GCP AI Platform support

Google Cloud’s AI Platform offers couple services that simplify Machine Learning tasks with use of Kubeflow based components.

Using kedro with AI Platform Notebooks

AI Platform Notebooks provides an easy way to manage and host JupyterLab based data science workbench environment. What we’ve found out is that the default images provided by a service cause some dependency conflicts. To avoid this issues make sure you use isolated virtual environment, e.g. virtualenv. New virual environment can be created by simply invoking python -m virtualenv venv command.

Using kedro-kubeflow with AI Platform Pipelines

AI Platform Pipelines is a service that allows to easily deploy Kubeflow Pipelines on new or existing Google Kubernetes Engine clusters.

In general kedro-kubeflow plugin should work with AI Platform Pipelines out of the box, with the only exception is that it requires authentication. Note that the host variable should point to a dashbard URL generated by AI Platform Pipelines service (e.g. https://653hddae86eb7b0-dot-europe-west1.pipelines.googleusercontent.com/), just open the dashboard from the service page and copy url from the browser.

Below is the list of authentication scenarios supported so far:

1. Connecting to AI Pipelines from AI Platform Notebooks

In this scenario authentication works out of the box with default credentials mechanism.

2. Authentication to AI Pipelines from local environment

To interact with AI Platform Pipelines from local environment you can use the mechanisms provided by Google Cloud SDK. After installing the SDK run google cloud application-default login to initialize default credentials on your local machine.

You can use service account key for authentication as well. To make that work just set GOOGLE_APPLICATION_CREDENTIALS environment variable to the path of where the service account key file is stored.

3. Authenticating through IAP Proxy

Identity Aware Proxy is a product that allows securing your cloud based applications with Google Identity.

To authenticate with IAP find out which oauth client ID is the proxy configured with and then save it in IAP_CLIENT_ID environment variable. The authentication should work seamlessly assuming identity you are using has been granted access to the application.

The above will work if you are connecting from within GCP VM or locally with specified service account credentials. It will NOT work for credentials obtained with google cloud application-default login.

Using kedro-kubeflow with Vertex AI Pipelines (DEPRECATED)

Vertex AI Pipelines support in kedro-kubeflow has been deprecated in favour of the new plugin kedro-vertexai