Snowflake Interview Questions

An array holds people and cakes; assign each person a cake. The trap is that it looks local and is not: you cannot answer for the queried person alone, because the assignment has to be solved for everyone first. One report works through the smallest example that breaks the obvious answer — with two people and two cakes side by side, the second person is nearest the first cake, so the first person ends up with the second. The same report notes asking several AI assistants afterwards and getting wrong answers from all of them. The other half of the round is a space-optimisation follow-up on the base scan. This is the free preview — full statement, follow-ups, and a solution run against its test cases.

The most reported question in the catalog, and it arrives in several costumes: course prerequisites, task completion order, build-and-test pipeline ordering. Cycle detection is the baseline; the follow-ups are where the round is decided, and reports name them consistently — produce the order, not just a yes or no; handle courses with individual durations; and say how you would run independent work in parallel. One report describes the interviewer asking for the layer-by-layer output specifically because the context was services starting in dependency order.

Design a scheduler that runs one-off and recurring jobs across a fleet, survives losing the node that owns the timer, and does not fire the same job twice. The reports push on the parts that are easy to hand-wave: which component is allowed to decide a job is due, what happens to a job that is mid-run when its worker dies, and how a paused job resumes on schedule. The walkthrough is explicit that a lease alone does not make leader election safe — a paused holder still believes it leads — and that advancing the next fire time only after a success stops a failing job recurring forever.