Skip to content

On-premise users: click in-app to access the full platform documentation for your version of DataRobot.

Predictions > Portable prediction methods > Portable Prediction Server > Portable batch predictions

Portable batch predictions

Availability information

The Portable Prediction Server is a premium feature exclusive to DatRobot MLOps. Contact your DataRobot representative or administrator for information on enabling this feature.

Portable batch predictions (PBP) let you score large amounts of data on disconnected environments. Before you can use portable batch predictions, you need to configure the Portable Prediction Server (PPS), a DataRobot execution environment for DataRobot model packages (.mlpkg files) distributed as a self-contained Docker image. Portable batch predictions use the same Docker image as the PPS but run it in a different mode.

Scoring methods

Portable batch predictions can use the following adapters to score datasets:

  • Filesystem
  • JDBC
  • AWS S3
  • Azure Blob
  • GCS
  • Snowflake
  • Synapse

To run portable batch predictions, you need the following artifacts:

After you prepare these artifacts, you can run portable batch predictions. See also additional examples of running portable batch predictions.

Job definitions

You can define jobs using a JSON config file in which you describe prediction_endpoint, intake_settings, output_settings, timeseries_settings (optional) for time series scoring, and jdbc_settings (optional) for JDBC scoring.

Self-Managed AI Platform only: Prediction endpoint SSL configuration

If you need to disable SSL verification for the prediction_endpiont, you can set ALLOW_SELF_SIGNED_CERTS to True. This configuration disables SSL certificate verification for requests made by the application to the web server. This is useful if you have SSL encryption enabled on your cluster and are using certificates that are not signed by a globally trusted Certificate Authority (self-signed).

The prediction_endpoint describes how to access the PPS and is constructed as <schema>://<hostname>:<port>, where you define the following parameters:

Parameter Type Description
schema string http or https
hostname string The hostname of the instance where your PPS is running.
port string The port of the prediction API running inside the PPS.

The jdbc_setting has the following attributes:

Parameter Type Description
url string The URL to connect via the JDBC interface.
class_name string The class name used as an entry point for JDBC communication.
driver_path string The path to the JDBC driver on your filesystem (available inside the PBP container).
template_name string The name of the template in case of write-back. To obtain the names of the support templates, contact your DataRobot representative.

The other parameters are similar to those available for standard batch predictions, however, they are in snake_case, not camelCase:

Parameter Type Description
abort_on_error boolean Enable or disable cancelling the portable batch prediction job if an error occurs.
Example: true
chunk_size string Chunk the dataset for scoring in sequence as asynchronous tasks. In most cases, the default value will produce the best performance. Bigger chunks can be used to score very fast models and smaller chunks can be used to score very slow models.
Example: "auto"
column_names_remapping array Rename or remove columns from the output for this job. Set an output_name for the column to null or false to remove it.
Example: [{'input_name': 'isbadbuy_1_PREDICTION', 'output_name':'prediction'}, {'input_name': 'isbadbuy_0_PREDICTION', 'output_name': null}]
csv_settings object Set the delimiter, character encoding, and quote character for comma separated value (CSV) files.
Example: { "delimiter": ",", "encoding": "utf-8", "quotechar": "\"" }
deployment_id string Define the ID of the deployment associated with the portable batch predictions.
Example: 61f05aaf5f6525f43ed79751
disable_row_level_error_handling boolean Enable or disable error handling by prediction row.
Example: false
include_prediction_status boolean Enable or disable including the prediction_status column in the output; defaults to false.
Example: false
include_probabilities boolean Enable or disable returning probabilities for all classes.
Example: true
include_probabilities_classes array Define the classes to provide class probabilities for.
Example: [ 'setosa', 'versicolor', 'virginica' ]
intake_settings object Set the intake options required for the input type.
Example: { "type": "localFile" }
num_concurrent integer Set the maximum number chunks to score concurrently on the prediction instance specified by the deployment.
Example: 1
output_settings object Set the output options required for the output type.
Example: { "credential_id": "string", "format": "csv", "partitionColumns": [ "string" ], "type": "azure", "url": "string" }
passthrough_columns array Define the scoring dataset columns to include in the prediction response. This option is mutually exclusive with passthrough_columns_set.
Example: [ "column1", "column2" ]
passthrough_columns_set string Enable including all scoring dataset columns in the prediction response. The only option is all. This option is mutually exclusive with passthrough_columns.
Example: "all"
prediction_warning_enabled boolean Enable or disable prediction warnings.
Example: true
skip_drift_tracking boolean Enable or disable drift tracking for this batch of predictions. This allows you to make test predictions without affecting deployment stats.
Example: false
timeseries_settings object Define the settings required for time series predictions.
Example: { "forecast_point": "2019-08-24T14:15:22Z", "relax_known_in_advance_features_check": false, "type": "forecast" }

