> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# Select Tables

> **Prerelease documentation**
>
> This Kumo Studio information is prerelease and is not yet supported. Stay tuned for the Docker and Kubernetes releases.

After your connector is set up, select the source tables to ingest into Kumo.

## Connecting a Table

1. Navigate to **Tables** in the side menu and click **Add Table**.
2. Next, select a **Source**:
   * **Connector** – Choose an existing data connector.
   * [Local Upload](local-data-upload) – Upload a local CSV or Parquet file.

## Selecting a Connector Source

If you choose **Connector** as the source type:

1. Select a connector from the drop-down.
2. Kumo loads available tables in the **Source Table** drop-down.
3. *(For Native Databricks users)* – Enter the **Schema Name** to populate the table list.

![Add New Table page with Connector selected as the source and a connector chosen](/sdgm/_fern-img/8fc1bdb90f4ab9ad2bf34cfe7fe08da5fc3d572f07174fe0bab1bb0d0edc62ea.webp)

## Column Selection

Select columns that meet the following criteria for best results:

* **Clean:** Remove fake or synthetic data, predictions from other ML models, and any data known to be unreliable or inaccurate. Also remove columns whose definitions have changed over time - for example, where an attribute ID maps to different things in different periods.
* **Relevant and Mutually Exclusive:** Remove columns with duplicated, similar, or irrelevant information.
  The larger the graph size (the total column count across all tables in a graph), the higher the compute cost.
* **Complete:** Cover the full history for the timeframe in question (for example, a complete record of purchases rather than only a user's first or last purchase).
  If this results in an oversized dataset, provide a compressed version that tracks changes in aggregate metrics over time (for example, per day, week, or month).

> 🚧 Using wrong or unnecessary columns can degrade model performance (due to noisy pQuery inputs or data leakage) and increase pQuery training costs.

## Column Preprocessing

After selecting a table, you can define **column types and preprocessing steps** to ensure proper data handling. For details, refer to [Column Preprocessing](column-preprocessing).

## Debugging & Data Validation

To verify a table’s schema and metadata:

1. Navigate to the **Tables** page.
2. Click on the table name to view details.
3. Click on a **column** to view the detailed **Stats**.

This provides:

* **Column statistics** (for example, missing values, cardinality, distributions).
* **Sample rows** for verification.

Kumo computes this information from a smaller sample of the dataset. Complete statistics are available after full data ingestion during model training.

![Table detail panel for DimCustomer, listing each column's data type, semantic type, and share of missing values](/sdgm/_fern-img/098790ce2c0df4acd0c2f6dde29eef0a9f318cf1d8d175b2affc1d58604601b3.webp)

### Handling Missing Data

* Kumo treats **blank entries** as missing values.
* It does **not** automatically recognize special strings (`"NaN"`, `"none"`, `"N/A"`) as missing.
* For numeric columns, missing values are often filled with placeholder values (e.g., `-1`). These should be cleared or left blank if you want them treated as missing.