use of org.apache.flink.api.java.functions.SampleInCoordinator in project flink by apache.
the class DataSetUtils method sampleWithSize.
/**
* Generate a sample of DataSet which contains fixed size elements.
*
* <p><strong>NOTE:</strong> Sample with fixed size is not as efficient as sample with fraction,
* use sample with fraction unless you need exact precision.
*
* @param withReplacement Whether element can be selected more than once.
* @param numSamples The expected sample size.
* @param seed Random number generator seed.
* @return The sampled DataSet
*/
public static <T> DataSet<T> sampleWithSize(DataSet<T> input, final boolean withReplacement, final int numSamples, final long seed) {
SampleInPartition<T> sampleInPartition = new SampleInPartition<>(withReplacement, numSamples, seed);
MapPartitionOperator mapPartitionOperator = input.mapPartition(sampleInPartition);
// There is no previous group, so the parallelism of GroupReduceOperator is always 1.
String callLocation = Utils.getCallLocationName();
SampleInCoordinator<T> sampleInCoordinator = new SampleInCoordinator<>(withReplacement, numSamples, seed);
return new GroupReduceOperator<>(mapPartitionOperator, input.getType(), sampleInCoordinator, callLocation);
}
use of org.apache.flink.api.java.functions.SampleInCoordinator in project flink by apache.
the class RangePartitionRewriter method rewriteRangePartitionChannel.
private List<Channel> rewriteRangePartitionChannel(Channel channel) {
final List<Channel> sourceNewOutputChannels = new ArrayList<>();
final PlanNode sourceNode = channel.getSource();
final PlanNode targetNode = channel.getTarget();
final int sourceParallelism = sourceNode.getParallelism();
final int targetParallelism = targetNode.getParallelism();
final Costs defaultZeroCosts = new Costs(0, 0, 0);
final TypeComparatorFactory<?> comparator = Utils.getShipComparator(channel, this.plan.getOriginalPlan().getExecutionConfig());
// 1. Fixed size sample in each partitions.
final int sampleSize = SAMPLES_PER_PARTITION * targetParallelism;
final SampleInPartition sampleInPartition = new SampleInPartition(false, sampleSize, SEED);
final TypeInformation<?> sourceOutputType = sourceNode.getOptimizerNode().getOperator().getOperatorInfo().getOutputType();
final TypeInformation<IntermediateSampleData> isdTypeInformation = TypeExtractor.getForClass(IntermediateSampleData.class);
final UnaryOperatorInformation sipOperatorInformation = new UnaryOperatorInformation(sourceOutputType, isdTypeInformation);
final MapPartitionOperatorBase sipOperatorBase = new MapPartitionOperatorBase(sampleInPartition, sipOperatorInformation, SIP_NAME);
final MapPartitionNode sipNode = new MapPartitionNode(sipOperatorBase);
final Channel sipChannel = new Channel(sourceNode, TempMode.NONE);
sipChannel.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
final SingleInputPlanNode sipPlanNode = new SingleInputPlanNode(sipNode, SIP_NAME, sipChannel, DriverStrategy.MAP_PARTITION);
sipNode.setParallelism(sourceParallelism);
sipPlanNode.setParallelism(sourceParallelism);
sipPlanNode.initProperties(new GlobalProperties(), new LocalProperties());
sipPlanNode.setCosts(defaultZeroCosts);
sipChannel.setTarget(sipPlanNode);
this.plan.getAllNodes().add(sipPlanNode);
sourceNewOutputChannels.add(sipChannel);
// 2. Fixed size sample in a single coordinator.
final SampleInCoordinator sampleInCoordinator = new SampleInCoordinator(false, sampleSize, SEED);
final UnaryOperatorInformation sicOperatorInformation = new UnaryOperatorInformation(isdTypeInformation, sourceOutputType);
final GroupReduceOperatorBase sicOperatorBase = new GroupReduceOperatorBase(sampleInCoordinator, sicOperatorInformation, SIC_NAME);
final GroupReduceNode sicNode = new GroupReduceNode(sicOperatorBase);
final Channel sicChannel = new Channel(sipPlanNode, TempMode.NONE);
sicChannel.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
final SingleInputPlanNode sicPlanNode = new SingleInputPlanNode(sicNode, SIC_NAME, sicChannel, DriverStrategy.ALL_GROUP_REDUCE);
sicNode.setParallelism(1);
sicPlanNode.setParallelism(1);
sicPlanNode.initProperties(new GlobalProperties(), new LocalProperties());
sicPlanNode.setCosts(defaultZeroCosts);
sicChannel.setTarget(sicPlanNode);
sipPlanNode.addOutgoingChannel(sicChannel);
this.plan.getAllNodes().add(sicPlanNode);
// 3. Use sampled data to build range boundaries.