You can also configure Prediction Explanations for portable batch predictions:

Parameter Type Description
max_explanations int/str Set the number of explanations returned by the prediction server. For SHAP explanations, a special constant all is also accepted.
Example: 1
explanation_algorithm string Define the algorithm used for Prediction Explanations, either SHAP or XEMP.
Example: "shap"
explanation_class_names array Define the class names to explain for each row. This setting is only applicable to XEMP Prediction Explanations for multiclass models and it is mutually exclusive with explanation_num_top_classes.
Example: [ "class1", "class2" ]
explanation_num_top_classes integer Set the number of top predicted classes, by prediction value, to explain for each row. This setting is only applicable to XEMP Prediction Explanations for multiclass models and it is mutually exclusive with explanation_class_names.
Example: 1
threshold_low float Set the lower threshold for requiring a Prediction Explanation. Predictions must be below this value (or above the threshold_high value) for Prediction Explanations to compute.
Example: 0.678
threshold_high float Set the upper threshold for requiring a Prediction Explanation. Predictions must be above this value (or below the threshold_low value) for Prediction Explanations to compute.
Example: 0.345

The following outlines a JDBC example that scores to and from Snowflake using single-mode PPS running locally and can be defined as a job_definition_jdbc.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "intake_settings": {
        "type": "jdbc",
        "table": "SCORING_DATA",
        "schema": "PUBLIC"
    },
    "output_settings": {
        "type": "jdbc",
        "table": "SCORED_DATA",
        "statement_type": "create_table",
        "schema": "PUBLIC"
    },
    "passthrough_columns_set": "all",
    "include_probabilities": true,
    "jdbc_settings": {
        "url": "jdbc:snowflake://my_account.snowflakecomputing.com/?warehouse=WH&db=DB&schema=PUBLIC",
        "class_name": "net.snowflake.client.jdbc.SnowflakeDriver",
        "driver_path": "/tmp/portable_batch_predictions/jdbc/snowflake-jdbc-3.12.0.jar",
        "template_name": "Snowflake"
    }
}

Credentials environment variables

If you are using JDBC or private containers in cloud storage, you can specify the required credentials as environment variables. The following table shows which variables names are used:

Name Type Description
AWS_ACCESS_KEY_ID string AWS Access key ID
AWS_SECRET_ACCESS_KEY string AWS Secret access key
AWS_SESSION_TOKEN string AWS token
GOOGLE_STORAGE_KEYFILE_PATH string Path to GCP credentials file
AZURE_CONNECTION_STRING string Azure connection string
JDBC_USERNAME string Username for JDBC
JDBC_PASSWORD string Password for JDBC
SNOWFLAKE_USERNAME string Username for Snowflake
SNOWFLAKE_PASSWORD string Password for Snowflake
SYNAPSE_USERNAME string Username for Azure Synapse
SYNAPSE_PASSWORD string Password for Azure Synapse

