org.apache.flink.optimizer.testfunctions.IdentityGroupReducer Java Examples

@Test
public void testReduce() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();

	try {
		OptimizedPlan oPlan = compileNoStats(plan);
		JobGraphGenerator jobGen = new JobGraphGenerator();
		jobGen.compileJobGraph(oPlan);
	} catch(CompilerException ce) {
		ce.printStackTrace();
		fail("The pact compiler is unable to compile this plan correctly");
	}
}
 

/**
 * Source -> Map -> Reduce -> Cross -> Reduce -> Cross -> Reduce ->
 * |--------------------------/                  /
 * |--------------------------------------------/
 * 
 * First cross has SameKeyFirst output contract
 */
@Test
public void testTicket158() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.map(new IdentityMapper<Long>()).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.cross(set1).with(new IdentityCrosser<Long>()).withForwardedFieldsFirst("*").name("Cross1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce2")
			.cross(set1).with(new IdentityCrosser<Long>()).name("Cross2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce3")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();
	OptimizedPlan oPlan = compileNoStats(plan);

	JobGraphGenerator jobGen = new JobGraphGenerator();
	jobGen.compileJobGraph(oPlan);
}
 

@Test
public void testReduce() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();

	try {
		OptimizedPlan oPlan = compileNoStats(plan);
		JobGraphGenerator jobGen = new JobGraphGenerator();
		jobGen.compileJobGraph(oPlan);
	} catch(CompilerException ce) {
		ce.printStackTrace();
		fail("The pact compiler is unable to compile this plan correctly");
	}
}
 

@Test
public void testReduce() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();

	try {
		OptimizedPlan oPlan = compileNoStats(plan);
		JobGraphGenerator jobGen = new JobGraphGenerator();
		jobGen.compileJobGraph(oPlan);
	} catch(CompilerException ce) {
		ce.printStackTrace();
		fail("The pact compiler is unable to compile this plan correctly");
	}
}
 

/**
 * Source -> Map -> Reduce -> Cross -> Reduce -> Cross -> Reduce ->
 * |--------------------------/                  /
 * |--------------------------------------------/
 * 
 * First cross has SameKeyFirst output contract
 */
@Test
public void testTicket158() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.map(new IdentityMapper<Long>()).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.cross(set1).with(new IdentityCrosser<Long>()).withForwardedFieldsFirst("*").name("Cross1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce2")
			.cross(set1).with(new IdentityCrosser<Long>()).name("Cross2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce3")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();
	OptimizedPlan oPlan = compileNoStats(plan);

	JobGraphGenerator jobGen = new JobGraphGenerator();
	jobGen.compileJobGraph(oPlan);
}
 

/**
 * Source -> Map -> Reduce -> Cross -> Reduce -> Cross -> Reduce ->
 * |--------------------------/                  /
 * |--------------------------------------------/
 * 
 * First cross has SameKeyFirst output contract
 */
@Test
public void testTicket158() {
	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(DEFAULT_PARALLELISM);
	DataSet<Long> set1 = env.generateSequence(0,1);

	set1.map(new IdentityMapper<Long>()).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce1")
			.cross(set1).with(new IdentityCrosser<Long>()).withForwardedFieldsFirst("*").name("Cross1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce2")
			.cross(set1).with(new IdentityCrosser<Long>()).name("Cross2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce3")
			.output(new DiscardingOutputFormat<Long>()).name("Sink");

	Plan plan = env.createProgramPlan();
	OptimizedPlan oPlan = compileNoStats(plan);

	JobGraphGenerator jobGen = new JobGraphGenerator();
	jobGen.compileJobGraph(oPlan);
}
 

@Test
public void testBranchingUnion() {
	try {
		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Long> source1 = env.generateSequence(0,1);
		DataSet<Long> source2 = env.generateSequence(0,1);

		DataSet<Long> join1 = source1.join(source2).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 1");

		DataSet<Long> map1 = join1.map(new IdentityMapper<Long>()).name("Map 1");

		DataSet<Long> reduce1 = map1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 1");

		DataSet<Long> reduce2 = join1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 2");

		DataSet<Long> map2 = join1.map(new IdentityMapper<Long>()).name("Map 2");

		DataSet<Long> map3 = map2.map(new IdentityMapper<Long>()).name("Map 3");

