Netflix Software Engineer Interview Questions

“I was sure that showing more items on the first screen would help people find something faster. We tested it on a share of users. People with the fuller screen took longer to choose and left more often. My first thought was a bug in the test, so I checked the group split and the event logging. Both were fine. I then watched a few session recordings and saw people scrolling back and forth, unable to decide. We kept the old layout and I wrote up what we learned. The outcome was your result.”

“Our app retried failed requests three times, and during an outage those retries made things worse. Nobody owned the retry logic, so I took it. Before changing anything I read past incident notes and asked two engineers from other teams what they would worry about. I wrote a short note with the plan: fewer retries, longer waits between them, and a limit across the whole app. I shared it, waited a day for objections, and shipped it behind a switch. The result was your result. If it had gone badly, the note made clear it was my call.”

“The film is stored as short pieces, and each piece exists in several qualities. Before each request the player picks one. I would use two signals. One is recent download speed, averaged so a single slow piece does not cause panic. The other is how many seconds are already buffered. With a full buffer I can risk higher quality. With a nearly empty one I drop quality quickly, because a pause is worse than a soft picture. I would step up slowly and step down fast, and avoid changing on every piece, since constant switching is also unpleasant to watch.”

“I would treat each title as the set of profiles that watched most of it. Two titles are similar when many of the same profiles watched both. Raw overlap favours very popular titles, since they overlap with everything, so I would divide by the sizes of both sets. Computing every pair at request time is too slow. A batch job does it regularly and stores the best matches per title. Then I remove what the viewer has already seen and what is not available in their country. New titles have few viewers, so for them I would fall back on genre and cast.”

“I would index each title under its original name, its translated names, and its cast. Text is normalised first, such as lower case and accents removed, with rules per language. Viewers type slowly on a television, so the search must work on prefixes after two or three letters. For spelling mistakes I would match terms within a small edit distance. The country filter must be part of the search itself. If I filter afterwards, a query may return a full page of titles the viewer cannot watch and then show nothing. Results are ranked by match quality, then popularity in that country.”

“The video file is stored encrypted, and a separate licence on the device holds the key and the rules, such as when it expires. At play, the app checks the licence without needing a network. The device clock can be changed by the user, so I would rely on the protected clock that the media system provides, or compare with the last known server time. When the device is online again, licences renew quietly in the background. Before a download expires I would warn the viewer in the downloads list, because discovering it on the plane is the worst moment.”

“I would first check that the viewing record was saved and marked as finished. If the viewer stopped during the credits, the film may count as unfinished, which is a rule problem, not a bug. If the record is fine, I would look at when the suggestions were computed. They are likely built by a batch job that ran before last night, then cached. The fix is not to run the heavy job more often. I would keep the batch results and apply a light filter at request time that removes recently finished titles, using fresh viewing data.”

“I am drawn to problems where engineering quality is felt directly by a person. In streaming, a good decision in the player means a film that does not stop, and a good decision in suggestions means someone finds a story they love in a language they did not expect. I have worked on your project, where I learned to test ideas with real users and accept the answer. I want to work where that habit is normal and the scale makes small gains matter. I would like to go deeper into playback, which I only know from the outside.”