Snowflake Architecture
Everything on the exam sits on top of this part: warehouses, caching, pricing, Time Travel, cloning, and sharing are all consequences of three architectural decisions Snowflake made in 2012, so learn the decisions and half the exam becomes derivable instead of memorizable. Domain 1 carries 31% of the exam. The signature trap is number collision, near-identical figures planted side by side: 16 MB micro-partition, 128 MB VARIANT, 100 to 250 MB load files, 10 clusters, 10% cloud services. The reflex to build: for any behavior, name the layer that did it and the cache that answered, without hesitating.
The Snowflake Architecture
This chapter covers 31% of the COF-C03 exam directly (the "Snowflake AI Data Cloud Features and Architecture" domain) and indirectly underpins every other domain. Read it slowly. Every later chapter refers back to the mental model built here.
- Shared-disk: all compute nodes read one central disk, and queue for it.
- Shared-nothing: each node owns a slice of the data on its own disk, so adding a node means moving data.
- MPP (massively parallel processing): one query split across many machines working at the same time.
- OLAP: analytics, big scans over many rows. OLTP: many small transactions, like placing an order.
- ACID: a transaction happens in full or not at all.
- DML changes rows (
INSERT,UPDATE,DELETE). DDL changes objects (CREATE,ALTER,DROP). - RBAC (role-based access control): privileges go to roles, and users get roles. Part 3 covers it.
- SSO: logging in through your company's login page.
Why Snowflake Exists
Snowflake is a data platform delivered as a fully managed SaaS service, built natively for the cloud, that separates storage from compute so each can scale, be paid for, and fail independently. It runs on AWS, Azure, and GCP, and there's no hardware, software installation, indexing, or tuning of physical structures for the customer to manage.
The problem it was built to solve
To understand why Snowflake looks the way it does, you need the two architectures that came before it, because the contrast explains every design choice.
Shared-disk architecture (classic Oracle RAC style): every compute node reads from one central storage layer. Data is in one place, which is convenient, but all nodes fight over the same disk and the same locks. Add more compute and you don't get more throughput, you get more contention. Storage is the bottleneck.
Shared-nothing architecture (Teradata, Netezza, early Redshift, Hadoop): each node owns a slice of the data on its own local disk. Queries run in parallel across slices, which scales well, until it doesn't. Compute and storage are welded together: to add storage you must add compute (and pay for it), to add compute you must redistribute data across nodes, which can take hours or days on a large cluster. Resizing is disruptive, and a node failure means its data slice is temporarily gone.
Snowflake's answer, the multi-cluster shared data architecture: keep the single central data copy from shared-disk, but make it infinitely scalable cloud object storage (S3, Azure Blob, GCS). Keep the parallel compute from shared-nothing, but make compute clusters (virtual warehouses) stateless, disposable, and independent of the data. The result is a hybrid: one copy of data, any number of independent compute clusters reading it simultaneously without contention, coordinated by a brain layer on top.
SHARED-DISK
Node
Node
Node
STORAGEOne disklocks + I/O queue
- nodes queue on one disk
- one consistent copy
SHARED-NOTHING
Node
Slice 1own disk
Node
Slice 2own disk
Node
Slice 3own disk
- resize = move data
- parallel scans
SNOWFLAKE · SHARED DATA
WH A
WH B
WH C
STORAGEObject storageone copy of the data
- add compute in seconds
- no warehouse waits on another
- One data copy, always consistent
- Every node fights over the same disk and locks
- More compute → more contention, not more throughput
- Storage is the ceiling on scale
- Parallel scans across independent slices
- Compute and storage welded together
- Resizing means redistributing data: hours, sometimes days
- One node down, one slice of data temporarily gone
- One data copy, on infinitely scalable object storage
- Any number of stateless compute clusters read it at once
- Add or drop compute in seconds, no rebalancing
- A cluster failing loses nothing: data was never on it
Why this matters in production
A concrete scenario you'll live daily as a data engineer. The BI team runs dashboards all day on a Medium warehouse. At 02:00, ELT loads hammer a Large warehouse. At month end, finance runs a brutal reconciliation on an X-Large. In a shared-nothing system these three workloads would queue on the same cluster, and the ELT load would freeze the dashboards.
In Snowflake they run on three separate warehouses reading the same tables at the same instant, with zero copies of the data and zero interference. When finance is done, their warehouse suspends and billing stops that second. Nobody files a ticket asking IT to "give finance more power for month-end," because finance already has their own warehouse, sized for exactly that job, paying for exactly that job.
COMPUTE LAYER · ONE BILL PER WAREHOUSE
- reads
WAREHOUSEBI_WH · Mediumdashboards all daybilled to BI
- writes
WAREHOUSEELT_WH · Largeloads at 02:00billed to Data Eng
- reads
WAREHOUSEFINANCE_WH · X-Largemonth-end closebilled to Finance
STORAGE LAYER · ONE COPY
TABLESSALES and ORDERS tablesmicro-partitions in cloud object storage
What stateless compute actually buys
Because a warehouse holds nothing but a cache, three operational moves become trivial that were projects in shared-nothing systems.
Resizing is a metadata change plus node provisioning, seconds instead of a weekend of data redistribution. Failure recovery is replacement: a dead node held no data slice, so Snowflake swaps it and running queries retry. And isolation becomes an org chart decision rather than a capacity fight. Give finance FINANCE_WH, tag it, and their month-end reconciliation appears as a line on their cost center. Chargeback in Teradata took a consultant; here it's a warehouse name.
What "SaaS" means concretely on the exam
- No hardware to provision, no software to install, patch, or upgrade. Snowflake releases updates weekly with no downtime and no customer action.
- No vacuuming, no statistics jobs, no partitioning schemes to define, and no index you have to build for pruning to work. Snowflake maintains its own storage structures (micro-partitions, covered later in this part).
- You cannot access the underlying cloud storage or VMs directly. Data is only reachable through SQL via Snowflake.
- Snowflake runs entirely on public cloud infrastructure. There is no on-premises Snowflake and no hybrid deployment.
- What stays with you: choosing warehouse sizes and suspend settings, designing roles and grants, and deciding what to load and how long to keep it. Snowflake handles everything else above.
Options such as "Snowflake can be deployed on-premises for regulated industries" or "pruning needs an index you build and tune" are false. Snowflake has no on-prem option (Virtual Private Snowflake is still Snowflake-hosted, just in a dedicated environment), and pruning works from micro-partition metadata, with no index required. Optional secondary indexes (CREATE INDEX, Enterprise edition) and hybrid-table indexes exist, but Snowflake maintains them. The usual tuning levers are clustering keys, search optimization, and warehouse sizing, each a managed service.
- Snowflake = shared-disk's single data copy + shared-nothing's MPP compute, minus the drawbacks of both.
- Storage and compute scale independently and are billed independently.
- Zero management of hardware, indexes, or physical partitions. Ever.

