Snowflake Architecture

Snowflake est un data warehouse cloud-native lancé en 2014 (founders ex-Oracle), pionnier de l'architecture séparant compute et storage, run on AWS / Azure / GCP. IPO 2020 (largest software IPO ever). Architecture révolutionnaire popularisée standard moderne data warehousing.

Architecture 3-couches :
(1) **Storage layer** — données stockées dans S3/Blob/GCS en format columnar propriétaire (FDN — Snowflake's micro-partitions). Auto-managed (partitioning, clustering, micro-partitions). Pay per storage TB/month.
(2) **Compute layer** — **Virtual Warehouses** (VWs) — clusters MPP indépendants. Multiple warehouses peuvent attaquer same data simultanément sans interférence. Auto-suspend après inactivity, auto-resume on query. Pay per credit consumed (depend on warehouse size XS/S/M/L/XL/2XL/.../6XL).
(3) **Cloud services layer** — orchestration, query optimization, metadata, transaction management, security, sharing.

Forces uniques :
(1) **Separation compute/storage** — scale independently. Pause compute (no cost) while keeping data.
(2) **Multi-cluster warehouses** — auto-scale concurrent users (each cluster handles independent workload).
(3) **Zero-copy cloning** — `CREATE TABLE clone CLONE source;` instant clone via metadata, no data copy until divergence. Massive for dev/test environments.
(4) **Time Travel** — query data as-of past timestamp (default 1 day, up to 90 days enterprise) — undo accidental deletes/updates.
(5) **Fail-safe** — additional 7-day retention for disaster recovery.
(6) **Secure Data Sharing** — share live data with external partners without copies (Snowflake Marketplace).
(7) **Multi-cloud + cross-region replication** native.
(8) **Snowpark** — Python/Java/Scala dans Snowflake (UDFs, stored procs, dataframes API).
(9) **Iceberg tables, Polaris Catalog** — open table format support, interop with non-Snowflake tools.
(10) **Cortex AI** — built-in LLM functions (`SNOWFLAKE.CORTEX.COMPLETE()` etc.).

Use cases : enterprise data warehousing, customer analytics, financial reporting, marketing analytics, supply chain. Adopted Capital One, Adobe, Pfizer, Sony, BlackRock, etc.

Vs BigQuery / Redshift : (1) Snowflake very strong on multi-cloud, sharing, ease of use ; (2) BigQuery serverless purer (no warehouse sizing), tighter Google ecosystem ; (3) Redshift cheapest for steady AWS workloads with RIs, less elastic. Compétences DEA-C01, DP-203.