		DataSet<Long> join2 = reduce1.union(reduce2).union(map2).union(map3)
				.join(map2, JoinHint.REPARTITION_SORT_MERGE).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 2");

		join2.output(new DiscardingOutputFormat<Long>());

		Plan plan = env.createProgramPlan();
		OptimizedPlan oPlan = compileNoStats(plan);

		JobGraphGenerator jobGen = new JobGraphGenerator();
		
		//Compile plan to verify that no error is thrown
		jobGen.compileJobGraph(oPlan);
	} catch (Exception e) {
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void checkPropertyHandlingWithDecreasingParallelism() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);

	env
		.generateSequence(0, 1).setParallelism(p * 2)
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map1")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce1")
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce2")
		.output(new DiscardingOutputFormat<Long>()).setParallelism(p).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);

	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();

	Assert.assertTrue("The no sorting local strategy.",
			LocalStrategy.SORT == red2Node.getInput().getLocalStrategy() ||
					LocalStrategy.SORT == map2Node.getInput().getLocalStrategy());

	Assert.assertTrue("The no partitioning ship strategy.",
			ShipStrategyType.PARTITION_HASH == red2Node.getInput().getShipStrategy() ||
					ShipStrategyType.PARTITION_HASH == map2Node.getInput().getShipStrategy());
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, such that more tasks are on one instance.
 * Expected to re-establish partitioning between map and reduce via a local hash.
 */
@Test
public void checkPropertyHandlingWithIncreasingLocalParallelism() {
	final int p = DEFAULT_PARALLELISM * 2;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertTrue("Invalid ship strategy for an operator.", 
			(ShipStrategyType.PARTITION_RANDOM ==  mapIn && ShipStrategyType.PARTITION_HASH == reduceIn) || 
			(ShipStrategyType.PARTITION_HASH == mapIn && ShipStrategyType.FORWARD == reduceIn));
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 2nd map and 2nd reduce, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between map and reduce (hash).
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism2() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, reduceIn);
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between reduce and map, via hash, because random is a full network
 * transit as well.
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism1() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType redIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, redIn);
}
 

private Plan getTestPlan(boolean joinPreservesSolutionSet, boolean mapBeforeSolutionDelta) {

		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Tuple2<Long, Long>> solSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Solution Set");
		DataSet<Tuple2<Long, Long>> workSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Workset");
		DataSet<Tuple2<Long, Long>> invariantInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Invariant Input");

		DeltaIteration<Tuple2<Long, Long>, Tuple2<Long, Long>> deltaIt = solSetInput.iterateDelta(workSetInput, 100, 0).name(ITERATION_NAME);

		DataSet<Tuple2<Long, Long>> join1 = deltaIt.getWorkset().join(invariantInput).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.withForwardedFieldsFirst("*").name(JOIN_WITH_INVARIANT_NAME);

		DataSet<Tuple2<Long, Long>> join2 = deltaIt.getSolutionSet().join(join1).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.name(JOIN_WITH_SOLUTION_SET);
		if(joinPreservesSolutionSet) {
			((JoinOperator<?,?,?>)join2).withForwardedFieldsFirst("*");
		}

		DataSet<Tuple2<Long, Long>> nextWorkset = join2.groupBy(0).reduceGroup(new IdentityGroupReducer<Tuple2<Long, Long>>())
				.withForwardedFields("*").name(NEXT_WORKSET_REDUCER_NAME);

		if(mapBeforeSolutionDelta) {

			DataSet<Tuple2<Long, Long>> mapper = join2.map(new IdentityMapper<Tuple2<Long, Long>>())
					.withForwardedFields("*").name(SOLUTION_DELTA_MAPPER_NAME);

			deltaIt.closeWith(mapper, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long,Long>>());
		}
		else {
			deltaIt.closeWith(join2, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
		}

		return env.createProgramPlan();
	}
 

@Test
public void testMultipleIterations() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);
	
	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
	
	DataSet<String> reduced = input
			.map(new IdentityMapper<String>())
			.reduceGroup(new Top1GroupReducer<String>());
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()).withBroadcastSet(reduced, "bc1"))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()).withBroadcastSet(reduced, "bc2"))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()).withBroadcastSet(reduced, "bc3"))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void testMultipleIterationsWithClosueBCVars() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);

	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void testMultipleIterationsWithClosueBCVars() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);

	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void testMultipleIterations() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);
	
