Invoke Vertex AI online predictions from AlloyDB databases

Preview

This product is covered by the Pre-GA Offerings Terms of the Google Cloud Terms of Service. Pre-GA products might have limited support, and changes to pre-GA products might not be compatible with other pre-GA versions. For more information, see the launch stage descriptions.

This page shows how to invoke Vertex AI online predictions from an AlloyDB database, including the steps required to enable this capability.

About invoking online predictions

AlloyDB provides you the ability to get Vertex AI online predictions in your SQL code by calling the ml_predict_row function so that you don't have to manually create a custom integration with Vertex AI.

Note: The ml_predict_row function supports invoking predictions only on tabular or custom models.

Vertex AI online prediction is a service that is optimized to run data through hosted models with low latency. You send small batches of data to the service and it returns your predictions in the response.

Limitations that apply to AlloyDB Vertex AI online predictions are available in Vertex AI quotas and limits.

For more information about Vertex AI, see Introduction to Vertex AI.

How to enable and invoke online predictions

After database administrators enable database access to Vertex AI online predictions and grant database users the access to predictions, the database users can then invoke predictions using the ml_predict_row SQL function.

1. Enable database access to predictions.

   To enable AlloyDB database access to Vertex AI online predictions, you grant the AlloyDB service agent IAM permissions to access Vertex AI, and then you create the google_ml_integration extension in the database containing the data you want to run predictions on.

2. Grant database users access to predictions.

   In the database where you've created the extension, grant the database users who are going to run predictions the right to execute the ml_predict_row function.

3. Invoke predictions using SQL.

   Authorized database users invoke the ml_predict_row SQL function to run Vertex AI online predictions against database data.

Enable database access to predictions

To enable AlloyDB database access to Vertex AI online predictions, you grant the AlloyDB service agent IAM permissions to access Vertex AI, and then you create the google_ml_integration extension in the database containing the data you want to run predictions on.

1. Add Vertex AI permissions to the AlloyDB service agent for the project where the AlloyDB database's cluster is located:

   Console

   1. Get the project number of the project that has AlloyDB clusters or instances by following the instructions available in Identifying projects.
   2. In the Google Cloud console, go to the IAM page.

      Go to IAM

   3. Select the project that has Vertex AI endpoints.
   4. Enable the Include Google-provided role grants checkbox.
   5. Click Add.
   6. In the New principals field, enter:

      [e-mail]

      where, PROJECT_NUMBER is the project number you got earlier.

   7. In the Role field, enter Vertex AI User.
   8. Click Save.

   gcloud

   gcloud projects add-iam-policy-binding PROJECT_ID \
     --member="serviceAccount:[e-mail]" \
     --role="roles/aiplatform.user"
     

   • PROJECT_ID: The ID of the project that has the Vertex AI endpoint.
   • PROJECT_NUMBER: The project number of the project that has AlloyDB clusters or instances.
   Note: In general, the policy change takes effect within 60 seconds. However, under certain circumstances, it can take up to 7 minutes for such changes to fully propagate across the system.
2. Create the google_ml_integration extension in the database containing the data you want to run predictions on:
   1. Connect a psql client to the cluster's primary instance, as described in Connect a psql client to an instance.
      Note: You must connect to the primary instance even if you intend to invoke predictions while connected to a read pool instance.
   2. At the psql command prompt, connect to the database and create the extension:

      \c DB_NAME
      CREATE EXTENSION IF NOT EXISTS google_ml_integration;
             

      DB_NAME: The name of the database on which the extension should be created.

Note: The availability of the google_ml_integration extension is controlled by the google_ml_integration.enabled database flag, which is set to on by default when you create an AlloyDB instance.

Grant database users access to invoke predictions

Grant permission for database users to execute the ml_predict_row function to run predictions:

1. Connect a psql client to the cluster's primary instance, as described in Connect a psql client to an instance.

   Note: You must connect to the primary instance even if you intend to invoke predictions while connected to a read pool instance.

2. At the psql command prompt, connect to the database and grant permissions:

   \c DB_NAME
   GRANT EXECUTE ON FUNCTION ml_predict_row TO USER_NAME;
   

   • DB_NAME: The name of the database on which the permissions should be granted.
   • USER_NAME: The name of the user for whom the permissions should be granted.

Invoke predictions using SQL

Invoke the ml_predict_row SQL function to invoke Vertex AI online predictions against database data. The function has the following syntax:

ML_PREDICT_ROW (MODEL_ENDPOINT TEXT,
               ARGS JSON) RETURNS JSON
       

• MODEL_ENDPOINT: The Vertex AI online prediction endpoint to send the request to, specified in this format:

  projects/PROJECT_ID/locations/REGION_ID/endpoints/ENDPOINT_NUMBER
   

  For more information about this endpoint format, see Path parameters.
• ARGS: The content to run the prediction on, with additional prediction specific parameters, specified as a JSON request in this format:

  
  {
     "instances": [
       value
     ],
     "parameters": value
  }
  

  For more information about this request format, see Online predictions with the API.
• RETURN VALUE: The predictions as a JSON response in this format:

  
  {
  
    "predictions": [
       value
    ],
    "model": string,
    "modelDisplayName": string
  }
  
  

  For more information about this format, see Response body.