Keep "Column description" for Virtual Dataset

[emmanuel-bos]

Hi,
I'm new to Superset and still discovering it. Seems great ! I'm connecting it to a PostgreSQL datasource. This PostgreSQL instance has comments on each column set with the non standard SQL command "Comment on ". I was happily surprised to find that when creating a physical dataset, those comments were retrieved in the column descrition. I expect this to be very useful when I will test the MCP server functionality. But when I create un Virtual Dataset from on join query in the SQL Lab, I see the SQL comments are not retrieved in the column description. It seems pity as I will have to fill it myself to interact to best use the MCP functionality. Is there a reason to that or is it just a bug ?

[dosubot[bot]]

Hi @emmanuel-bos! I'm Dosu and I'm helping the Apache Superset team.

This is expected behavior, not a bug — but it's a limitation worth knowing about.

For physical datasets, Superset uses SQLAlchemy's inspector (database.get_columns()) to fetch table metadata, which includes the database-level comment field. A recent PR (#34359) added logic in fetch_metadata() to read that comment field and store it as the column description. PostgreSQL's COMMENT ON COLUMN values are surfaced through this inspector path.

For virtual datasets, metadata is obtained differently: Superset runs your SQL query with a LIMIT 1 and derives column metadata from the cursor description [1]. The cursor description provides the column name and data type, but not any database-level comments. Since your JOIN query can alias columns from multiple tables, there's no general mechanism to trace each output column back to its source table and retrieve the original COMMENT ON COLUMN value.

The gap is architectural: physical datasets go through the inspector (which knows about column comments), while virtual datasets go through query execution (which doesn't). The external_metadata() method on SqlaTable shows this branching clearly — it calls get_virtual_table_metadata() when SQL is present, and get_physical_table_metadata() otherwise [2].

Workaround: You can manually fill in column descriptions in the virtual dataset editor (Dataset → Edit → Columns → description field), since TableColumn has a description column in the ORM [3]. While tedious for many columns, this is currently the path to get descriptions visible for MCP and other tooling. Filing a GitHub issue to request automatic propagation of column comments for virtual datasets would be a great way to get it tracked.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.