	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
	
	DataSet<String> reduced = input
			.map(new IdentityMapper<String>())
			.reduceGroup(new Top1GroupReducer<String>());
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()).withBroadcastSet(reduced, "bc1"))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()).withBroadcastSet(reduced, "bc2"))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()).withBroadcastSet(reduced, "bc3"))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void testBranchingUnion() {
	try {
		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Long> source1 = env.generateSequence(0,1);
		DataSet<Long> source2 = env.generateSequence(0,1);

		DataSet<Long> join1 = source1.join(source2).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 1");

		DataSet<Long> map1 = join1.map(new IdentityMapper<Long>()).name("Map 1");

		DataSet<Long> reduce1 = map1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 1");

		DataSet<Long> reduce2 = join1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 2");

		DataSet<Long> map2 = join1.map(new IdentityMapper<Long>()).name("Map 2");

		DataSet<Long> map3 = map2.map(new IdentityMapper<Long>()).name("Map 3");

		DataSet<Long> join2 = reduce1.union(reduce2).union(map2).union(map3)
				.join(map2, JoinHint.REPARTITION_SORT_MERGE).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 2");

		join2.output(new DiscardingOutputFormat<Long>());

		Plan plan = env.createProgramPlan();
		OptimizedPlan oPlan = compileNoStats(plan);

		JobGraphGenerator jobGen = new JobGraphGenerator();
		
		//Compile plan to verify that no error is thrown
		jobGen.compileJobGraph(oPlan);
	} catch (Exception e) {
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between reduce and map, via hash, because random is a full network
 * transit as well.
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism1() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType redIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, redIn);
}
 

@Test
public void testMultipleIterations() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);
	
	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
	
	DataSet<String> reduced = input
			.map(new IdentityMapper<String>())
			.reduceGroup(new Top1GroupReducer<String>());
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()).withBroadcastSet(reduced, "bc1"))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()).withBroadcastSet(reduced, "bc2"))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()).withBroadcastSet(reduced, "bc3"))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

private Plan getTestPlan(boolean joinPreservesSolutionSet, boolean mapBeforeSolutionDelta) {

		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Tuple2<Long, Long>> solSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Solution Set");
		DataSet<Tuple2<Long, Long>> workSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Workset");
		DataSet<Tuple2<Long, Long>> invariantInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Invariant Input");

		DeltaIteration<Tuple2<Long, Long>, Tuple2<Long, Long>> deltaIt = solSetInput.iterateDelta(workSetInput, 100, 0).name(ITERATION_NAME);

		DataSet<Tuple2<Long, Long>> join1 = deltaIt.getWorkset().join(invariantInput).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.withForwardedFieldsFirst("*").name(JOIN_WITH_INVARIANT_NAME);

		DataSet<Tuple2<Long, Long>> join2 = deltaIt.getSolutionSet().join(join1).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.name(JOIN_WITH_SOLUTION_SET);
		if(joinPreservesSolutionSet) {
			((JoinOperator<?,?,?>)join2).withForwardedFieldsFirst("*");
		}

		DataSet<Tuple2<Long, Long>> nextWorkset = join2.groupBy(0).reduceGroup(new IdentityGroupReducer<Tuple2<Long, Long>>())
				.withForwardedFields("*").name(NEXT_WORKSET_REDUCER_NAME);

		if(mapBeforeSolutionDelta) {

			DataSet<Tuple2<Long, Long>> mapper = join2.map(new IdentityMapper<Tuple2<Long, Long>>())
					.withForwardedFields("*").name(SOLUTION_DELTA_MAPPER_NAME);

			deltaIt.closeWith(mapper, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long,Long>>());
		}
		else {
			deltaIt.closeWith(join2, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
		}

		return env.createProgramPlan();
	}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 2nd map and 2nd reduce, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between map and reduce (hash).
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism2() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, reduceIn);
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, such that more tasks are on one instance.
 * Expected to re-establish partitioning between map and reduce via a local hash.
 */
@Test
public void checkPropertyHandlingWithIncreasingLocalParallelism() {
	final int p = DEFAULT_PARALLELISM * 2;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertTrue("Invalid ship strategy for an operator.", 
			(ShipStrategyType.PARTITION_RANDOM ==  mapIn && ShipStrategyType.PARTITION_HASH == reduceIn) || 
			(ShipStrategyType.PARTITION_HASH == mapIn && ShipStrategyType.FORWARD == reduceIn));
}
 

