Data virtualization

Data virtualization provides a single logical interface — usually queried with SQL — over data that physically remains in its original, separate systems: a warehouse, an operational database, a file store, a SaaS API. A query engine sitting in front of these sources translates an incoming query into the calls each underlying system understands and combines the results, so the user never has to know where the data actually lives.

This differs from ETL or ELT, which physically copy and consolidate data into one warehouse before it can be queried; virtualization instead queries the sources live, trading the cost and latency of a copy step for always-current results and no duplicated storage. The trade-off is performance and complexity: a virtualized query is often only as fast as its slowest underlying source, and pushing filters and joins down efficiently to each source is a hard optimization problem.

It is used when copying data everywhere is impractical or against policy — sensitive data that must stay in place for compliance reasons, or systems too numerous or fast-changing to keep fully replicated — and it underpins architectural patterns like data fabric and parts of data mesh. A common pitfall is expecting virtualization to match the performance of a purpose-built warehouse on heavy analytical workloads; it generally does not, and is better suited to moderate-volume, cross-system lookups than to large-scale aggregation.