Google Backend Developer Interview Questions

“I would split it into a store of known addresses, a scheduler and many fetch workers. The scheduler groups addresses by website and allows each site only a few requests in a given time, following the rules the site publishes for crawlers. How often a page is revisited depends on how often it changed before: a news front page often, an old article rarely. Workers fetch, store the content and extract new links. I would detect duplicate content with a hash of the page, and cap depth on each site to avoid endless generated pages. If a worker dies, its addresses return to the queue after a time-out.”

“Uploads from phones break often, so I would accept the file in chunks and let the app resume from the last confirmed chunk. When complete, the original goes to durable storage and a message goes on a queue. Workers split the video into short segments and convert each segment to several qualities in parallel, so a long video does not wait on one machine. A failed segment is retried alone. I would make the video watchable as soon as a low quality is ready, and add higher qualities as they finish. A status record lets the uploader see progress. Other checks run as separate jobs from the same queue.”

“Saving whole documents would mean the second save wipes out the first person's words. So each client sends small operations, like insert these letters at this position, along with the version it was based on. The server puts the operations in one order. If an operation was based on an older version, the server adjusts its position to account for the edits that came in between, then applies it and sends it to everyone. Each client does the same adjustment for its own pending edits. This way all copies reach the same text. For someone who was offline, the client replays its queued operations on reconnect.”

“I would build an inverted index: for each word, a sorted list of the pages that contain it, with positions if we want phrases. A query with two words intersects two lists, which is fast because both are sorted. The lists are compressed, since page numbers close together can be stored as small gaps. The index is too large for one machine, so I would split it by page across many machines, query them all and merge the results. Rewriting a huge index for each changed page is too costly, so new and changed pages go into a small fresh index that is searched as well, then merged into the main one.”

“I worked on your project, a service that had two caches and a fallback path to an old database. Most of our incidents came from those parts disagreeing with each other. I measured and found the second cache saved very little time, and the fallback path had not been used for real in a long while. I proposed removing both. To be safe, I first logged every time they were used, then turned them off behind a flag for two weeks before deleting the code. Afterwards the service had your result, and new team members could understand it in a day.”

“I had written a batch job for your project and sent it for review feeling done. The reviewer asked me to restructure it so each step could be retried alone, which meant rewriting half. My first reaction was annoyance, because it worked. I waited a day, then asked for a short call to understand the concern. She showed me a past incident where a similar job failed at the last step and had to rerun for hours. I made the change. Two months later my job failed mid-way and resumed from the failed step. I now ask for design review before writing the code.”

“People may be missing important mail, so I would first roll the configuration back, even before I fully understand the fault. Then I would measure the damage: which users, and which messages were marked as spam during that window that would not have been before. Stopping the fault is not enough. Those messages should be moved back to inboxes, or the users told clearly. Then I would find why the change passed checks. I would want configuration changes released to a small share of users first, with an alarm on any jump in the rate of mail marked as spam.”

“The problems I enjoy most are about finding the right small piece of data in a very large store, quickly. Search, mail and video all have that at their centre. In my current work on your project, I learned how much an index design or a storage layout changes what is possible. Here those choices matter at a size where a small waste is multiplied enormously, so engineers have to be careful and measure everything. I want that discipline. I also like that the products are used by ordinary people every day, so the work is easy to explain to my family.”