A data science team wants to read a large governed lakehouse table into notebooks for feature engineering. They ask whether they should copy the table to a local disk first for speed. What guidance fits the watsonx.data consumption model?

The correct guidance is to query the table through a supported engine from the notebook, filtering and aggregating server-side rather than copying it. Notebooks are a legitimate consumer of lakehouse data, and the supported pattern is to connect to an engine and let it do the scanning, filtering and aggregation, returning only the result set. That keeps the read governed, auditable and current, and it usually moves far less data than copying the table would. Copying the table locally once per sprint is incorrect because the performance premise is weak and the governance cost is real. A per-sprint copy is stale for most of its life, and features engineered from it silently drift from the governed source. Requesting read access to the bucket and loading the Parquet files directly into the notebook is incorrect because it bypasses the catalog and the access policies attached to it. It also requires the notebook to resolve Iceberg's current snapshot correctly, which is the engine's job. Having an administrator export the table nightly to a shared drive the team can mount is incorrect because it turns an ad hoc copy into a standing one, and a nightly export to a shared drive is an ungoverned dataset by construction - no policy, no lineage, no revocation. For the exam, associate consumption with querying in place through a supported interface, whatever the client happens to be.