Snowflake

Snowflake is a cloud-native data platform built from scratch to run on cloud infrastructure (AWS, Azure, Google Cloud) without adapting legacy on-premises database architecture. Its defining architectural innovation is the strict separation of three independent layers: centralised storage, multi-cluster compute (virtual warehouses), and a global metadata service.

Virtual warehouses are independent compute clusters that can be started, paused, resized, and auto-suspended in seconds. Multiple warehouses can read from the same underlying data simultaneously with no contention — a BI dashboard and an ETL pipeline can run concurrently without one blocking the other. You pay per second of warehouse runtime, with credits billed by warehouse size (X-Small to 6X-Large, doubling credits per size).

Credit pricing varies by edition and cloud region. On-demand rates: Standard ~$2/credit, Enterprise ~$3/credit, Business Critical ~$4/credit. Annual pre-purchased credits reduce the effective rate by 30–40%. Storage is billed separately at ~$23/TB/month (on-demand) or ~$40/TB/month for 30-day failsafe retention.

Time Travel and Fail-Safe: Snowflake automatically retains historical versions of data for 0–90 days (configurable by edition), allowing queries against past states and accidental delete recovery. An additional 7-day fail-safe period is maintained by Snowflake for disaster recovery.

Data sharing and Marketplace: Snowflake's data sharing model lets organizations share live, governed data with partners or customers without copying it. The Snowflake Data Marketplace has thousands of third-party datasets (weather, financial, demographic) consumable directly in SQL.

Snowpark extends Snowflake's compute to run Python, Java, and Scala code natively inside Snowflake — enabling data engineers and ML engineers to write transformations and ML pipelines in familiar languages without moving data out.

Snowflake Cortex (2024–2025) adds LLM-powered SQL functions (COMPLETE, SUMMARIZE, TRANSLATE, SENTIMENT) that run directly in SQL against Snowflake-hosted models, plus integration with external model providers.

Iceberg Tables (generally available 2024) allow Snowflake to read and write Apache Iceberg format stored in the customer's own object storage (S3, ADLS, GCS), reducing vendor lock-in for storage.

Snowflake editions — Standard, Enterprise, Business Critical, VPS — add features for compliance (HIPAA, PCI DSS, SOC 2), multi-cluster warehouses, materialized views, and data masking.