Metaxy

Metaxy is a metadata layer for multimodal Data and ML pipelines that manages and tracks feature versions, dependencies, and data lineage across complex computational graphs.

Core Concepts

Feature Definitions

To define a feature, create a class inheriting from mx.BaseFeature with a FeatureSpec metaclass argument:

import metaxy as mx


class Video(
    mx.BaseFeature,
    spec=mx.FeatureSpec(
        key="media/video",
        id_columns=["video_id"],
        fields=["audio", "frames"],  # Logical fields: describe data contents for versioning
    ),
):
    # Metadata columns: stored in the metadata store, tracked by metaxy
    video_id: str
    path: str
    duration: float
    height: int
    width: int

Important distinction: fields in FeatureSpec are logical field specs that describe the data contents for versioning and lineage tracking. Class attributes are metadata columns stored in the metadata store. They serve different purposes and should not overlap.

To add dependencies between features, use the deps parameter with FeatureDep. To specify field-level lineage (for partial data dependencies), use FieldSpec with FieldDep or FieldsMapping.

Data Versioning

Metaxy automatically tracks sample versions and propagates changes through the dependency graph. To trigger recomputation when code changes, set code_version on FieldSpec:

fields = [
    mx.FieldSpec(key="embedding", code_version="2"),  # Bump to invalidate downstream
]

Metadata Stores

To configure a metadata store, create a metaxy.toml file or use programmatic configuration:

config = mx.MetaxyConfig(
    stores={"dev": mx.StoreConfig(
        type="metaxy.ext.polars.handlers.delta.DeltaMetadataStore",
        config={"root_path": "/tmp/metaxy"},
    )}
)
with config.use() as cfg:
    store = cfg.get_store("dev")

Supported backends: DuckDB, ClickHouse, BigQuery, LanceDB, Delta Lake.

Feature Graph

To visualize and manage the feature dependency graph, use the CLI:

mx graph render            # Terminal visualization
mx push --store dev        # Push graph to store

CLI

Metaxy provides a CLI (metaxy or mx alias) for managing features, metadata, and migrations:

mx list features --verbose     # List features with dependencies
mx graph render                # Visualize feature graph
mx metadata status --all-features  # Check metadata freshness (expensive!)
mx mcp                         # Start MCP server for AI assistants

Multi-Environment & Feature Locking

Cross-project dependencies typically resolve automatically via Python packages — if project B depends on project A as a Python dependency, its features are discovered at import time. Feature locking (mx lock) is needed for multi-environment setups where projects cannot be installed into each other (e.g., separate deployment environments). In that case, use mx push to publish definitions to a shared store, and mx lock to fetch them into a metaxy.lock file. Set locked = true (or METAXY_LOCKED=1) in production to enforce version consistency.

For configuration patterns and CLI usage, see examples/configuration.md and examples/cli.md.

Testing

To test features in isolation, use context managers to avoid polluting the global registry:

import pytest
import metaxy as mx


@pytest.fixture
def metaxy_env(tmp_path):
    with mx.FeatureGraph().use():
        with mx.MetaxyConfig(
            stores={"test": mx.StoreConfig(
                type="metaxy.ext.polars.handlers.delta.DeltaMetadataStore",
                config={"root_path": str(tmp_path / "delta_test")},
            )}
        ).use() as config:
            yield config

Map Datatype (Experimental)

To enable native Arrow Map column support (recommended for stores that support it), set enable_map_datatype = true in metaxy.toml. Requires the polars-map package. See https://docs.metaxy.io/stable/guide/concepts/metadata-stores/#map-datatype

When working with Map columns, use metaxy.utils.collect_to_polars, metaxy.utils.collect_to_arrow, or metaxy.utils.switch_implementation_to_polars to materialize or convert frames. These utilities preserve Map column types that would otherwise be lost with standard Narwhals backend conversions. See https://docs.metaxy.io/stable/guide/concepts/metadata-stores/#map-datatype

Examples

For complete code examples, see:

• examples/feature-definitions.md - Feature classes with dependencies and field-level deps
• examples/configuration.md - TOML and programmatic configuration
• examples/metadata-stores.md - Store operations
• examples/testing.md - Test isolation patterns
• examples/cli.md - CLI command reference

Documentation

For comprehensive documentation: https://docs.metaxy.io/stable/

Key pages:

• Quickstart: https://docs.metaxy.io/stable/guide/quickstart/quickstart/
• Feature Definitions: https://docs.metaxy.io/stable/guide/concepts/definitions/features/
• Data Versioning: https://docs.metaxy.io/stable/guide/concepts/versioning/
• Metadata Stores: https://docs.metaxy.io/stable/guide/concepts/metadata-stores/
• Projects: https://docs.metaxy.io/stable/guide/concepts/projects/
• External Features: https://docs.metaxy.io/stable/guide/concepts/definitions/external-features/
• CLI Reference: https://docs.metaxy.io/stable/reference/cli/
• API Reference: https://docs.metaxy.io/stable/reference/api/