Got rejected after the OpenAI loop, Sora system design was rough
Got rejected after the OpenAI loop, Sora system design was rough
Went through OpenAI's full process for a senior SWE role recently, 6 rounds total counting the phone screen, and just got the rejection call so figured I'd write up what I saw since I couldn't find a ton of detail before going in. For context I've got 6 YOE, mostly infra work, in case that colors how I read any of this.
Phone screen was system design plus coding. SD was a payment system, the coffee shop point of sale version rather than anything online, they hand you a doc to read through yourself and the interviewer wants you thinking about idempotency from the jump. I spent a good chunk of time on how you'd record per bank transactions without ever touching the card data directly, since apparently they deliberately don't store which bank a card routes to on their end so you can't just look it up. Coding was the infected plants matrix problem, basically an epidemic spread simulation. First three parts went fine, ran out of time before the immunity twist showed up.
Onsite had 5 more rounds. First coding round was GPU Credits, a resource allocation type problem, I'd practiced something similar beforehand so it felt manageable and I finished with a few minutes to spare.
System design was Sora, the video scheduling service, and this is where I think I lost the round, ngl. Sounds intimidating from the name but once you're in it, it's really a job scheduler with a GPU pool underneath. Got the basic architecture down fast, queue plus workers plus a metadata store, but the interviewer kept pushing on what happens when a worker dies mid job. Restart from a checkpoint or eat the full retry, how do you avoid wasting a GPU that was most of the way done, how do you re-queue a preempted job without double notifying the user. I had answers but they weren't as sharp as they should've been, and in hindsight I needed a cleaner checkpointing story before jumping straight to the notification layer.
Cross function round was new to me, hadn't seen it mentioned anywhere before I went in. Interviewer was a PM, not an engineer, and it was framed around working with non-technical stakeholders. Felt more like a standard collab story round once we actually got into it though.
Behavioral was the usual leadership and conflict stuff plus the OpenAI specific one, why here versus anywhere else. Deep dive was on a past infra project, and the interviewer didn't have much background in that particular stack so we ended up talking more generally about tradeoffs than getting into the weeds with him.
Heard back about two weeks later, rejected. No real specifics beyond something like strong in some areas, not enough signal in others, which I'm guessing points at the SD round given how deep they went on failure handling there. Honestly kind of a bummer since everything else felt solid, but if you get something similar, worth spending real prep time on worker failure scenarios beyond just the happy path design, that seemed to be where they wanted more depth out of me specifically.
Anyone else get the Sora question recently, curious how deep it usually goes for other people.