Ce contenu n'est pas disponible dans la langue sélectionnée.

Chapter 156. Apache Spark

Apache Spark component

This documentation page covers the Apache Spark component for Apache Camel. The main purpose of the Spark integration with Camel is to provide a bridge between Camel connectors and Spark tasks. In particular Camel connector provides a way to route message from various transports, dynamically choose a task to execute, use incoming message as input data for that task and finally deliver the results of the execution back to the Camel pipeline.

Note

Apache Spark is not supported on Apache Karaf.

Supported architectural styles

Spark component can be used as a driver application deployed into an application server (or executed as a fat jar).

Spark component can also be submitted as a job directly into the Spark cluster.

While Spark component is primary designed to work as a long running job serving as an bridge between Spark cluster and the other endpoints, you can also use it as a fire-once short job.

Running Spark in OSGi servers

Currently the Spark component doesn't support execution in the OSGi container. Spark has been designed to be executed as a fat jar, usually submitted as a job to a cluster. For those reasons running Spark in an OSGi server is at least challenging and is not support by Camel as well.

URI format

Currently the Spark component supports only producers - it it intended to invoke a Spark job and return results. You can call RDD, data frame or Hive SQL job.

spark:{rdd|dataframe|hive}

RDD jobs

To invoke an RDD job, use the following URI:

spark:rdd?rdd=#testFileRdd&rddCallback=#transformation

Where rdd option refers to the name of an RDD instance (subclass of org.apache.spark.api.java.JavaRDDLike) from a Camel registry, while rddCallback refers to the implementation of org.apache.camel.component.spark.RddCallback interface (also from a registry). RDD callback provides a single method used to apply incoming messages against the given RDD. Results of callback computations are saved as a body to an exchange.

public interface RddCallback<T> {
    T onRdd(JavaRDDLike rdd, Object... payloads);
}

The following snippet demonstrates how to send message as an input to the job and return results:

String pattern = "job input";
long linesCount = producerTemplate.requestBody("spark:rdd?rdd=#myRdd&rddCallback=#countLinesContaining", pattern, long.class);

The RDD callback for the snippet above registered as Spring bean could look as follows:

@Bean
RddCallback<Long> countLinesContaining() {
    return new RddCallback<Long>() {
        Long onRdd(JavaRDDLike rdd, Object... payloads) {
            String pattern = (String) payloads[0];
            return rdd.filter({line -> line.contains(pattern)}).count();
        }
    }
}

The RDD definition in Spring could looks as follows:

@Bean
JavaRDDLike myRdd(JavaSparkContext sparkContext) {
  return sparkContext.textFile("testrdd.txt");
}

RDD jobs options

Option	Description	Default value
`rdd`	RDD instance (subclass of `org.apache.spark.api.java.JavaRDDLike`).	`null`
`rddCallback`	Instance of `org.apache.camel.component.spark.RddCallback` interface.	`null`

Void RDD callbacks

If your RDD callback doesn't return any value back to a Camel pipeline, you can either return null value or use VoidRddCallback base class:

@Bean
RddCallback<Void> rddCallback() {
  return new VoidRddCallback() {
        @Override
        public void doOnRdd(JavaRDDLike rdd, Object... payloads) {
            rdd.saveAsTextFile(output.getAbsolutePath());
        }
    };
}

Converting RDD callbacks

If you know what type of the input data will be sent to the RDD callback, you can use ConvertingRddCallback and let Camel to automatically convert incoming messages before inserting those into the callback:

@Bean
RddCallback<Long> rddCallback(CamelContext context) {
  return new ConvertingRddCallback<Long>(context, int.class, int.class) {
            @Override
            public Long doOnRdd(JavaRDDLike rdd, Object... payloads) {
                return rdd.count() * (int) payloads[0] * (int) payloads[1];
            }
        };
    };
}

Annotated RDD callbacks

Probably the easiest way to work with the RDD callbacks is to provide class with method marked with @RddCallback annotation:

import static org.apache.camel.component.spark.annotations.AnnotatedRddCallback.annotatedRddCallback;
 
