Prediction intake options¶

You can configure a prediction source using the Predictions > Job Definitions tab or the Batch Prediction API. This topic describes both the UI and API intake options.

If you are using a custom CSV format, any intake option dealing with CSV will adhere to that format.

Local file streaming¶

Local file intake does not have any special options. This intake option requires you to upload the job's scoring data using a PUT request to the URL specified in the csvUpload link in the job data. This starts the job (or queues it for processing if the prediction instance is already occupied).

If there is no other queued job for the selected prediction instance, scoring will start while you are still uploading.

Refer to this sample use case.

Multipart upload¶

Because the local file intake process requires that you upload scoring data for a job using a PUT request to the URL specified in the csvUpload parameter, by default, a single PUT request starts the job (or queues it for processing if the prediction instance is occupied). Multipart upload for batch predictions allows you to override the default behavior to upload scoring data through multiple files. This upload process requires multiple PUT requests followed by a single POST request (finalizeMultipart) to finalize the multipart upload manually. This feature can be helpful when you want to upload large datasets over a slow connection or if you experience frequent network instability.

Multipart upload endpoints¶

This feature adds the following multipart upload endpoints to the batch prediction API:

Local file intake settings¶

The intake settings for the local file adapter added two new properties to support multipart upload for the batch prediction API:

Multipart intake setting¶

To enable the new multipart upload workflow, configure the intakeSettings for the localFile adapter as shown in the following sample request:

{
    "intakeSettings": {
        "type": "localFile",
        "multipart": true
    }
}

• Upload any number of sequentially numbered files.

• Finalize the upload to indicate that all required files uploaded successfully.

Async intake setting¶

To enable the new multipart upload workflow with async enabled, configure the intakeSettings for the localFile adapter as shown in the following sample request:

{
    "intakeSettings": {
        "type": "localFile",
        "multipart": true,
        "async": false
    }
}

A defining feature of batch predictions is that the scoring job starts on the initial file upload, and only one batch prediction job at a time can run for any given prediction instance. This functionality may cause issues when uploading large datasets over a slow connection. In these cases, the client's upload speed could create a bottleneck and block the processing of other jobs. To avoid this potential bottleneck, you can set async to false, as shown in the example above. This configuration postpones submitting the batch prediction job to the queue.

When "async": false, the point at which a job enters the batch prediction queue depends on the multipart setting:

• If "multipart": true, the job is submitted to the queue after the POST request for finalizeMultipart resolves.

• If "multipart": false, the job is submitted to the queue after the initial file intake PUT request resolves.

Example multipart upload requests¶

The batch prediction API requests required to upload a 3 part multipart batch prediction job would be:

PUT /api/v2/batchPredictions/:id/csvUpload/part/0/

PUT /api/v2/batchPredictions/:id/csvUpload/part/1/

PUT /api/v2/batchPredictions/:id/csvUpload/part/2/

POST /api/v2/batchPredictions/:id/csvUpload/finalizeMultipart/

Each uploaded part is a complete CSV file with a header.

Abort a multipart upload¶

If you start a multipart upload that you don't want to finalize, you can use a DELETE request to the existing batchPredictions abort route:

DELETE /api/v2/batchPredictions/:id/

AI Catalog dataset scoring¶

To read input data from an AI Catalog / Data Registry dataset, the following options are available:

Wrangler recipe dataset scoring¶

The following options are available to read input data from a wrangler recipe created in the DataRobot NextGen Workbench from a Snowflake data connection:

Amazon S3 scoring¶

For larger files, S3 is the preferred method for intake. DataRobot can ingest files from both public and private buckets. To score from Amazon S3, you must set up a credential with DataRobot consisting of an access key (ID and key) and, optionally, a session token.

AWS credentials are encrypted and only decrypted when used to set up the client for communication with AWS during scoring.

Azure Blob Storage scoring¶

Another scoring option for large files is Azure. To score from Azure Blob Storage, you must configure credentials with DataRobot using an Azure Connection String.

Azure credentials are encrypted and are only decrypted when used to set up the client for communication with Azure during scoring.

Google Cloud Storage scoring¶

DataRobot also supports the Google Cloud Storage adapter. To score from Google Cloud Storage, you must set up a credential with DataRobot consisting of a JSON-formatted account key.

GCP credentials are encrypted and are only decrypted when used to set up the client for communication with GCP during scoring.

HTTP scoring¶

In addition to the cloud storage adapters, you can also point batch predictions to a regular URL so DataRobot can stream the data for scoring:

The URL can optionally contain a username and password, such as https://username:password@example.com/datasets/scoring.csv.

The http adapter can be used for ingesting data from pre-signed URLs from either S3, Azure, or GCP.

JDBC scoring¶

DataRobot supports reading from any JDBC-compatible database for Batch Predictions.

To use JDBC with the Batch Prediction API, specify jdbc as the intake type. Since no file is needed for a PUT request, scoring will start immediately, transitioning the job to RUNNING if preliminary validation succeeds. To support this, the Batch Prediction API integrates with external data sources using credentials securely stored in data credentials.

Supply data source details using the Predictions > Job Definitions tab or the Batch Prediction API (intakeSettings) as described in the table below.

Refer to the example section for a complete API example.

Allowed source IP addresses¶

Any connection initiated from DataRobot originates from one of the following IP addresses:

BigQuery scoring¶

To use BigQuery for scoring, supply data source details using the Predictions > Job Definitions tab or the Batch Prediction API (intakeSettings) as described in the table below.

Refer to the example section for a complete API example.

Snowflake scoring¶

Using JDBC to transfer data can be costly in terms of IOPS (input/output operations per second) and expense for data warehouses. This adapter reduces the load on database engines during prediction scoring by using cloud storage and bulk insert to create a hybrid JDBC-cloud storage solution.

Supply data source details using the Predictions > Job Definitions tab or the Batch Prediction API (intakeSettings) as described in the table below.

Refer to the example section for a complete API example.

Synapse scoring¶

To use Synapse for scoring, supply data source details using the Predictions > Job Definitions tab or the Batch Prediction API (intakeSettings) as described in the table below.

Refer to the example section for a complete API example.