Datasets

A dataset is a table rowlab has read in and keeps as Parquet. Your source, a file on disk, a URL or an object in S3, is read once and never written to.

Sources

CSV, JSON (an array of objects), JSONL and Parquet, plain or gzipped. The format is detected from the content, so the file name doesn’t matter. Sources:

• Upload from the browser, or the API’s multipart endpoint.
• A URL, fetched by the server. On a shared server, private addresses are refused.
• A path on the server’s disk, locally. On a shared server this is off unless the operator turns ROWLAB_SERVER_FILES on.
• S3, through a connection (Settings → Storage) that holds the bucket’s credentials. Members use connections without seeing their secrets.

The tree

A run never changes a dataset. It makes a new one, with every row of the original and the answer columns added, and records which dataset it came from and how. Names say where a dataset sits: reviews → reviews__1 → reviews__1.2. The datasets list shows the tree, and each dataset’s page says what made it: the SQL, the tasks and the models.

This is deliberately not versioning. rowlab is a workspace between your storage and your storage; keep the source and the exports where you keep data, and delete datasets here when you are done.

Profile

Each column has a profile: its type, min and max, average, number of distinct values and the share of empty cells. It is the first thing to read when writing a prompt, since it says which columns hold the text, which the ids, and where the gaps are.

Limits

Files up to ROWLAB_MAX_FILE (1 GB by default) are read, with ROWLAB_DUCKDB_MEMORY (1 GB) of memory per query and the rest spilled to disk. Raise them in Configuration for bigger data.