Kale Versions Save

New features:

• Introduce an SDK to programmatically create pipelines from Python modules (#326)
• Handle pipeline versions (#330)
• Create KFP artifacts using the SDK (#333)
• Support both v1alpha3 and v1beta1 Katib versions (#336)

Notable fixes/improvements:

• Detect docker image SHA instead of tag (#328)
• Bump MLMD version (#335)
• Various utils improvements (#337)
• Improve InferenceService handling when fields are missing (#343)

Fixes:

• Installation with pip >= 20.3 no longer breaks (#277)
• Dockerfiles now work as expected (#280)

New features:

• Serve KFServing models with a single API #256
• New OpenVaccine example #235
• New XGBoost marshalling backend #252
• Specify storage class and access modes for PVCs #255

Notable fixes/improvements:

• Refactoring of the marshal package #254
• Fixed state management in labextension #249
• Fix potential name conflicts in templates #203
• Various improvements to examples #205
• Fix labextension sidebar width #257
• Marshalling improvements #218 #228 #236 #238 #243 #268

This minor release mainly focuses on bug fixing and overall UX improvements.

Notably:

• Better log reporting
• Improved Katib jobs monitoring
• Relaxed KFP requirements to support newer versions
• New DogBreed example supporting TensorFlow 2.x

New features:

• Run hyperparameter tuning experiments from JupyterLab
• Jupyter rich outputs are displayed in the KFP dashboard
• Pipeline parameters and pipeline metrics
• Run pipeline steps on GPU
• Revamped cell tag UI and app theme

Improvements and bug fixes:

• Better data passing detection
• Faster UI renders
• Better notebook validation - fail early in case of errors
• Many more bug fixes and UX improvements

This release includes the code presented at KubeCon NA 2019 during the Kale workshop, with some additional bug fixes and performance improvements

This release includes the necessary machinery to convert a Jupyter Notebook to a Kubeflow Pipelines deployment.

This release provides four main modules:

• nbparser: notebook parse module; tagging-language; generation of code graph
• static_analysis: run static analysis over code blocks to detect data dependencies
• marshal: functions to (de)serialize objects of any type with dynamic dispatchers
• codegen: generate kfp Python code using templates, based on the graph produced by nbparser module

Flask Server

The api module provides a simple Flask app that exposes the /kale API that accepts a JupyterNotebook in raw format and call the Kale core module to create a KFP deployment.

JupyterLab extension

The kale-toolbar-runner provides a deployment button in the JupyterNotebook's toolbar. By clicking the deployment button Jupyter will send a POST request to localhost:5000/kale with the currently active raw notebook.