@Test
public void checkPropertyHandlingWithDecreasingParallelism() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);

	env
		.generateSequence(0, 1).setParallelism(p * 2)
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map1")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce1")
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce2")
		.output(new DiscardingOutputFormat<Long>()).setParallelism(p).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);

	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();

	Assert.assertTrue("The no sorting local strategy.",
			LocalStrategy.SORT == red2Node.getInput().getLocalStrategy() ||
					LocalStrategy.SORT == map2Node.getInput().getLocalStrategy());

	Assert.assertTrue("The no partitioning ship strategy.",
			ShipStrategyType.PARTITION_HASH == red2Node.getInput().getShipStrategy() ||
					ShipStrategyType.PARTITION_HASH == map2Node.getInput().getShipStrategy());
}
 

@Test
public void testBranchingUnion() {
	try {
		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Long> source1 = env.generateSequence(0,1);
		DataSet<Long> source2 = env.generateSequence(0,1);

		DataSet<Long> join1 = source1.join(source2).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 1");

		DataSet<Long> map1 = join1.map(new IdentityMapper<Long>()).name("Map 1");

		DataSet<Long> reduce1 = map1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 1");

		DataSet<Long> reduce2 = join1.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>()).name("Reduce 2");

		DataSet<Long> map2 = join1.map(new IdentityMapper<Long>()).name("Map 2");

		DataSet<Long> map3 = map2.map(new IdentityMapper<Long>()).name("Map 3");

		DataSet<Long> join2 = reduce1.union(reduce2).union(map2).union(map3)
				.join(map2, JoinHint.REPARTITION_SORT_MERGE).where("*").equalTo("*")
				.with(new IdentityJoiner<Long>()).name("Join 2");

		join2.output(new DiscardingOutputFormat<Long>());

		Plan plan = env.createProgramPlan();
		OptimizedPlan oPlan = compileNoStats(plan);

		JobGraphGenerator jobGen = new JobGraphGenerator();
		
		//Compile plan to verify that no error is thrown
		jobGen.compileJobGraph(oPlan);
	} catch (Exception e) {
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

@Test
public void checkPropertyHandlingWithDecreasingParallelism() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);

	env
		.generateSequence(0, 1).setParallelism(p * 2)
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map1")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce1")
		.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
		.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce2")
		.output(new DiscardingOutputFormat<Long>()).setParallelism(p).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);

	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();

	Assert.assertTrue("The no sorting local strategy.",
			LocalStrategy.SORT == red2Node.getInput().getLocalStrategy() ||
					LocalStrategy.SORT == map2Node.getInput().getLocalStrategy());

	Assert.assertTrue("The no partitioning ship strategy.",
			ShipStrategyType.PARTITION_HASH == red2Node.getInput().getShipStrategy() ||
					ShipStrategyType.PARTITION_HASH == map2Node.getInput().getShipStrategy());
}
 

@Test
public void testMultipleIterationsWithClosueBCVars() {
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(100);

	DataSet<String> input = env.readTextFile(IN_FILE).name("source1");
		
	IterativeDataSet<String> iteration1 = input.iterate(100);
	IterativeDataSet<String> iteration2 = input.iterate(20);
	IterativeDataSet<String> iteration3 = input.iterate(17);
	
	
	iteration1.closeWith(iteration1.map(new IdentityMapper<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration2.closeWith(iteration2.reduceGroup(new Top1GroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	iteration3.closeWith(iteration3.reduceGroup(new IdentityGroupReducer<String>()))
			.output(new DiscardingOutputFormat<String>());
	
	Plan plan = env.createProgramPlan();
	
	try{
		compileNoStats(plan);
	}catch(Exception e){
		e.printStackTrace();
		Assert.fail(e.getMessage());
	}
}
 

private Plan getTestPlan(boolean joinPreservesSolutionSet, boolean mapBeforeSolutionDelta) {

		// construct the plan
		ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
		env.setParallelism(DEFAULT_PARALLELISM);
		DataSet<Tuple2<Long, Long>> solSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Solution Set");
		DataSet<Tuple2<Long, Long>> workSetInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Workset");
		DataSet<Tuple2<Long, Long>> invariantInput = env.readCsvFile("/tmp/sol.csv").types(Long.class, Long.class).name("Invariant Input");

