Hey @anhnnguyen, you can add libraries a few ways when building a notebook-based ETL pipeline:

The best practice, scalable approach to add libraries across multiple workloads or clusters is to use Policy-scoped libraries. Any compute that uses the cluster policy you define will add any dependencies to the cluster at runtime. Check this: Policy-scoped libraries 

If you only need to add libraries to a single workload or cluster, use compute-scoped libraries.
Check this: Compute-scoped Libraries