Here's an example of the credentials.env file used for JDBC scoring:

export JDBC_USERNAME=TEST_USER
export JDBC_PASSWORD=SECRET

Run portable batch predictions

Portable batch predictions run inside a Docker container. You need to mount job definitions, files, and datasets (if you are going to score from a host filesystem and set a path inside the container) onto Docker. Using a JDBC job definition and credentials from previous examples, the following outlines a complete example of how to start a portable batch predictions job to score to and from Snowflake.

docker run --rm \
    -v /host/filesystem/path/job_definition_jdbc.json:/docker/container/filesystem/path/job_definition_jdbc.json \
    --network host \
    --env-file /host/filesystem/path/credentials.env \
    datarobot-portable-predictions-api batch /docker/container/filesystem/path/job_definition_jdbc.json

Here is another example of how to run a complete end-to-end flow, including PPS and a write-back job status into the DataRobot platform for monitoring progress.

#!/bin/bash

# This snippet starts both the PPS service and PBP job using the same PPS docker image
# available from Developer Tools.

#################
# Configuration #
#################

# Specify path to directory with mlpkg(s) which you can download from deployment
MLPKG_DIR='/host/filesystem/path/mlpkgs'
# Specify job definition path
JOB_DEFINITION_PATH='/host/filesystem/path/job_definition.json'
# Specify path to file with credentials if needed (for cloud storage adapters or JDBC)
CREDENTIALS_PATH='/host/filesystem/path/credentials.env'
# For DataRobot integration, specify API host and Token
API_HOST='https://app.datarobot.com'
API_TOKEN='XXXXXXXX'

# Run PPS service in the background
PPS_CONTAINER_ID=$(docker run --rm -d -p 127.0.0.1:8080:8080 -v $MLPKG_DIR:/opt/ml/model datarobot/datarobot-portable-prediction-api:<version>)
# Wait some time before PPS starts up
sleep 15
# Run PPS in batch mode to start PBP job
docker run --rm -v $JOB_DEFINITION_PATH:/tmp/job_definition.json \
    --network host \
    --env-file $CREDENTIALS_PATH \
    datarobot/datarobot-portable-prediction-api:<version> batch /tmp/job_definition.json
        --api_host $API_HOST --api_token $API_TOKEN
# Stop PPS service
docker stop $PPS_CONTAINER_ID

More examples

In all of the following examples, assume that PPS is running locally on port 8080, and the filesystem structure has the following format:

/host/filesystem/path/portable_batch_predictions/
├── job_definition.json
├── credentials.env
├── datasets
|   └── intake_dataset.csv
├── output
└── jdbc
    └── snowflake-jdbc-3.12.0.jar

Filesystem scoring with single-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "intake_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/datasets/intake_dataset.csv"
    },
    "output_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/output/results.csv"
    }
}

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json

Filesystem scoring with multi-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "deployment_id": "lending_club",
    "intake_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/datasets/intake_dataset.csv"
    },
    "output_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/output/results.csv"
    }
}

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json

Filesystem scoring with multi-model mode PPS and integration with DR job status tracking

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "deployment_id": "lending_club",
    "intake_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/datasets/intake_dataset.csv"
    },
    "output_settings": {
        "type": "filesystem",
        "path": "/tmp/portable_batch_predictions/output/results.csv"
    }
}

For the PPS MLPKG, in config.yaml, specify the deployment ID of the deployment for which you are running the portable batch prediction job.

