How Snowflake's layered architecture separates storage, compute, and services

Snowflake uses a unique multi-cluster shared data architecture that combines elements of shared-disk and shared-nothing architectures. It consists of three independently scalable layers: database storage, query processing (virtual warehouses), and cloud services. This design allows storage and compute to scale independently, enabling elastic performance and concurrency.

1 · Learn the must-know

• 1The storage layer holds all data in compressed, columnar, micro-partitioned format in cloud storage (S3, Azure Blob, or GCS), fully managed and optimized by Snowflake.
• 2The compute layer consists of virtual warehouses (MPP clusters) that independently process queries; multiple warehouses can access the same data simultaneously without contention.
• 3The cloud services layer handles authentication, metadata management, query parsing/optimization, security, and infrastructure management, coordinating across the entire system.
• 4Virtual warehouses can be resized, suspended, resumed, or set to auto-suspend/auto-resume, and multi-cluster warehouses can scale out automatically to handle concurrency spikes.
• 5Because compute and storage are separated, you only pay for storage used and compute time consumed, and scaling one does not require scaling the other.
• 6Snowflake runs natively on AWS, Azure, and GCP, and its architecture allows for seamless cross-cloud and cross-region features like replication without changing the underlying data model.

2 · Check your understanding