use of org.apache.flink.api.common.aggregators.Aggregator in project flink by apache.
the class CollectionExecutor method executeBulkIteration.
@SuppressWarnings("unchecked")
private <T> List<T> executeBulkIteration(BulkIterationBase<?> iteration) throws Exception {
Operator<?> inputOp = iteration.getInput();
if (inputOp == null) {
throw new InvalidProgramException("The iteration " + iteration.getName() + " has no input (initial partial solution).");
}
if (iteration.getNextPartialSolution() == null) {
throw new InvalidProgramException("The iteration " + iteration.getName() + " has no next partial solution defined (is not closed).");
}
List<T> inputData = (List<T>) execute(inputOp);
// get the operators that are iterative
Set<Operator<?>> dynamics = new LinkedHashSet<Operator<?>>();
DynamicPathCollector dynCollector = new DynamicPathCollector(dynamics);
iteration.getNextPartialSolution().accept(dynCollector);
if (iteration.getTerminationCriterion() != null) {
iteration.getTerminationCriterion().accept(dynCollector);
}
// 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();
List<T> currentResult = inputData;
final int maxIterations = iteration.getMaximumNumberOfIterations();
for (int superstep = 1; superstep <= maxIterations; superstep++) {
// set the input to the current partial solution
this.intermediateResults.put(iteration.getPartialSolution(), currentResult);
// set the superstep number
iterationSuperstep = superstep;
// grab the current iteration result
currentResult = (List<T>) execute(iteration.getNextPartialSolution(), superstep);
// evaluate the termination criterion
if (iteration.getTerminationCriterion() != null) {
execute(iteration.getTerminationCriterion(), superstep);
}
// 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();
return currentResult;
}
use of org.apache.flink.api.common.aggregators.Aggregator in project flink by apache.
the class CollectionExecutor method executeDeltaIteration.
@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;
}
use of org.apache.flink.api.common.aggregators.Aggregator in project flink by apache.
the class CollectionExecutor method executeDeltaIteration.
@SuppressWarnings("unchecked")
private <T> List<T> executeDeltaIteration(DeltaIterationBase<?, ?> iteration, JobID jobID) 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, jobID);
List<T> worksetInputData = (List<T>) execute(worksetInput, jobID);
// 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, jobID);
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, jobID);
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;
}
use of org.apache.flink.api.common.aggregators.Aggregator in project flink by apache.
the class CollectionExecutor method executeBulkIteration.
@SuppressWarnings("unchecked")
private <T> List<T> executeBulkIteration(BulkIterationBase<?> iteration, JobID jobID) throws Exception {
Operator<?> inputOp = iteration.getInput();
if (inputOp == null) {
throw new InvalidProgramException("The iteration " + iteration.getName() + " has no input (initial partial solution).");
}
if (iteration.getNextPartialSolution() == null) {
throw new InvalidProgramException("The iteration " + iteration.getName() + " has no next partial solution defined (is not closed).");
}
List<T> inputData = (List<T>) execute(inputOp, jobID);
// get the operators that are iterative
Set<Operator<?>> dynamics = new LinkedHashSet<Operator<?>>();
DynamicPathCollector dynCollector = new DynamicPathCollector(dynamics);
iteration.getNextPartialSolution().accept(dynCollector);
if (iteration.getTerminationCriterion() != null) {
iteration.getTerminationCriterion().accept(dynCollector);
}
// 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();
List<T> currentResult = inputData;
final int maxIterations = iteration.getMaximumNumberOfIterations();
for (int superstep = 1; superstep <= maxIterations; superstep++) {
// set the input to the current partial solution
this.intermediateResults.put(iteration.getPartialSolution(), currentResult);
// set the superstep number
iterationSuperstep = superstep;
// grab the current iteration result
currentResult = (List<T>) execute(iteration.getNextPartialSolution(), superstep, jobID);
// evaluate the termination criterion
if (iteration.getTerminationCriterion() != null) {
execute(iteration.getTerminationCriterion(), superstep, jobID);
}
// 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();
return currentResult;
}
use of org.apache.flink.api.common.aggregators.Aggregator in project flink by apache.
the class EventWithAggregatorsTest method testSerializationOfEventWithAggregateValues.
@Test
public void testSerializationOfEventWithAggregateValues() {
StringValue stringValue = new StringValue("test string");
LongValue longValue = new LongValue(68743254);
String stringValueName = "stringValue";
String longValueName = "longValue";
Aggregator<StringValue> stringAgg = new TestAggregator<StringValue>(stringValue);
Aggregator<LongValue> longAgg = new TestAggregator<LongValue>(longValue);
Map<String, Aggregator<?>> aggMap = new HashMap<String, Aggregator<?>>();
aggMap.put(stringValueName, stringAgg);
aggMap.put(longValueName, longAgg);
Set<String> allNames = new HashSet<String>();
allNames.add(stringValueName);
allNames.add(longValueName);
Set<Value> allVals = new HashSet<Value>();
allVals.add(stringValue);
allVals.add(longValue);
// run the serialization
AllWorkersDoneEvent e = new AllWorkersDoneEvent(aggMap);
IterationEventWithAggregators deserialized = pipeThroughSerialization(e);
// verify the result
String[] names = deserialized.getAggregatorNames();
Value[] aggregates = deserialized.getAggregates(cl);
Assert.assertEquals(allNames.size(), names.length);
Assert.assertEquals(allVals.size(), aggregates.length);
// check that all the correct names and values are returned
for (String s : names) {
allNames.remove(s);
}
for (Value v : aggregates) {
allVals.remove(v);
}
Assert.assertTrue(allNames.isEmpty());
Assert.assertTrue(allVals.isEmpty());
}
Aggregations