Extending Custom Models

You can extend custom models by implementing the ModelController interface.

ModelController Interface

ModelController is the core interface for all model controllers, defining model lifecycle management methods.

Interface Definition

public interface ModelController {
    String getModelName();
    List<FieldSchema> getOutputFields(ModelConfig model);
    String checkModel(ModelConfig model);
    String genModelTrainK8sYaml(ModelConfig model, ModelTrainConf trainConf);
    List<String> getExportCheckpoints(ModelExportConf exportConf);
    String genModelExportK8sYaml(ModelConfig model, ModelExportConf exportConf);
    String getServiceUrl(ModelConfig model, ServiceConfig serviceConf);
    String getServiceK8sYaml(ModelConfig model, ServiceConfig serviceConf);
}

Method Description

Method	Description
getModelName()	Return model type name, used to identify model type
getOutputFields(ModelConfig)	Return model output field list
checkModel(ModelConfig)	Check if model configuration is valid, return null if valid
genModelTrainK8sYaml(ModelConfig, ModelTrainConf)	Generate training task Kubernetes YAML
getExportCheckpoints(ModelExportConf)	Get exported checkpoint name list
genModelExportK8sYaml(ModelConfig, ModelExportConf)	Generate export task Kubernetes YAML
getServiceUrl(ModelConfig, ServiceConfig)	Get service access URL
getServiceK8sYaml(ModelConfig, ServiceConfig)	Generate service deployment Kubernetes YAML

Implementation Steps

1. Create Model Configuration Class (optional)

public class MyModelConfig {
    public static final ConfigOption<String> CUSTOM_PARAM = new ConfigOption<>(
        "custom_param",
        "default_value", 
        "Custom parameter description", 
        null, 
        String.class
    );
}

2. Implement ModelController Interface

public class MyModel implements ModelController {
    
    @Override
    public String getModelName() {
        return "my_model";
    }
    
    @Override
    public List<FieldSchema> getOutputFields(ModelConfig model) {
        return Arrays.asList(
            new FieldSchema("prediction", "DOUBLE"),
            new FieldSchema("confidence", "DOUBLE")
        );
    }
    
    @Override
    public String checkModel(ModelConfig model) {
        return null;
    }
    
    @Override
    public String genModelTrainK8sYaml(ModelConfig model, ModelTrainConf trainConf) {
        return "...";
    }
    
    @Override
    public List<String> getExportCheckpoints(ModelExportConf exportConf) {
        return Arrays.asList(exportConf.getCheckpointName() + "_export");
    }
    
    @Override
    public String genModelExportK8sYaml(ModelConfig model, ModelExportConf exportConf) {
        return "...";
    }
    
    @Override
    public String getServiceUrl(ModelConfig model, ServiceConfig serviceConf) {
        return "http://" + serviceConf.getId() + ".svc.cluster.local:80/predict";
    }
    
    @Override
    public String getServiceK8sYaml(ModelConfig model, ServiceConfig serviceConf) {
        return "...";
    }
}

3. Register Model Controller (SPI method)

Create file com.sqlrec.common.model.ModelController in src/main/resources/META-INF/services/ directory, file content is the fully qualified name of the implementation class:

com.example.MyModel

Using Custom Models

CREATE MODEL my_custom_model (
    feature1 VARCHAR,
    feature2 DOUBLE
) WITH (
    model = 'my_model',
    custom_param = 'value'
);