Snowflake AI Data Cloud Features and Architecture
The three layers
Snowflake's architecture separates storage (cloud object storage holding micro-partitions), compute (virtual warehouses), and cloud services (metadata, optimizer, transactions, security, authentication). Micro-partitions are immutable, columnar, compressed units of roughly 50–500 MB of uncompressed data. Snowflake records min/max values and distinct counts for each column in each partition, which is exactly what makes partition pruning possible without indexes or user maintenance.
Warehouses run from X-Small to 6X-Large, each step roughly doubling compute and credit burn. Scale up (bigger warehouse) for a single slow, heavy query; scale out (multi-cluster warehouse) for many concurrent queries. Warehouses also cache table data on local SSD - that cache belongs to the warehouse and does not follow you to another one. The result cache is account-wide, lasts 24 hours, and is invalidated when the underlying data changes.
Traps
- A multi-cluster warehouse does not make one query faster; it reduces queueing.
- Metadata-only operations (SHOW, DESCRIBE, DDL, unfiltered COUNT(*)) do not require a running warehouse.
- Resizing helps a query that spills; it does not fix a table that cannot prune.
Rule of thumb
If the Query Profile shows "Bytes spilled to remote storage", the warehouse is too small - double the size before rewriting SQL.