Choosing a Data Warehouse

Vendor-published benchmarks are constructed to win, and independent ones rarely reflect your query patterns. On realistic mixed workloads the three major platforms land close enough that performance is not the deciding factor for most organisations.

What actually differs is who can operate the platform, what happens to cost when usage grows, and whether machine learning is a first-class workload or an occasional one.

Snowflake asks the least of its operators. If your data team is analyst-heavy and SQL-first, it removes almost all infrastructure decisions, at the price of consumption costs that require active governance.

Databricks asks considerably more and returns considerably more control. If you employ data engineers who are comfortable with cluster configuration and Spark, and machine learning is a real workload, it is the stronger platform.

BigQuery removes capacity planning entirely. Inside Google Cloud, with spiky and unpredictable workloads, its serverless model is the simplest thing to operate, provided someone enforces partitioning before the bytes-scanned bill arrives.

All three platforms bill on consumption, which means an inefficient query is a billable event rather than a blocked one. The organisations that stay within budget assign an owner to warehouse cost in the first month, not the first crisis.

Practical controls: query timeouts, mandatory partitioning on large tables, separate compute for ad-hoc exploration, and a monthly review of the most expensive queries. None of these are platform features; all of them are the difference between a predictable bill and a board conversation.

Iceberg and Delta support has reduced storage-level lock-in meaningfully across all three platforms. The remaining lock-in is at the semantic and pipeline layer, where transformation code, access policies and orchestration accumulate.

That is where migration cost actually lives, and no table format solves it.