E-MapReduce:Integrate EMR Serverless Spark with an external model service

Procedure

The following steps walk you through deploying the Qwen3-0.6B model in PAI-EAS, registering it in EMR Serverless Spark, and calling it for AI inference.

If you have already deployed a service in the PAI console, you can skip to Obtain credentials.

Deploy a service

The following example deploys the Qwen3-0.6B public model.

1. Log on to the PAI console. Select a region on the top of the page. Then, select the desired workspace and click Elastic Algorithm Service (EAS).

2. On the Inference Services tab, click Components. In the Scenario-based Model Deployment section, click Deploy LLM.

3. On the Deploy LLM page, configure the following key parameters.

   • Model configuration: Select Public Model. Search for and select Qwen3-0.6B from the list.

   • Inference engine: SGLang/vLLM is recommended for its high compatibility with the OpenAI API standard. This example uses vLLM. For more information, see Choose a suitable inference engine.

   • Deployment template: Select Standalone. The template automatically provides recommended parameters, such as instance specification and image.

4. Click Deploy. The deployment process takes about 5 minutes. When the service status changes to Running, the deployment is successful.

   Note

   If the deployment fails, see Troubleshoot service deployment and status issues.

Obtain credentials

After the service is deployed, obtain its VPC endpoint and token. You need these credentials to register the service in EMR Serverless Spark.

1. On the Inference Services tab, click the name of your service to go to the Overview page. In the Basic Information section, click View Invocation Information.

2. In the Invocation Information dialog box, select the shared gateway tab to view and copy the public endpoint, VPC endpoint, and token.

Register the service

Register the PAI-EAS service in EMR Serverless Spark to allow the ai_query() function in Spark SQL to call it.

1. Navigate to the model service page.

   1. Log on to the E-MapReduce console.

   2. In the left navigation pane, choose EMR Serverless > Spark.

   3. On the Spark page, click the name of your target workspace.

   4. On the EMR Serverless Spark page, click AI center > > model service in the left navigation pane.

2. On the model service tab, click Create External Model Service and configure the following parameters:

   Parameter	Example Value	Description
   Model service name	my_qwen_service	Used as the endpointName parameter in AI function calls. Must be unique within the workspace and cannot be modified after creation.
   Endpoint	http://12*******39.vpc.cn-hangzhou.pai-eas.aliyuncs.com/api/predict/<ServiceName>/v1	Enter the VPC endpoint from the previous step and manually append /v1 to the end.
   Model name	Qwen3-0.6B	The model name used in service calls.
   Model type	Chat	Select Chat or Embedding based on the deployed model type.
   API key	nMzI**********************Zg==	Enter the Token that you obtained in the previous step.
   Description	Latest Qwen multimodal model service	Enter a brief description for easy identification.

3. After configuring the parameters, click create to register the model service.

Call the model

After the model service is registered, use the built-in ai_query() function in EMR Serverless Spark to call the model with standard SQL statements. The following example demonstrates data masking.

1. Create a Spark SQL job and enable the AI function

   1. On the Development tab, click the (New) icon.

   2. In the dialog box that appears, enter a Name, select SparkSQL as the type, and then click OK.

   3. In the drop-down list in the upper-right corner, click Connect to SQL Session and configure the following information:

      Parameter	Description
      Engine version	Select one of the following versions. esr-4.x: esr-4.6.0 or later. esr-3.x: esr-3.5.0 or later. esr-2.x: esr-2.9.0 or later.
      Advanced configurations	Add the spark.emr.serverless.ai.function.enable true configuration to enable the AI feature.

2. Write SQL to call the model

   On the data development page, call the model by using the ai_query() function in SQL.

   -- The second parameter, 'my_qwen_service', is the Model Service Name you set during registration.
   select ai_query('Perform data masking on the following text according to these rules:
   1) Replace all Chinese names with ""
   2) Keep the first 5 digits of phone numbers and replace the rest with "*"
   3) Replace complete addresses with "*****"
   4) Keep all other text unchanged
   5) Output only the masked text without any explanation
   Original text: My name is Zhang San, my phone number is 12345678900, please navigate to Smart Home, Longgang District, Shenzhen City', 'my_qwen_service');

3. View the masked result

   The following output shows the masked result.

   My name is , my phone number is 12345*****, please navigate to *****

Related documentation

• For more information about deploying custom models on PAI-EAS, see Quick start for Elastic Algorithm Service (EAS).

• PAI-EAS provides a one-stop solution for LLM deployment. For details, see Deploy large language models (LLMs).