Snowflake

Snowflake runs the same warehouse on AWS, Azure, and Google Cloud, which is exactly why cloud-agnostic organizations standardize on it. Storage is centralized and billed on average compressed usage; compute comes as independently sized “virtual warehouses” that suspend when idle and bill per second, so a nightly dbt run and an analyst’s ad-hoc queries never contend for the same resources.

The warehouse model is simple to reason about: nine T-shirt sizes from X-Small (1 credit/hour) to 6X-Large, with credit consumption roughly doubling at each step. Auto-suspend and auto-resume are on by default, and Snowflake’s own docs are blunt that larger is not necessarily faster for small queries — right-sizing down is usually the first cost win. Multi-cluster warehouses, which add clusters automatically under concurrency pressure, require Enterprise edition, as does extending Time Travel beyond one day.

Time Travel is the operational feature teams lean on most: every account gets 1 day of retention by default (Standard edition caps there; Enterprise allows up to 90 days on permanent objects), which enables UNDROP for accidentally deleted tables, AT/BEFORE queries against historical states, and zero-copy clones at a point in time. After retention expires, data falls into Fail-safe, where you can no longer query it yourself — plan retention deliberately for tables your pipelines mutate.

For tracking data specifically, it is a first-class target for Snowplow loaders, Segment and RudderStack warehouse destinations, and the source that Hightouch and Census activate from. Semi-structured event payloads live comfortably in VARIANT columns — Snowflake ingests JSON, Avro, ORC, Parquet, and XML into an internal columnar representation that queries efficiently without predefined schemas — and get flattened into typed models in dbt. Very large individual payloads are the exception: docs suggest restructuring once a single value approaches ~128 MB.

The discipline it demands is cost governance. The 60-second minimum per warehouse start means chatty automation that wakes a warehouse for one-second queries pays for a full minute each time; aggressive auto-suspend can therefore backfire on frequent short workloads. Auto-suspend tuning, warehouse sizing, and resource monitors are not optional at scale, and per-credit price varies by edition and region, so model costs against your actual contract rather than list examples.

Use it when you need one warehouse across clouds, workload isolation between pipelines and analysts, or data sharing with partners without building an API. Look past it when you are all-in on Google Cloud with GA4 (BigQuery is the shorter path), you need open-source or self-hosted infrastructure, or your workload is real-time event serving rather than batch analytics.