@Test
public void testCorrectTranslation() {
try {
final String JOB_NAME = "Test JobName";
final String ITERATION_NAME = "Test Name";
final String BEFORE_NEXT_WORKSET_MAP = "Some Mapper";
final String AGGREGATOR_NAME = "AggregatorName";
final int[] ITERATION_KEYS = new int[] {2};
final int NUM_ITERATIONS = 13;
final int DEFAULT_DOP = 133;
final int ITERATION_DOP = 77;
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// ------------ construct the test program ------------------
{
env.setDegreeOfParallelism(DEFAULT_DOP);
@SuppressWarnings("unchecked")
DataSet<Tuple3<Double, Long, String>> initialSolutionSet =
env.fromElements(new Tuple3<Double, Long, String>(3.44, 5L, "abc"));
@SuppressWarnings("unchecked")
DataSet<Tuple2<Double, String>> initialWorkSet =
env.fromElements(new Tuple2<Double, String>(1.23, "abc"));
DeltaIteration<Tuple3<Double, Long, String>, Tuple2<Double, String>> iteration =
initialSolutionSet.iterateDelta(initialWorkSet, NUM_ITERATIONS, ITERATION_KEYS);
iteration.name(ITERATION_NAME).parallelism(ITERATION_DOP);
iteration.registerAggregator(AGGREGATOR_NAME, new LongSumAggregator());
// test that multiple workset consumers are supported
DataSet<Tuple2<Double, String>> worksetSelfJoin =
iteration
.getWorkset()
.map(new IdentityMapper<Tuple2<Double, String>>())
.join(iteration.getWorkset())
.where(1)
.equalTo(1)
.projectFirst(0, 1)
.types(Double.class, String.class);
DataSet<Tuple3<Double, Long, String>> joined =
worksetSelfJoin
.join(iteration.getSolutionSet())
.where(1)
.equalTo(2)
.with(new SolutionWorksetJoin());
DataSet<Tuple3<Double, Long, String>> result =
iteration.closeWith(
joined, joined.map(new NextWorksetMapper()).name(BEFORE_NEXT_WORKSET_MAP));
result.print();
result.writeAsText("/dev/null");
}
Plan p = env.createProgramPlan(JOB_NAME);
// ------------- validate the plan ----------------
assertEquals(JOB_NAME, p.getJobName());
assertEquals(DEFAULT_DOP, p.getDefaultParallelism());
// validate the iteration
GenericDataSinkBase<?> sink1, sink2;
{
Iterator<? extends GenericDataSinkBase<?>> sinks = p.getDataSinks().iterator();
sink1 = sinks.next();
sink2 = sinks.next();
}
DeltaIterationBase<?, ?> iteration = (DeltaIterationBase<?, ?>) sink1.getInput();
// check that multi consumer translation works for iterations
assertEquals(iteration, sink2.getInput());
// check the basic iteration properties
assertEquals(NUM_ITERATIONS, iteration.getMaximumNumberOfIterations());
assertArrayEquals(ITERATION_KEYS, iteration.getSolutionSetKeyFields());
assertEquals(ITERATION_DOP, iteration.getDegreeOfParallelism());
assertEquals(ITERATION_NAME, iteration.getName());
MapOperatorBase<?, ?, ?> nextWorksetMapper =
(MapOperatorBase<?, ?, ?>) iteration.getNextWorkset();
JoinOperatorBase<?, ?, ?, ?> solutionSetJoin =
(JoinOperatorBase<?, ?, ?, ?>) iteration.getSolutionSetDelta();
JoinOperatorBase<?, ?, ?, ?> worksetSelfJoin =
(JoinOperatorBase<?, ?, ?, ?>) solutionSetJoin.getFirstInput();
MapOperatorBase<?, ?, ?> worksetMapper =
(MapOperatorBase<?, ?, ?>) worksetSelfJoin.getFirstInput();
assertEquals(IdentityMapper.class, worksetMapper.getUserCodeWrapper().getUserCodeClass());
assertEquals(
NextWorksetMapper.class, nextWorksetMapper.getUserCodeWrapper().getUserCodeClass());
if (solutionSetJoin.getUserCodeWrapper().getUserCodeObject() instanceof WrappingFunction) {
WrappingFunction<?> wf =
(WrappingFunction<?>) solutionSetJoin.getUserCodeWrapper().getUserCodeObject();
assertEquals(SolutionWorksetJoin.class, wf.getWrappedFunction().getClass());
} else {
assertEquals(
SolutionWorksetJoin.class, solutionSetJoin.getUserCodeWrapper().getUserCodeClass());
}
assertEquals(BEFORE_NEXT_WORKSET_MAP, nextWorksetMapper.getName());
assertEquals(
AGGREGATOR_NAME,
iteration.getAggregators().getAllRegisteredAggregators().iterator().next().getName());
} catch (Exception e) {
System.err.println(e.getMessage());
e.printStackTrace();
fail(e.getMessage());
}
}
@SuppressWarnings("unchecked")
private <T> List<T> executeDeltaIteration(DeltaIterationBase<?, ?> iteration) throws Exception {
