You are viewing documentation for Kubeflow 0.5

This is a static snapshot from the time of the Kubeflow 0.5 release.
For up-to-date information, see the latest version.

Component Specification

Definition of a Kubeflow Pipelines component

This specification describes the container component data model for Kubeflow Pipelines. The data model is serialized to a file in YAML format for sharing.

Below are the main parts of the component definition:

  • Metadata: Name, description, and other metadata.
  • Interface (inputs and outputs): Name, type, default value.
  • Implementation: How to run the component, given the input arguments.

Example of a component specification

A component specification takes the form of a YAML file, component.yaml. Below is an example:

name: xgboost4j - Train classifier
description: Trains a boosted tree ensemble classifier using xgboost4j

inputs:
- {name: Training data URI}
- {name: Rounds, type: Integer, default: '30', help: Number of training rounds}

outputs:
- {name: Trained model, type: XGBoost model, help: Trained XGBoost model}

implementation:
  container:
    image: gcr.io/ml-pipeline/xgboost-classifier-train@sha256:b3a64d57
    command: [
      /ml/train.py,
      --train-uri, {inputValue: Training data URI},
      --rounds,    {inputValue: Rounds},
      --out-model, {outputPath: Trained model},
    ]

See some examples of real-world component specifications.

Detailed specification (ComponentSpec)

This section describes the ComponentSpec.

Metadata

  • name: Human-readable name of the component.

  • description: Description of the component.

  • metadata: Standard object’s metadata:

    • annotations: An unstructured key value map stored with a resource. External tools can set this property to store and retrieve arbitrary metadata. Annotations are not queryable and should be preserved when modifying objects. See more information in the Kubernetes user guide.

    • labels: A map of string keys and values that can be used to organize and categorize (scope and select) objects. May match selectors of replication controllers and services. See more information in the Kubernetes user guide.

Interface

  • inputs and outputs: Specifies the list of inputs/outputs and their properties. Each input or output has the following properties:

    • name: Human-readable name of the input/output. Name must be unique inside the inputs or outputs section, but an output may have the same name as an input.
    • description: Human-readable description of the input/output.
    • default: Specifies the default value for an input. Only valid for inputs.
    • type: Specifies the type of input/output. The types are used as hints for pipeline authors and can be used by the pipeline system/UI to validate arguments and connections between components. Basic types are String, Integer, Float, and Bool. See the full list of types defined by the Kubeflow Pipelines SDK.

Implementation

  • implementation: Specifies how to execute the component instance. There are two implementation types, container and graph. (The latter is not in scope for this document.) In future we may introduce more implementation types like daemon or K8sResource.

    • container: Describes the Docker container that implements the component. A portable subset of the Kubernetes Container v1 spec.

      • image: Name of the Docker image.
      • command: Entrypoint array. The Docker image’s ENTRYPOINT is used if this is not provided. Each item is either a string or a placeholder. The most common placeholders are {inputValue: Input name} and {outputPath: Output name}.
      • args: Arguments to the entrypoint. The Docker image’s CMD is used if this is not provided. Each item is either a string or a placeholder. The most common placeholders are {inputValue: Input name} and {outputPath: Output name}.
      • env: Map of environment variables to set in the container.
      • fileOutputs: Legacy property that is only needed in cases where the container always stores the output data in some hard-coded non-configurable local location. This property specifies a map between some outputs and local file paths where the program writes the output data files. Only needed for components that have hard-coded output paths. Such containers need to be fixed by modifying the program or adding a wrapper script that copies the output to a configurable location. Otherwise the component may be incompatible with future storage systems.

You can set all other Kubernetes container properties when you use the component inside a pipeline.

Using placeholders for command-line arguments

Consuming input by value (parameter)

The placeholder is replaced by the value of the input argument:

  • In component.yaml:

    command: [program.py, --rounds, {inputValue: Rounds}]
    
  • In the pipeline code:

    task1 = component1(rounds=150)
    
  • Resulting command-line code (showing the value of the input argument that has replaced the placeholder):

    program.py --rounds 150
    

Outputs

Output paths are filled in by the pipeline system. The outputPath placeholder is replaced by a path. (The path can point to a mounted output volume, for example.) The parent directories of the path may or may not not exist. Your program must handle both cases without error.

  • In component.yaml:

    command: [program.py, --out-model, {outputPath: trained_model}]
    
  • In the pipeline code:

    task1 = component1()
    
  • Resulting command-line code (the placeholder is replaced by the generated path):

    program.py --out-model /outputs/trained_model/data