Snowflake Architecture
Snowflake · Snowflake Fundamentals
Theory
Snowflake uses a multi-cluster shared data architecture made of three independent layers. The Storage layer keeps every table as compressed, columnar micro-partitions in cloud object storage. The Compute layer is made of virtual warehouses, clusters that read that storage to run queries. The Cloud Services layer sits on top and handles login, access control, metadata, query optimization and transactions.
Because the layers are separate, each one scales on its own. A finance team and a data science team can each run their own warehouse against the same orders table at the same time, and neither slows the other down, while the data itself is stored only once.
Why we use it
Traditional warehouses tie storage and compute to the same machines, so more data or more users means a bigger, costlier cluster and teams fighting for resources. Splitting the layers lets you add compute without moving data and isolate workloads from each other.
When to use it
Use this model to reason about almost every Snowflake question: cost (compute vs storage billing), performance (warehouse size, pruning), concurrency (separate or multi-cluster warehouses) and features like cloning and Time Travel, which come from the storage design.
The Storage Layer: Where Data Actually Lives
Every table you create is automatically compressed, converted to a columnar format, and split into micro-partitions — immutable chunks of roughly 50–500MB — inside cloud object storage (S3, Azure Blob, or GCS). You never touch these files directly or manage their layout; Snowflake handles compression, organization, and the statistics used for optimization entirely on its own.
The Compute Layer: Independent Virtual Warehouses
Queries run on virtual warehouses — isolated MPP (massively parallel processing) clusters sized from X-Small to 6X-Large. Because each warehouse is its own cluster, ten teams can query the same tables through ten different warehouses at the same time with zero contention, something a single shared-cluster database can't offer. Compute is billed per-second while a warehouse runs, independent of how much data it reads.
The Cloud Services Layer: The Coordination Brain
Sitting above both storage and compute, Cloud Services handles everything that isn't raw data crunching: authenticating users, enforcing role-based access, parsing and optimizing SQL, tracking metadata and statistics, and managing transactions. It runs on Snowflake-managed infrastructure, so this coordination layer is always available even when no warehouse is running.
Why This Split Is the Core Idea
Separating storage from compute — and separating both from the services that coordinate them — is what lets Snowflake scale each piece independently, avoid resource contention between teams, and offer features like instant zero-copy cloning and Time Travel with no extra infrastructure to manage.
How it works
- You submit a SQL query
- Cloud Services authenticates you and checks your role's privileges
- Optimizer uses micro-partition metadata to plan the query
- Is a warehouse running for this session? Yes: Use the running warehouse. No: Auto-resume it (if allowed) or fail.
- Warehouse reads only the needed micro-partitions from storage
- Cloud Services returns the result to you
Example
-- Snowflake's three layers, illustrated conceptually:
-- Storage: compressed, columnar micro-partitions in cloud object storage
-- Compute: independent virtual warehouses that read that storage
-- Services: parsing, optimization, security, metadata (coordinates both)
SHOW WAREHOUSES; -- lists your compute layer
SHOW DATABASES; -- your storage layer, organized into DBs/schemas/tablesKey points
- Storage layer holds compressed columnar micro-partitions, managed entirely by Snowflake
- Compute layer runs as independent virtual warehouses that can be paused and resized
- Cloud Services handles metadata, security, optimization and transactions
- Compute and storage are billed separately
- Many warehouses can read the same data at once without contention