent weeks on a new layout and expected a clear win. The test showed no real difference in how much people used the product. Before saying anything, I checked the group sizes, the length of the test and the logging, so I could trust the result. Then I met the team lead first, in private, and walked through the numbers. In the wider meeting I showed what we had learned: which groups liked it and which did not. I did not soften the main finding. The team kept one part of the design. The outcome was your result.”

“A weekly number rose sharply for one kind of device and nothing in the product explained it. I looked at raw events for a few users and saw the same start event sent twice, a moment apart. A new app version retried the message even when the first one had arrived. I worked with the engineers to add a unique id to each event, and I removed duplicates in the reporting table. Two reports had already gone out with the high number. I sent a correction with the old and new values and a short note on the cause. The result was your result.”

“I would start from the purpose. If the number is used to judge whether people chose a title and stayed, a few seconds should not count. One option is a minimum share of the running time. Another is total hours watched divided by running time, which gives a number comparable between a short film and a long one. I would test each rule on edge cases: a short cartoon that a child repeats, a series where one episode is sampled, a film paused and finished over a week. Whichever rule we choose, I would publish it and not change it quietly.”

“From the play events I would first build one row per member who finished the first episode, with the time they finished and their country. That group is my denominator. Next I build one row per member with the earliest time they started the second episode. I left join the first set to the second, so members who never continued stay in the result. Then I mark a member as continued when the start falls within seven days after the finish. Grouping by country, I count continued members and divide by all members. I would exclude members who finished less than a week ago.”

“Clicks alone can mislead. A striking poster may win clicks and then disappoint, so people stop after a few minutes. My main metric would be plays of the title that continue past a meaningful point, per member who saw the poster. As a guardrail I would watch total viewing, since a poster may only pull attention from other titles. Members must be split at random and always see the same poster. I would decide the length of the test in advance and not stop the day one side looks ahead. I would also check whether the result differs by country.”

“First I would turn regularly into a rule, for example at least a few active days in each of the three earlier months. From the daily viewing table I count active days per member per month for that earlier period, and keep members who meet the rule in every month. Then I check the most recent month with a not exists condition: no viewing rows for that member in that period. I would join to the accounts table to drop members who cancelled or joined recently, since they are a different story. I would check a handful of members by hand before sharing the list.”

“A doubling with no cause is more likely a data problem than good news. I would split the number by app version, country and hour to see where it begins. If it starts with one app version at one time, I suspect a logging change. Then I look at raw events for a few accounts: are sessions impossibly long, or are events repeated? I would compare with an independent source, such as the amount of video data served to those devices. If that did not double, viewing did not double. Meanwhile I mark the dashboard number as under review so nobody quotes it.”

“Viewing data is a record of real choices, not opinions. People say they want documentaries and then watch comedy, and the data shows what they did. I find that honest and hard to read well, because a stop can mean boredom or a crying baby. In your project I learned how much the definition of a metric shapes the decision made from it. I would like to work where tests are a normal way to decide and where the questions are about human attention. I want to get better at designing experiments and at explaining uncertain results plainly.”