Went through the Snowflake L5 senior SWE loop earlier this year. The system design round is the one people get least prepared for and it's also the one that carries the most weight at this level. Sharing what I learned.
The prompt I got was something like: "Design a query execution engine that can run analytical queries across hundreds of terabytes of data concurrently for thousands of tenants."
Yeah. It's very on-brand.
A few things I noticed about what they were actually looking for:
They want you to know Snowflake's actual architecture conceptually. Not to parrot their marketing, but the ideas: separation of storage and compute, virtual warehouses as isolated compute clusters, the fact that cache is per-warehouse. If you can ground your design in those principles you come across as someone who's already thought about the problem domain.
Concurrency and isolation matter a lot. I spent time on multi-tenancy: how do you prevent one heavy query from starving another tenant. The interviewer lit up when I brought up resource queuing and auto-suspend.
Data format and compression. Columnar storage, why Parquet/ORC-style formats help analytical workloads, skip indexes. You don't need to go deep on implementation but you should be able to name the concepts and explain why.
They'll push on scale. The "what breaks first" question came up. I talked about metadata overhead when you have millions of micro-partitions, and the interviewer confirmed that was a real problem they deal with.
My prep: I read the Snowflake whitepaper (2016, still relevant for fundamentals), did a few general distributed systems design exercises, and spent an afternoon thinking specifically about multi-tenant analytical databases vs. OLTP systems. That was probably worth more than grinding more LeetCode at this stage.
The round was 60 minutes. About 10 min requirement clarification, 35 min design, 15 min deep-dive into one component. Go into it with a structured approach but be willing to pivot based on what the interviewer finds interesting.