Operator<?> solutionInput = iteration.getInitialSolutionSet();
Operator<?> worksetInput = iteration.getInitialWorkset();
if (solutionInput == null) {
throw new InvalidProgramException(
"The delta iteration " + iteration.getName() + " has no initial solution set.");
}
if (worksetInput == null) {
throw new InvalidProgramException(
"The delta iteration " + iteration.getName() + " has no initial workset.");
}
if (iteration.getSolutionSetDelta() == null) {
throw new InvalidProgramException(
"The iteration "
+ iteration.getName()
+ " has no solution set delta defined (is not closed).");
}
if (iteration.getNextWorkset() == null) {
throw new InvalidProgramException(
"The iteration " + iteration.getName() + " has no workset defined (is not closed).");
}
List<T> solutionInputData = (List<T>) execute(solutionInput);
List<T> worksetInputData = (List<T>) execute(worksetInput);
// get the operators that are iterative
Set<Operator<?>> dynamics = new LinkedHashSet<Operator<?>>();
DynamicPathCollector dynCollector = new DynamicPathCollector(dynamics);
iteration.getSolutionSetDelta().accept(dynCollector);
iteration.getNextWorkset().accept(dynCollector);
BinaryOperatorInformation<?, ?, ?> operatorInfo = iteration.getOperatorInfo();
TypeInformation<?> solutionType = operatorInfo.getFirstInputType();
int[] keyColumns = iteration.getSolutionSetKeyFields();
boolean[] inputOrderings = new boolean[keyColumns.length];
TypeComparator<T> inputComparator =
((CompositeType<T>) solutionType)
.createComparator(keyColumns, inputOrderings, 0, executionConfig);
Map<TypeComparable<T>, T> solutionMap =
new HashMap<TypeComparable<T>, T>(solutionInputData.size());
// fill the solution from the initial input
for (T delta : solutionInputData) {
TypeComparable<T> wrapper = new TypeComparable<T>(delta, inputComparator);
solutionMap.put(wrapper, delta);
}
List<?> currentWorkset = worksetInputData;
// register the aggregators
for (AggregatorWithName<?> a : iteration.getAggregators().getAllRegisteredAggregators()) {
aggregators.put(a.getName(), a.getAggregator());
}
String convCriterionAggName =
iteration.getAggregators().getConvergenceCriterionAggregatorName();
ConvergenceCriterion<Value> convCriterion =
(ConvergenceCriterion<Value>) iteration.getAggregators().getConvergenceCriterion();
final int maxIterations = iteration.getMaximumNumberOfIterations();
for (int superstep = 1; superstep <= maxIterations; superstep++) {
List<T> currentSolution = new ArrayList<T>(solutionMap.size());
currentSolution.addAll(solutionMap.values());
// set the input to the current partial solution
this.intermediateResults.put(iteration.getSolutionSet(), currentSolution);
this.intermediateResults.put(iteration.getWorkset(), currentWorkset);
// set the superstep number
iterationSuperstep = superstep;
// grab the current iteration result
List<T> solutionSetDelta = (List<T>) execute(iteration.getSolutionSetDelta(), superstep);
this.intermediateResults.put(iteration.getSolutionSetDelta(), solutionSetDelta);
// update the solution
for (T delta : solutionSetDelta) {
TypeComparable<T> wrapper = new TypeComparable<T>(delta, inputComparator);
solutionMap.put(wrapper, delta);
}
currentWorkset = execute(iteration.getNextWorkset(), superstep);
if (currentWorkset.isEmpty()) {
break;
}
// evaluate the aggregator convergence criterion
if (convCriterion != null && convCriterionAggName != null) {
Value v = aggregators.get(convCriterionAggName).getAggregate();
if (convCriterion.isConverged(superstep, v)) {
break;
}
}
// clear the dynamic results
for (Operator<?> o : dynamics) {
intermediateResults.remove(o);
}
// set the previous iteration's aggregates and reset the aggregators
for (Map.Entry<String, Aggregator<?>> e : aggregators.entrySet()) {
previousAggregates.put(e.getKey(), e.getValue().getAggregate());
e.getValue().reset();
}
}
previousAggregates.clear();
aggregators.clear();
List<T> currentSolution = new ArrayList<T>(solutionMap.size());
currentSolution.addAll(solutionMap.values());
return currentSolution;
}
