Load data using Mosaic Streaming

This article describes how to use Mosaic Streaming to convert data from Apache Spark to a format compatible with PyTorch.

Mosaic Streaming is an open source data loading library. It enables single-node or distributed training and evaluation of deep learning models from datasets that are already loaded as Apache Spark DataFrames. Mosaic Streaming primarily supports Mosaic Composer, but also integrates with native PyTorch, PyTorch Lightning, and the TorchDistributor. Mosaic Streaming provides a series of benefits over traditional PyTorch DataLoaders including:

  • Compatibility with any data type, including images, text, video, and multimodal data.

  • Support for major cloud storage providers (AWS, OCI, GCS, Azure, Databricks UC Volume, and any S3 compatible object store such as Cloudflare R2, Coreweave, Backblaze b2, etc.)

  • Maximizing correctness guarantees, performance, flexibility, and ease of use. For more information, view their key features page.

For general information about Mosaic Streaming, view the Streaming API documentation.


Mosaic Streaming has been pre-installed into all versions of the Databricks Runtime 15.2 ML and higher.

Load data from Spark DataFrames using Mosaic Streaming

Mosaic Streaming provides a straightforward workflow for converting from Apache Spark to the Mosaic Data Shard (MDS) format which can then be loaded for use in a distributed environment.

The recommended workflow is:

  1. Use Apache Spark to load and optionally preprocess data.

  2. Use streaming.base.converters.dataframe_to_mds to save the dataframe to disk for transient storage and/or to a Unity Catalog volume for persistent storage. This data will be stored in the MDS format and can be further optimized with support for compression and hashing. Advanced use cases can also include the preprocessing of data using UDFs. View the Spark DataFrame to MDS tutorial for more information.

  3. Use streaming.StreamingDataset to load the necessary data to memory. StreamingDataset is a version of PyTorch’s IterableDataset that features elastically deterministic shuffling, which enables fast mid-epoch resumption. View the StreamingDataset documentation for more information.

  4. Use streaming.StreamingDataLoader to load the necessary data for training/evaluation/testing. StreamingDataLoader is a version of PyTorch’s DataLoader which provides an additional checkpoint/resumption interface, for which it tracks the number of samples seen by the model in this rank.

For an end-to-end example, see the following notebook:

Simplify data loading from Spark to PyTorch using Mosaic Streaming notebook

Open notebook in new tab