Skip to main content

workflow.variables.types

Workflow variable types for artifact storage and push results.

ModelArtifact Objects​

@dataclass
class ModelArtifact()

A packaged model artifact ready for upload.

Both the raw model package and the serving-ready deployable artifact are represented as ModelArtifact instances. Packaging must be complete before passing to the pusher — packaging is an assembler-time concern (e.g. a Ray worker with GPU access).

Attributes:

  • path - Absolute local filesystem path to the packaged artifact file or directory.
  • metadata - Typed metadata forwarded to the model registry at registration time. Subclass ModelMetadata to add provider-specific fields.

Example:

from michelangelo.workflow.variables.metadata import ModelMetadata meta = ModelMetadata(training_framework="xgboost", deployable=True) artifact = ModelArtifact(path="/tmp/model", metadata=meta) artifact.metadata.training_framework 'xgboost'

FeaturePackageArtifact Objects​

@dataclass
class FeaturePackageArtifact()

A feature package preceding a model's feature-computation stage.

Assemblers fuse this into the model's own schema/sample data to produce the end-to-end serving contract (see fuse_e2e_schema, build_e2e_sample_data).

Attributes:

  • path - Absolute local filesystem path to the feature package.
  • metadata - Typed metadata describing the feature package's schema and sample data.

Example:

from michelangelo.workflow.variables.metadata import ( ... FeaturePackageMetadata, ... ) package = FeaturePackageArtifact( ... path="/tmp/features", metadata=FeaturePackageMetadata() ... ) package.metadata.schema FeatureSchema(input_schema=[], feature_store_features_schema=[], derived_features_schema=[])

AssembledModel Objects​

@dataclass
class AssembledModel()

A trained model transmitted between workflow tasks.

raw_model is required. deployable_model is optional — omit it for models that are not packaged for serving (e.g. research checkpoints or models where ModelMetadata.deployable is False). When absent, the pusher skips the deployable upload and sets ModelPushResult.deployable_artifact_uri to None.

Packaging is the assembler's responsibility. The pusher only uploads and registers pre-packaged artifacts.

Attributes:

  • raw_model - Raw model package (weights + sample data) intended for offline validation and reproducibility.

  • deployable_model - Optional serving-ready bundle (e.g. Triton config + weights) intended for deployment to a model server. None when the model has not been packaged for serving.

  • feature_package - Optional feature package fused into the deployable model's end-to-end schema/sample data during assembly. None when the model has no upstream feature-computation stage.

    Example (with deployable):

    artifact = ModelArtifact(path="/tmp/model.ubj") assembled = AssembledModel( ... raw_model=artifact, ... deployable_model=artifact, ... ) assembled.raw_model.path '/tmp/model.ubj'

    Example (raw only):

    assembled = AssembledModel(raw_model=ModelArtifact(path="/tmp/model.ubj")) assembled.deployable_model is None True

NativeTransformResult Objects​

@dataclass
class NativeTransformResult()

The result of the native transform task (tabular_native_transform).

Contains the transformed datasets and the PyTorch transform module (model). For incremental-training flows, the transform spec and feature stats are stored on model.metadata (transform_spec and feature_stats) so downstream tasks (assembler, pusher) can persist them as a transform checkpoint for a future incremental run.

Attributes:

  • transformed_datasets - Mapping of dataset name (e.g. "train", "validation", "test") to its transformed DatasetVariable. Datasets that had no transform spec applied (empty inputs) are passed through unchanged.
  • model - The materialized transform module, wrapped as a ModelVariable, or None when the transform spec produced no layers (e.g. an empty spec).

Example:

result = NativeTransformResult( ... transformed_datasets={"train": DatasetVariable.create(None)}, ... ) result.model is None True

PusherResult Objects​

@dataclass
class PusherResult()

The outcome of a single plugin execution.

Attributes:

  • name - Artifact name from PusherPluginConfig.name.
  • plugin - Plugin name that was invoked (e.g. "model_plugin").
  • success - True if the plugin completed without error.
  • value - Plugin-specific return data. Empty dict when success is False.
  • error - Human-readable error description when success is False. None when success is True.

Example:

result = PusherResult( ... name="model", ... plugin="model_plugin", ... success=True, ... value={"model_name": "clf-v1", "version": "1"}, ... ) result.success True result.error is None True