Architecture Deep Dive
Snowflake · Snowflake Fundamentals
Theory
Each layer has a clear job. Storage keeps data once, as immutable micro-partitions plus metadata such as min/max values per column. Compute is a set of virtual warehouses; each warehouse is its own cluster with its own CPU, memory and local disk cache, and it can be started, paused or resized at any time. Cloud Services parses SQL, checks privileges, builds the plan from metadata, coordinates transactions and keeps the result cache.
This is why there is no contention between teams: the ETL warehouse loading data and the BI warehouse running dashboards share only storage, never CPU. A traditional shared-nothing database would make both workloads compete on the same nodes.
Why we use it
Knowing what each layer does tells you which knob to turn. Slow scans point to compute size or pruning, high login or metadata activity points to Cloud Services, and data volume affects storage cost only, not query capacity.
When to use it
Use it when designing warehouse layout (one warehouse per workload), diagnosing performance or cost, and answering exam questions that ask which layer performs a task such as authentication, caching or pruning.
Storage, Compute, Cloud Services in Detail
Storage holds the data once, compute (virtual warehouses) reads it on demand, and Cloud Services coordinates everything in between — parsing SQL, planning queries, managing transactions, and enforcing security.
Why No Contention
Because each layer scales independently, multiple warehouses can read the same underlying storage simultaneously without locking each other out, which is the core reason Snowflake supports high concurrency without manual tuning.
How This Differs From Traditional Databases
A traditional database ties compute and storage to the same server, so scaling one means scaling (and paying for) both. Snowflake's separation is what makes elastic, pay-per-second compute possible.
How it works
- Query arrives at Cloud Services
- Parse SQL and check the role's privileges
- Build a plan using metadata and prune micro-partitions
- Is the identical result already in the result cache? Yes: Return the cached result with no warehouse. No: Send the plan to a virtual warehouse.
- Warehouse scans storage, using its local disk cache where it can
- Result is returned and cached by Cloud Services
Example
-- Multiple virtual warehouses can query
-- the same data simultaneously without blockingKey points
- Multiple warehouses can query the same data without contention
- Pause and resume compute independently to control cost
- Cloud Services uses metadata to optimize and prune automatically
- The result cache lives in Cloud Services, not in a warehouse
- Each warehouse has its own local disk cache, lost when it suspends