#!/bin/bash

docker run --rm \
    --network host
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json \
        --api_host https://app.datarobot.com --api_token XXXXXXXXXXXXXXXXXXX

JDBC scoring with single-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "deployment_id": "lending_club",
    "intake_settings": {
        "type": "jdbc",
        "table": "INTAKE_TABLE"
    },
    "output_settings": {
        "type": "jdbc",
        "table": "OUTPUT_TABLE",
        "statement_type": "create_table"
    },
    "passthrough_columns_set": "all",
    "include_probabilities": true,
    "jdbc_settings": {
        "url": "jdbc:snowflake://your_account.snowflakecomputing.com/?warehouse=SOME_WH&db=MY_DB&schema=MY_SCHEMA",
        "class_name": "net.snowflake.client.jdbc.SnowflakeDriver",
        "driver_path": "/tmp/portable_batch_predictions/jdbc/snowflake-jdbc-3.12.0.jar",
        "template_name": "Snowflake"
    }
}

credentials.env file:

JDBC_USERNAME=TEST
JDBC_PASSWORD=SECRET

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    --env-file /host/filesystem/path/credentials.env \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json

S3 scoring with single-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "intake_settings": {
        "type": "s3",
        "url": "s3://intake/dataset.csv",
        "format": "csv"
    },
    "output_settings": {
        "type": "s3",
        "url": "s3://output/result.csv",
        "format": "csv"
    }
}

credentials.env file:

AWS_ACCESS_KEY_ID=XXXXXXXXXXXX
AWS_SECRET_ACCESS_KEY=XXXXXXXXXXX

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    --env-file /path/to/credentials.env \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json

Snowflake scoring with multi-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "deployment_id": "lending_club",
    "intake_settings": {
        "type": "snowflake",
        "table": "INTAKE_TABLE",
        "schema": "MY_SCHEMA",
        "external_stage": "MY_S3_STAGE_IN_SNOWFLAKE"
    },
    "output_settings": {
        "type": "snowflake",
        "table": "OUTPUT_TABLE",
        "schema": "MY_SCHEMA",
        "external_stage": "MY_S3_STAGE_IN_SNOWFLAKE",
        "statement_type": "insert"
    },
    "passthrough_columns_set": "all",
    "include_probabilities": true,
    "jdbc_settings": {
        "url": "jdbc:snowflake://your_account.snowflakecomputing.com/?warehouse=SOME_WH&db=MY_DB&schema=MY_SCHEMA",
        "class_name": "net.snowflake.client.jdbc.SnowflakeDriver",
        "driver_path": "/tmp/portable_batch_predictions/jdbc/snowflake-jdbc-3.12.0.jar",
        "template_name": "Snowflake"
    }
}

credentials.env file:

# Snowflake creds for JDBC connectivity
SNOWFLAKE_USERNAME=TEST
SNOWFLAKE_PASSWORD=SECRET
# AWS creds needed to access external stage
AWS_ACCESS_KEY_ID=XXXXXXXXXXXX
AWS_SECRET_ACCESS_KEY=XXXXXXXXXXX

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions \
    --env-file /host/filesystem/path/credentials.env \
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json

Time series scoring over Azure Blob with multi-model mode PPS

job_definition.json file:

{
    "prediction_endpoint": "http://127.0.0.1:8080",
    "deployment_id": "euro_date_ts_mlpkg",
    "intake_settings": {
        "type": "azure",
        "url": "https://batchpredictionsdev.blob.core.windows.net/datasets/euro_date.csv",
        "format": "csv"
    },
    "output_settings": {
        "type": "azure",
        "url": "https://batchpredictionsdev.blob.core.windows.net/results/output_ts.csv",
        "format": "csv"
    },
    "timeseries_settings":{
        "type": "forecast",
        "forecast_point": "2007-11-14",
        "relax_known_in_advance_features_check": true
    }
}

credentials.env file:

# Azure Blob connection string
AZURE_CONNECTION_STRING='DefaultEndpointsProtocol=https;AccountName=myaccount;AccountKey=XXX;EndpointSuffix=core.windows.net'

#!/bin/bash

docker run --rm \
    --network host \
    -v /host/filesystem/path/portable_batch_predictions:/tmp/portable_batch_predictions
    --env-file /host/filesystem/path/credentials.env
    datarobot/datarobot-portable-prediction-api:<version> batch \
        /tmp/portable_batch_predictions/job_definition.json
Updated April 2, 2025