Snowflake has three layers that scale on their own: storage, compute and cloud services. The idea is that data lives in one place and any number of independent compute clusters can read it.
+-------------------------------------------------+ | Cloud services: metadata, optimizer, security, | | transactions, query compilation | +-------------------------------------------------+ | Compute: virtual warehouses (BI, ETL, data sci) | | WH_BI WH_ETL WH_ML | +-------------------------------------------------+ | Storage: micro-partitions in cloud object store | +-------------------------------------------------+
Storage layer
When you load a table, Snowflake converts it into compressed, columnar files called micro-partitions and stores them in S3, Azure Blob or GCS. You never see or manage these files. Storage is billed by size, separately from compute.
Compute layer
A virtual warehouse is a cluster of machines that runs queries. You pick a size (X-Small, Small, Medium and so on, each step doubling the power and the credits per hour) and you can have as many warehouses as you want. They all read the same storage, and they do not compete with each other. A warehouse can be suspended when idle, and then it costs nothing.
Cloud services layer
This is the brain: it keeps table metadata (including per-partition min and max values), parses and optimizes queries, handles authentication and access control, and coordinates transactions. It runs all the time, and heavy use of it beyond a daily allowance can show up on the bill.
What the separation gives you
Say finance runs a heavy month-end report while analysts run dashboards. They use two warehouses on the same tables, so neither slows the other, and no data is copied. If the ETL warehouse needs to be four times bigger for an hour, you resize it for that hour and then shrink it back.
In an interview, name the three layers, say what each one owns, and give one example of isolation, like BI and ETL warehouses sharing one copy of the data. That example is what most interviewers want to hear.