final RangeBoundaryBuilder rangeBoundaryBuilder = new RangeBoundaryBuilder(comparator, targetParallelism);
final TypeInformation<CommonRangeBoundaries> rbTypeInformation = TypeExtractor.getForClass(CommonRangeBoundaries.class);
final UnaryOperatorInformation rbOperatorInformation = new UnaryOperatorInformation(sourceOutputType, rbTypeInformation);
final MapPartitionOperatorBase rbOperatorBase = new MapPartitionOperatorBase(rangeBoundaryBuilder, rbOperatorInformation, RB_NAME);
final MapPartitionNode rbNode = new MapPartitionNode(rbOperatorBase);
final Channel rbChannel = new Channel(sicPlanNode, TempMode.NONE);
rbChannel.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
final SingleInputPlanNode rbPlanNode = new SingleInputPlanNode(rbNode, RB_NAME, rbChannel, DriverStrategy.MAP_PARTITION);
rbNode.setParallelism(1);
rbPlanNode.setParallelism(1);
rbPlanNode.initProperties(new GlobalProperties(), new LocalProperties());
rbPlanNode.setCosts(defaultZeroCosts);
rbChannel.setTarget(rbPlanNode);
sicPlanNode.addOutgoingChannel(rbChannel);
this.plan.getAllNodes().add(rbPlanNode);
// 4. Take range boundaries as broadcast input and take the tuple of partition id and record
// as output.
final AssignRangeIndex assignRangeIndex = new AssignRangeIndex(comparator);
final TypeInformation<Tuple2> ariOutputTypeInformation = new TupleTypeInfo<>(BasicTypeInfo.INT_TYPE_INFO, sourceOutputType);
final UnaryOperatorInformation ariOperatorInformation = new UnaryOperatorInformation(sourceOutputType, ariOutputTypeInformation);
final MapPartitionOperatorBase ariOperatorBase = new MapPartitionOperatorBase(assignRangeIndex, ariOperatorInformation, ARI_NAME);
final MapPartitionNode ariNode = new MapPartitionNode(ariOperatorBase);
final Channel ariChannel = new Channel(sourceNode, TempMode.NONE);
// To avoid deadlock, set the DataExchangeMode of channel between source node and this to
// Batch.
ariChannel.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.BATCH);
final SingleInputPlanNode ariPlanNode = new SingleInputPlanNode(ariNode, ARI_NAME, ariChannel, DriverStrategy.MAP_PARTITION);
ariNode.setParallelism(sourceParallelism);
ariPlanNode.setParallelism(sourceParallelism);
ariPlanNode.initProperties(new GlobalProperties(), new LocalProperties());
ariPlanNode.setCosts(defaultZeroCosts);
ariChannel.setTarget(ariPlanNode);
this.plan.getAllNodes().add(ariPlanNode);
sourceNewOutputChannels.add(ariChannel);
final NamedChannel broadcastChannel = new NamedChannel("RangeBoundaries", rbPlanNode);
broadcastChannel.setShipStrategy(ShipStrategyType.BROADCAST, DataExchangeMode.PIPELINED);
broadcastChannel.setTarget(ariPlanNode);
List<NamedChannel> broadcastChannels = new ArrayList<>(1);
broadcastChannels.add(broadcastChannel);
ariPlanNode.setBroadcastInputs(broadcastChannels);
// 5. Remove the partition id.
final Channel partChannel = new Channel(ariPlanNode, TempMode.NONE);
final FieldList keys = new FieldList(0);
partChannel.setShipStrategy(ShipStrategyType.PARTITION_CUSTOM, keys, idPartitioner, DataExchangeMode.PIPELINED);
ariPlanNode.addOutgoingChannel(partChannel);
final RemoveRangeIndex partitionIDRemoveWrapper = new RemoveRangeIndex();
final UnaryOperatorInformation prOperatorInformation = new UnaryOperatorInformation(ariOutputTypeInformation, sourceOutputType);
final MapOperatorBase prOperatorBase = new MapOperatorBase(partitionIDRemoveWrapper, prOperatorInformation, PR_NAME);
final MapNode prRemoverNode = new MapNode(prOperatorBase);
final SingleInputPlanNode prPlanNode = new SingleInputPlanNode(prRemoverNode, PR_NAME, partChannel, DriverStrategy.MAP);
partChannel.setTarget(prPlanNode);
prRemoverNode.setParallelism(targetParallelism);
prPlanNode.setParallelism(targetParallelism);
GlobalProperties globalProperties = new GlobalProperties();
globalProperties.setRangePartitioned(new Ordering(0, null, Order.ASCENDING));
prPlanNode.initProperties(globalProperties, new LocalProperties());
prPlanNode.setCosts(defaultZeroCosts);
this.plan.getAllNodes().add(prPlanNode);
// 6. Connect to target node.
channel.setSource(prPlanNode);
channel.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
prPlanNode.addOutgoingChannel(channel);
return sourceNewOutputChannels;
}
Aggregations