kfp-notebook 0.21.0

Jupyter Notebook operator for Kubeflow Pipelines

KFP-Notebook is an operator that enable running notebooks as part of a Kubeflow Pipeline.

Building kfp-notebook

make clean install

Usage

The example below can easily be added to a python script or jupyter notebook for testing purposes.

import os
import kfp
from kfp_notebook.pipeline import NotebookOp
from kubernetes.client.models import V1EnvVar

# KubeFlow Pipelines API Endpoint
kfp_url = 'http://dataplatform.ibm.com:32488/pipeline'

# S3 Object Storage
cos_endpoint = 'http://s3.us-south.cloud-object-storage.appdomain.cloud'
cos_bucket = 'test-bucket'
cos_username = 'test'
cos_password = 'test123'
cos_directory = 'test-directory'
cos_dependencies_archive = 'test-archive.tar.gz'

# Inputs and Outputs
inputs = []
outputs = []

# Container Image
image = 'tensorflow/tensorflow:latest'

def run_notebook_op(op_name, notebook_path):

    notebook_op = NotebookOp(name=op_name,
                             notebook=notebook_path,
                             cos_endpoint=cos_endpoint,
                             cos_bucket=cos_bucket,
                             cos_directory=cos_directory,
                             cos_dependencies_archive=cos_dependencies_archive,
                             pipeline_outputs=outputs,
                             pipeline_inputs=inputs,
                             image=image)

    notebook_op.container.add_env_variable(V1EnvVar(name='AWS_ACCESS_KEY_ID', value=cos_username))
    notebook_op.container.add_env_variable(V1EnvVar(name='AWS_SECRET_ACCESS_KEY', value=cos_password))
    notebook_op.container.set_image_pull_policy('Always')

    return op

def demo_pipeline():
    stats_op = run_notebook_op('stats', 'generate-community-overview')
    contributions_op = run_notebook_op('contributions', 'generate-community-contributions')
    run_notebook_op('overview', 'overview').after(stats_op, contributions_op)

# Compile the new pipeline
kfp.compiler.Compiler().compile(demo_pipeline,'pipelines/pipeline.tar.gz')

# Upload the compiled pipeline
client = kfp.Client(host=kfp_url)
pipeline_info = client.upload_pipeline('pipelines/pipeline.tar.gz',pipeline_name='pipeline-demo')

# Create a new experiment
experiment = client.create_experiment(name='demo-experiment')

# Create a new run associated with experiment and our uploaded pipeline
run = client.run_pipeline(experiment.id, 'demo-run', pipeline_id=pipeline_info.id)

Generated Kubeflow Pipelines