		DeltaIteration<Tuple2<Long, Long>, Tuple2<Long, Long>> deltaIt = solSetInput.iterateDelta(workSetInput, 100, 0).name(ITERATION_NAME);

		DataSet<Tuple2<Long, Long>> join1 = deltaIt.getWorkset().join(invariantInput).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.withForwardedFieldsFirst("*").name(JOIN_WITH_INVARIANT_NAME);

		DataSet<Tuple2<Long, Long>> join2 = deltaIt.getSolutionSet().join(join1).where(0).equalTo(0)
				.with(new IdentityJoiner<Tuple2<Long, Long>>())
				.name(JOIN_WITH_SOLUTION_SET);
		if(joinPreservesSolutionSet) {
			((JoinOperator<?,?,?>)join2).withForwardedFieldsFirst("*");
		}

		DataSet<Tuple2<Long, Long>> nextWorkset = join2.groupBy(0).reduceGroup(new IdentityGroupReducer<Tuple2<Long, Long>>())
				.withForwardedFields("*").name(NEXT_WORKSET_REDUCER_NAME);

		if(mapBeforeSolutionDelta) {

			DataSet<Tuple2<Long, Long>> mapper = join2.map(new IdentityMapper<Tuple2<Long, Long>>())
					.withForwardedFields("*").name(SOLUTION_DELTA_MAPPER_NAME);

			deltaIt.closeWith(mapper, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long,Long>>());
		}
		else {
			deltaIt.closeWith(join2, nextWorkset)
					.output(new DiscardingOutputFormat<Tuple2<Long, Long>>());
		}

		return env.createProgramPlan();
	}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between reduce and map, via hash, because random is a full network
 * transit as well.
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism1() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
				.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType redIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, redIn);
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 2nd map and 2nd reduce, so the hash partitioning from 1st reduce is not reusable.
 * Expected to re-establish partitioning between map and reduce (hash).
 */
@Test
public void checkPropertyHandlingWithIncreasingGlobalParallelism2() {
	final int p = DEFAULT_PARALLELISM;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.FORWARD, mapIn);
	Assert.assertEquals("Invalid ship strategy for an operator.", ShipStrategyType.PARTITION_HASH, reduceIn);
}
 

/**
 * Simple Job: Map -> Reduce -> Map -> Reduce. All functions preserve all fields (hence all properties).
 * 
 * Increases parallelism between 1st reduce and 2nd map, such that more tasks are on one instance.
 * Expected to re-establish partitioning between map and reduce via a local hash.
 */
@Test
public void checkPropertyHandlingWithIncreasingLocalParallelism() {
	final int p = DEFAULT_PARALLELISM * 2;

	// construct the plan
	ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
	env.setParallelism(p);
	DataSet<Long> set1 = env.generateSequence(0,1).setParallelism(p);

	set1.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p).name("Map1")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p).name("Reduce1")
			.map(new IdentityMapper<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Map2")
			.groupBy("*").reduceGroup(new IdentityGroupReducer<Long>())
			.withForwardedFields("*").setParallelism(p * 2).name("Reduce2")
			.output(new DiscardingOutputFormat<Long>()).setParallelism(p * 2).name("Sink");

	Plan plan = env.createProgramPlan();
	// submit the plan to the compiler
	OptimizedPlan oPlan = compileNoStats(plan);
	
	// check the optimized Plan
	// when reducer 1 distributes its data across the instances of map2, it needs to employ a local hash method,
	// because map2 has twice as many instances and key/value pairs with the same key need to be processed by the same
	// mapper respectively reducer
	SinkPlanNode sinkNode = oPlan.getDataSinks().iterator().next();
	SingleInputPlanNode red2Node = (SingleInputPlanNode) sinkNode.getPredecessor();
	SingleInputPlanNode map2Node = (SingleInputPlanNode) red2Node.getPredecessor();
	
	ShipStrategyType mapIn = map2Node.getInput().getShipStrategy();
	ShipStrategyType reduceIn = red2Node.getInput().getShipStrategy();
	
	Assert.assertTrue("Invalid ship strategy for an operator.", 
			(ShipStrategyType.PARTITION_RANDOM ==  mapIn && ShipStrategyType.PARTITION_HASH == reduceIn) || 
			(ShipStrategyType.PARTITION_HASH == mapIn && ShipStrategyType.FORWARD == reduceIn));
}