Managing datasets — creation from databases and files, schema browsing, metadata, and data preview
A dataset in the Workbench is a defined view of data. Physical datasets are backed by data found in database connections — tables and views in PostgreSQL, MySQL, MSSQL, or MS Access — or DuckDB tables created by file uploads. Virtual datasets are federated SQL views spanning one or more physical datasets across different storage systems. Recipe outputs also become datasets. Datasets are the bridge between your data sources and the recipe engine.
Datasets are created through five paths: Database browse — select a table or view from a connected database; File import — upload a CSV, JSON, or Excel file; Virtual dataset — define a federated SQL query spanning one or more physical datasets; Recipe output — the result of running a recipe becomes a dataset; SQL query — save a query result as a named dataset.
When you connect a database, the Workbench automatically creates datasets for every table and view in the default schema. You can refine these auto-datasets — add descriptions, rename them, or delete the ones you don't need. The underlying database table is never modified by dataset operations.
Every dataset has a preview panel showing the first 100 rows, column names, types, and null counts. The preview is a live query — it always reflects the current state of the underlying data, even for database-backed datasets. For large datasets (over 1M rows), preview uses a random sample.
Datasets carry metadata: name, description, source (database/file/recipe), creation date, creator, row count estimate, and column profiles. Metadata is editable — you can add descriptions, tags, and custom notes to any dataset. Metadata is searchable through the Algorithm Discovery feature.
Datasets follow the same sharing model as recipes. They can be private (owner only), shared with specific users or groups, or public (all users on the instance). Group admins control dataset visibility for their group members.