went through the coinbase senior loop (L5 equivalent) earlier this year. wanted to write up the system design round specifically because there's not much out there that's current.
the round is 60 minutes. one interviewer, sometimes a second observer. they give you a prompt in the first few minutes and expect you to drive. the prompts i saw leaned toward fintech/payments territory. mine was something like: design a system to process millions of crypto transactions per second with exactly-once delivery semantics. not surprising for coinbase, but you need to actually know what "exactly once" means in distributed systems and not just say the words.
what they care about, from what i could tell:
clarifying questions first. they noticed if you jumped straight to drawing boxes. spend 5 minutes on scale, consistency requirements, failure modes. my interviewer visibly relaxed when i asked about eventual vs. strong consistency before designing anything.
crypto-adjacent knowledge helps, but isn't required. i was asked follow-up questions about transaction idempotency that any senior distributed systems person should know. you don't need to know what a mempool is, but you should know how to design an idempotent payment processing layer.
tradeoff discussions are scored heavily. pick an approach, explain why, then genuinely engage with alternatives when pushed. i got pushback on my kafka choice and had to defend or pivot -- they wanted to see the thinking, not the final answer.
time management: they move fast. don't spend 30 minutes on the data model. sketch it, acknowledge the tradeoffs, move on. i ran out of time on the scaling section and felt like that hurt me, though i still got the offer.
one thing i didn't expect: they asked about observability late in the round. have an answer ready for how you'd monitor this thing in production. not just metrics, but specific failure conditions you'd alert on. that felt like an L5 vs L4 signal question.
overall the bar felt real but fair. L5 at coinbase is probably E5 at meta or SWE III at google in terms of what's expected. maybe slightly lower bar on pure leetcode but higher bar on actual distributed systems depth.