Five rounds, spread over two days: one coding round, two design rounds, two behavioral rounds.
For the coding round you were allowed to use AI, so I picked GPT 6.1 Sol to help me write it. That AI could edit the code directly, so I didn't need to copy and paste anything.
Part 1:
Given a grid where 0 is a fire station, -2 is a lake, and -1 is a residential area. There are multiple fire stations in the grid. Find the Manhattan distance from each residential area to the nearest fire station. Fire trucks can't go through the lake. I first told the interviewer my idea: run a BFS from each fire station to compute its distance to every residential area, then take the minimum. He asked what the complexity was. I said that if the number of fire stations is fixed, it is O(N), where N is the size of the grid. If the number of fire stations also varies, it is O(M*N), where M is the number of fire stations. He didn't say anything, so I told the AI my idea and asked if there was a better solution. The AI said I could use a multi-source BFS: put all the fire stations into the queue together with distance 0, then run the BFS, which is only O(N). I said OK, let's implement that, and the AI implemented it for me. I checked the AI-generated code line by line, and then the interviewer gave me Part 2:
The government pays for the distance from the fire trucks to the residential areas. There is a base cost, and at distance 0 there is some discount off the base, at distance 1 a different discount, at distance 2 I think it is just the base, at distance 3 you charge some extra percentage, and so on, and at the end there is a cap. Compute the cost from the fire trucks to each residential area. I brainstormed with the AI the same way and then had it implement it. At the end the interviewer asked what the space complexity of this implementation was. I said O(N), because it keeps adding residential areas to the queue without ever popping them, it only moves a pointer, so in the end the whole grid ends up in the queue. It could be changed to actually pop, to save space. I also told the AI that after visiting each node it could do queue = queue[1:] to pop it out (I implemented it in Go). The AI changed it accordingly.
I saw there was a Part 3, but there wasn't much time left, so they didn't let me continue.
Design 1 was with a Distinguished Engineer from their department. I needed to design a mechanism to figure out who needs to stamp a PR. I had to draw a diagram and also write pseudocode for the key logic. I assumed up front that it was one big monorepo, so I came up with a rough design and wrote pseudocode in half Go, half Python. Then he asked: what if the commit rate is very high, and you never asked me whether it is a monorepo, what if there are many repos, some big and some small? How do you make sure the system performs well? So I added things like a persistent cache directory and pre-warming for big repos. In the end he thought it was fine. This Distinguished Engineer came over from Google, where he was also a Distinguished Engineer.
Design 2: design a CI scheduling system. It had to be triggered from GitHub, find a suitable worker (CPU/GPU), and then handle what happens if there aren't enough workers during peak hours. I more or less followed the Buildkite design. I honestly don't know this area well, so I just followed the Hello Interview steps one by one so I wouldn't screw up. The interviewer just listened coldly from start to finish. At the end he let me ask questions, and I asked: I heard LinkedIn's culture has been getting worse in recent years, do you feel that? He said he thinks LinkedIn was too relaxed a few years ago and is less relaxed now, so people complain, but the workload is only what is normal at Google (he also came from Google).
Of the two behavioral rounds, one was a chat with the hiring manager, who is a director. Besides asking about the projects I had done, this guy kept talking tech with me without end, and dragged a one-hour interview into an hour and a half. He did most of the talking. Almost every time I mentioned a project he would cut in to say he had run into the same problem too.
In the end, as expected, I failed Design 2, but the director and I got along so well that they still decided to give me an offer.
Everyone here says LinkedIn's culture is going downhill, but I felt this Distinguished Engineer and the director still know their technology. Would a team like that be better?
Discussion
Loading comments…