FeatureQL language

FeatureQL is a SQL dialect for data transformations with three design principles:

Abstracted: Separates transformation logic from storage and execution engines. You can swap data sources or execution backends without rewriting transformations.

Atomic: Works at the column level instead of SQL's table level. This makes dependencies explicit and enables granular lineage tracking.

Functional: Enforces pure transformations that can be composed together. Each transformation is self-contained and predictable.

FeatureQL transpiles to native SQL for DuckDB, Trino, BigQuery, and Datafusion, so it works with your existing database infrastructure.

What problems it solves

Reusable business logic: Instead of copying the same revenue calculation across 50 queries, define it once as a feature. When the calculation changes, update one place, not 50.

Consistent metrics: Everyone uses the same definition of "active customer" or "churn rate" from the feature catalog. No more meetings about why numbers don't match.

Rapid experimentation: Compose features like building blocks. Test variations with VARIANT(), swap calculation methods instantly, and iterate without rewriting entire queries.

One language, all use cases: The same features power dashboards, ML models, real-time APIs, and operational systems. No translation between analytics SQL and application code.

Simpler complex queries: Working with nested data (like all orders per customer) is natural with FeatureQL's array operations. What takes 100 lines of SQL with multiple CTEs becomes 10 clear lines.

Faster debugging: Features are testable functions. You can test each feature in isolation instead of debugging a 500-line SQL query.

Better collaboration: Data teams publish verified features that analysts and engineers reuse. New team members are productive immediately instead of learning years of tribal knowledge.