Metaxy is a metadata layer for multimodal Data and ML pipelines. Metaxy tracks lineage and versioning across complex computational graphs for multimodal datasets. Metaxy can cache every single sample and scale to handle millions of them.
Metaxy manages metadata while data typically lives elsewhere:
┌─────────────────────────────────┐ ┌─────────────────────────┐
│ Metadata (Metaxy) │ │ Data (e.g., S3) │
├──────┬──────────┬──────┬────────┤ │ │
│ ID │ path │ size │version │ │ 📦 s3://my-bucket/ │
├──────┼──────────┼──────┼────────┤ │ │
│ img1 │ s3://... │ 2.1M │a3fdsf │ ────────>│ ├─ img1.jpg │
│ img2 │ s3://... │ 1.8M │b7e123 │ ────────>│ ├─ img2.jpg │
└──────┴──────────┴──────┴────────┘ └─────────────────────────┘
The feature that makes Metaxy stand out is the ability to track partial data dependencies and detect prunable updates — updates that don't trigger change propagation through certain paths in the dependency graph because they modify fields that aren't dependencies of those downstream features. For example, updating audio upstream of a face recognition step allows pruning the face recognition branch since it only depends on video frames. This problem is specific to multimodal pipelines and doesn't typically emerge in traditional data engineering.
Metaxy's goal is to provide a standard instrument for any kind of multimodal (or purely tabular) incremental pipelines, standardizing dependency specification, versioning, partial data dependencies, and manipulations over metadata. Or, in short, to be a universal glue for incremental data pipelines.
Metaxy is very reliable and is fanatically tested across all supported Python versions and platforms 1.
Documentation
Read the docs to learn more.
Metaxy is published in the Journal of Open Source Software.
Installation
Install Metaxy from PyPI:
uv add metaxy
Using Metaxy
Metaxy is highly pluggable and generally can be used with any kind of incremental pipelines, storage, metadata storage, and dataframe libraries.
Metaxy provides integrations with popular tools such as Dagster, Ray, ClickHouse, DeltaLake, SQLModel.
The full list can be found here.
Blog Posts
Contributing
See CONTRIBUTING.md.