@Bean
RddCallback<Long> rddCallback() {
    return annotatedRddCallback(new MyTransformation());
}
 
...
 
import org.apache.camel.component.spark.annotation.RddCallback;
 
public class MyTransformation {
 
    @RddCallback
    long countLines(JavaRDD<String> textFile, int first, int second) {
        return textFile.count() * first * second;
    }
 
}

If you will pass CamelContext to the annotated RDD callback factory method, the created callback will be able to convert incoming payloads to match the parameters of the annotated method:

import static org.apache.camel.component.spark.annotations.AnnotatedRddCallback.annotatedRddCallback;
 
@Bean
RddCallback<Long> rddCallback(CamelContext camelContext) {
    return annotatedRddCallback(new MyTransformation(), camelContext);
}
 
...

 
import org.apache.camel.component.spark.annotation.RddCallback;
 
public class MyTransformation {
 
    @RddCallback
    long countLines(JavaRDD<String> textFile, int first, int second) {
        return textFile.count() * first * second;
    }
 
}
 
...
 
// Convert String "10" to integer
long result = producerTemplate.requestBody("spark:rdd?rdd=#rdd&rddCallback=#rddCallback" Arrays.asList(10, "10"), long.class);

DataFrame jobs

Instead of working with RDDs Spark component can work with DataFrames as well.

To invoke an DataFrame job, use the following URI:

spark:dataframe?dataFrame=#testDataFrame&dataFrameCallback=#transformation

Where dataFrame option refers to the name of a DataFrame instance (instance of org.apache.spark.sql.DataFrame) from a Camel registry, while dataFrameCallback refers to the implementation of org.apache.camel.component.spark.DataFrameCallback interface (also from a registry). DataFrame callback provides a single method used to apply incoming messages against the given DataFrame. Results of callback computations are saved as a body to an exchange.

public interface DataFrameCallback<T> {
    T onDataFrame(DataFrame dataFrame, Object... payloads);
}

The following snippet demonstrates how to send message as an input to a job and return results:

String model = "Micra";
long linesCount = producerTemplate.requestBody("spark:dataFrame?dataFrame=#cars&dataFrameCallback=#findCarWithModel", model, long.class);

The DataFrame callback for the snippet above registered as Spring bean could look as follows:

@Bean
RddCallback<Long> findCarWithModel() {
    return new DataFrameCallback<Long>() {
    	@Override
    	public Long onDataFrame(DataFrame dataFrame, Object... payloads) {
        	String model = (String) payloads[0];
        	return dataFrame.where(dataFrame.col("model").eqNullSafe(model)).count();
    	}
	};
}

The DataFrame definition in Spring could looks as follows:

@Bean
DataFrame cars(HiveContext hiveContext) {
  	DataFrame jsonCars = hiveContext.read().json("/var/data/cars.json");
 	jsonCars.registerTempTable("cars");
	return jsonCars;
}

DataFrame jobs options

Option	Description	Default value
`dataFrame`	DataFrame instance (subclass of `org.apache.spark.sql.DataFrame`).	`null`
`dataFrameCallback`	Instance of `org.apache.camel.component.spark.DataFrameCallback` interface.	`null`

Hive jobs

Instead of working with RDDs or DataFrame Spark component can also receive Hive SQL queries as payloads. To send Hive query to Spark component, use the following URI:

spark:hive

The following snippet demonstrates how to send message as an input to a job and return results:

long carsCount = template.requestBody("spark:hive?collect=false", "SELECT * FROM cars", Long.class);
List<Row> cars = template.requestBody("spark:hive", "SELECT * FROM cars", List.class);

The table we want to execute query against should be registered in a HiveContext before we query it. For example in Spring such registration could look as follows:

@Bean
DataFrame cars(HiveContext hiveContext) {
  	DataFrame jsonCars = hiveContext.read().json("/var/data/cars.json");
 	jsonCars.registerTempTable("cars");
	return jsonCars;
}

Hive jobs options

Option	Description	Default value
`collect`	Indicates if results should be collected (as a list of `org.apache.spark.sql.Row` instances) or if `count()` should be called against those.	`true`