2.1 Snowflake AI Data Cloud Features and Architecture Overview

The Three-Layer Architecture

Snowflake's defining design is a hybrid of shared-disk and shared-nothing architectures. Like a shared-disk system, it uses a single, centralized repository for persisted data that every compute node can reach. Like a shared-nothing system, it processes queries using MPP (massively parallel processing) compute clusters where each node stores a portion of the working set locally. This blend delivers the simplicity of shared data with the performance of distributed processing.

The platform is built as three physically separated but logically integrated layers:

The critical exam concept is that these layers scale independently. You can resize compute (a warehouse) without touching storage, and storage grows automatically without you provisioning compute. Many warehouses can query the same table simultaneously with no copy of the data and no resource contention — this is the foundation of Snowflake's concurrency model.

The Cloud Services Layer

The cloud services layer is the coordinator that ties everything together. It runs on compute instances Snowflake provisions from the cloud provider, and it is multi-tenant — shared across accounts. Its responsibilities include:

• Authentication and federated identity (SSO, key-pair, OAuth)
• Access control (role-based access control, RBAC)
• Infrastructure and metadata management (tracking every micro-partition's min/max values and statistics)
• Query parsing and optimization (building the query plan, pruning partitions)
• Transaction management (ACID, multi-statement transactions)
• Security, including encryption key management
• The result cache, which serves identical query results for 24 hours without using a warehouse

Because metadata (row counts, min/max per column, distinct counts) lives here, queries like SELECT COUNT(*), MIN(), MAX(), and many SELECT of metadata can be answered without a running warehouse — the answer comes straight from the metadata store. Cloud services usage is generally free; Snowflake only bills cloud-services credits when daily consumption exceeds 10% of the daily compute (warehouse) credits, which is rare in practice.

The Centralized Storage Layer

When you load data, Snowflake reorganizes it into its internal optimized, compressed, columnar format and writes it to cloud object storage (Amazon S3, Azure Blob, or Google Cloud Storage depending on your account's cloud platform). Customers never see these files directly and cannot access them with object-storage tools — all access is through SQL. Snowflake fully manages organization, file size, structure, compression, metadata, and statistics. Storage is billed separately from compute, on a flat rate per terabyte of compressed data per month.

Snowflake as a SaaS Platform

Snowflake is delivered purely as Software-as-a-Service. There is no hardware (virtual or physical) to select, install, configure, or manage, and no software to install or patch. Snowflake handles all ongoing maintenance, management, upgrades, and tuning. The service runs entirely on cloud infrastructure — you cannot run Snowflake on a private, on-premises setup. Releases are deployed transparently with no downtime: Snowflake ships weekly behavioral-change and new-feature releases plus periodic patch releases.

This SaaS model is why the SnowPro Core exam frames so many questions around which built-in feature or default behavior applies rather than how to configure infrastructure. The platform abstracts the operational complexity, and your job — and the exam's focus — is knowing which capability to invoke and how Snowflake behaves by default.

Why This Domain Carries the Most Weight

At approximately 31% of the SnowPro Core blueprint, "Snowflake AI Data Cloud Features and Architecture" is the single largest scored domain. Expect questions on the three layers, what each does, virtual-warehouse behavior, micro-partitions and clustering, caching, storage features like Time Travel and Fail-safe, supported clouds and regions, editions, and the AI Data Cloud / Marketplace concept. Master this domain first: it has the highest point value and underpins everything tested in the SQL, performance, and data-loading domains.

How the Layers Cooperate on a Single Query

It helps to trace one query through all three layers, because exam stems often hinge on which layer did the work:

1. A client submits SQL through Snowsight, a driver, or a connector. The request lands in the cloud services layer, which authenticates the user, checks RBAC privileges, and parses the statement.
2. The optimizer (still in cloud services) builds a plan and uses micro-partition metadata to prune partitions the query cannot need. If the answer is already cached or derivable from metadata, no warehouse runs at all.
3. Otherwise the plan is handed to a virtual warehouse. Its MPP nodes read the surviving micro-partitions from centralized storage (or from the warehouse's warm local cache) and execute the work in parallel.
4. Results return to the client, and the result set is stored in the result cache for 24 hours.

The Object Hierarchy

A second foundational concept is Snowflake's container hierarchy, which the exam tests directly:

The fully qualified name of an object is therefore DATABASE.SCHEMA.OBJECT. Two special schemas — INFORMATION_SCHEMA and PUBLIC — are created in every database automatically. Understanding this hierarchy is the prerequisite for the access-control and SQL domains, where you grant privileges level by level.