Snowflake Architecture: Three-Layer Design Explained

Snowflake uses a multi-cluster shared-data architecture that differs from both traditional shared-disk and shared-nothing designs. Storage, compute, and cloud services are fully decoupled, and each layer scales independently. This article walks through the role of each layer, micro-partitions, the cache hierarchy, and the Cloud Services billing rule — the exam-relevant essentials plus what matters in production.

Three-Layer Architecture Overview

Snowflake's architecture is built from three layers. Each layer scales and fails over independently, freeing users from infrastructure management so they can focus on analytics.

┌─────────────────────────────────────────────┐
│         Cloud Services Layer                │
│  Auth / Access Control / Query Optimization │
│  Metadata Management / Transactions         │
├─────────────────────────────────────────────┤
│         Compute Layer (Virtual Warehouses)  │
│  WH-1 (XS)  │  WH-2 (L)  │  WH-3 (XL)    │
│  Independent clusters, isolated from each   │
├─────────────────────────────────────────────┤
│         Storage Layer                       │
│  Cloud Object Storage (S3 / Azure Blob / GCS)│
│  Micro-partitions (columnar, compressed)    │
└─────────────────────────────────────────────┘
      

Storage Layer

Data is stored on cloud-provider object storage (AWS S3, Azure Blob Storage, Google Cloud Storage) in a columnar format. Users never interact with storage directly — Snowflake handles all data management automatically.

How Micro-Partitions Work

Table data is automatically split into units called micro-partitions.

Micro-partition immutability is what enables Time Travel (point-in-time queries) and Fail-safe (disaster recovery). When you run UPDATE, the affected partition is recreated as a new version, and the previous version is retained for the Time Travel window.

Compute Layer

Virtual Warehouses handle query execution. Each warehouse is an independent compute cluster — workloads in one warehouse don't impact others.

Warehouses support multi-cluster scaling, automatically adding clusters as concurrent query count grows. Auto-suspend stops idle warehouses, and auto-resume starts them again on the next query.

Cloud Services Layer

The 'brain' of Snowflake. Users never see this layer directly, but it handles these critical responsibilities:

• Authentication & access control: User authentication, RBAC, and network policy enforcement
• Metadata management: Table definitions and micro-partition statistics
• Query optimization: Query parsing, optimization, and pruning decisions
• Transaction management: Guarantees ACID transactions
• Security: End-to-end AES-256 encryption with automatic key rotation

Cloud Services Billing (the 10% adjustment)

Example: if warehouses consume 200 credits in a day, the Cloud Services threshold is 20 credits. If Cloud Services consumed 25 credits, only 25 - 20 = 5 credits are billed on top.

The Three-Cache Hierarchy

Snowflake has three caches that improve query performance and reduce cost.

Result Cache is the most cost-effective — it returns results without starting a warehouse. It shines for repeated identical queries like scheduled BI dashboard refreshes. But Result Cache is invalidated as soon as DML runs on the underlying table.

End-to-End Query Processing

1. User submits SQL
2. Cloud Services Layer parses and optimizes the query
3. If Result Cache has a matching result, it returns immediately without a warehouse
4. If an aggregation query can be answered from Metadata Cache, it returns without a warehouse
5. Otherwise the query runs on the specified warehouse
6. Micro-partition pruning reads only the relevant data
7. Data already in Warehouse Cache skips storage access
8. Results are returned to the user and stored in Result Cache

Snowflake Editions and Architecture Features

Exam Focus Areas

On the SnowPro Core exam, the Architecture domain is about 25% of the test — the single biggest area. These are the points you should expect to be tested on:

• Role of each layer and which processing happens where
• Concrete benefits of decoupled storage and compute
• Micro-partition properties (size, immutability, columnar, automatic management)
• Differences across the three caches (location, warehouse start required or not, invalidation conditions)
• The Cloud Services Layer 10% billing rule
• Feature differences across editions (multi-cluster, Time Travel window, Tri-Secret Secure)

Check Your Understanding

Frequently Asked Questions