Machine learning production quest

This type of production is used for time-series forecasting use cases that work with larger datasets and the results are not expected real-time.

Select the Batch Production tab to access the batch production quest canvas. This tab is active when you have exported the selected activities from the training quest that you intend to use in the batch scoring pipeline.

After the batch production quest is created, you can continue to modify the pipeline if necessary. When the batch quest is initially created, the entry point activity is Ingest Batch Data. This activity can consume only datasets from Data Lake. It is assumed that the input dataset has the same schema as the dataset in the corresponding Import Data activity that is marked in the training quest. The same schema assumes the same variables of the same types.

Saving and running the batch production quest produces the predictions as the output of the Score Model activity.

The execution of the Batch Production can be scheduled for regular runs through an ION Workflow.

Realtime production is used to specify the final flow of activities and deploy the model as a REST API for real-time inferences.

Select the Realtime Production tab to access the real-time production quest canvas. This tab is active when you have exported the selected activities from the training quest that you intend to use in the final endpoint deployment process.

After the real-time production quest is created, you must define the schema. When the production quest is initially created, the entry point activity is Ingest Data. This activity assumes that the input dataset has the same schema as the dataset in the corresponding Import Data activity that is marked in the training quest. The same schema assumes the same variables of the same types.

You can use the Endpoint schema configurator to reduce the original set of variables and remove all redundant variables. When calling an endpoint, you must send the values of variables as defined in the endpoint schema, in the same order.

Saving and running the real-time production quest on a sample of data is necessary to test if predictions of the trained model are as expected.

The Ingest Data and the Prediction activity boxes cannot be deleted.

The Ingest Data activity represents a random sample of the initial dataset used in the training quest. This is to test the real-time production quest.

The output port of the Prediction activity provides the predictions from the trained model when unseen data is ingested to test the model.