Github user fhueske commented on a diff in the pull request:
https://github.com/apache/flink/pull/1255#discussion_r43956068
--- Diff:
flink-optimizer/src/main/java/org/apache/flink/optimizer/traversals/RangePartitionRewriter.java
---
@@ -0,0 +1,188 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.flink.optimizer.traversals;
+
+import org.apache.flink.api.common.distributions.CommonRangeBoundaries;
+import org.apache.flink.api.common.operators.UnaryOperatorInformation;
+import org.apache.flink.api.common.operators.base.GroupReduceOperatorBase;
+import org.apache.flink.api.common.operators.base.MapOperatorBase;
+import org.apache.flink.api.common.operators.base.MapPartitionOperatorBase;
+import org.apache.flink.api.common.typeinfo.BasicTypeInfo;
+import org.apache.flink.api.common.typeinfo.TypeInformation;
+import org.apache.flink.api.common.typeutils.TypeComparatorFactory;
+import org.apache.flink.api.java.functions.AssignRangeIndex;
+import org.apache.flink.api.java.functions.PartitionIDRemoveWrapper;
+import org.apache.flink.api.java.functions.RangeBoundaryBuilder;
+import org.apache.flink.api.java.functions.SampleInCoordinator;
+import org.apache.flink.api.java.functions.SampleInPartition;
+import org.apache.flink.api.java.sampling.IntermediateSampleData;
+import org.apache.flink.api.java.tuple.Tuple2;
+import org.apache.flink.api.java.typeutils.TupleTypeInfo;
+import org.apache.flink.api.java.typeutils.TypeExtractor;
+import org.apache.flink.optimizer.dag.GroupReduceNode;
+import org.apache.flink.optimizer.dag.MapNode;
+import org.apache.flink.optimizer.dag.MapPartitionNode;
+import org.apache.flink.optimizer.dag.TempMode;
+import org.apache.flink.optimizer.plan.Channel;
+import org.apache.flink.optimizer.plan.NamedChannel;
+import org.apache.flink.optimizer.plan.OptimizedPlan;
+import org.apache.flink.optimizer.plan.PlanNode;
+import org.apache.flink.optimizer.plan.SingleInputPlanNode;
+import org.apache.flink.optimizer.util.Utils;
+import org.apache.flink.runtime.operators.DriverStrategy;
+import org.apache.flink.runtime.operators.shipping.ShipStrategyType;
+import org.apache.flink.util.Visitor;
+
+import java.util.ArrayList;
+import java.util.LinkedList;
+import java.util.List;
+
+public class RangePartitionRewriter implements Visitor<PlanNode> {
+
+ OptimizedPlan plan;
+
+ public RangePartitionRewriter(OptimizedPlan plan) {
+ this.plan = plan;
+ }
+
+ @Override
+ public boolean preVisit(PlanNode visitable) {
+ return true;
+ }
+
+ @Override
+ public void postVisit(PlanNode visitable) {
+ List<Channel> outgoingChannels =
visitable.getOutgoingChannels();
+ List<Channel> appendOutgoingChannels = new LinkedList<>();
+ List<Channel> removeOutgoingChannels = new LinkedList<>();
+ for (Channel channel : outgoingChannels) {
+ ShipStrategyType shipStrategy =
channel.getShipStrategy();
+ if (shipStrategy == ShipStrategyType.PARTITION_RANGE) {
+ if (channel.getDataDistribution() == null) {
+ removeOutgoingChannels.add(channel);
+
appendOutgoingChannels.addAll(rewriteRangePartitionChannel(channel));
+ }
+ }
+ }
+ outgoingChannels.addAll(appendOutgoingChannels);
+ for (Channel channel : removeOutgoingChannels) {
+ outgoingChannels.remove(channel);
+ }
+ }
+
+ private List<Channel> rewriteRangePartitionChannel(Channel channel) {
+ List<Channel> appendOutgoingChannels = new LinkedList<>();
+ PlanNode sourceNode = channel.getSource();
+ PlanNode targetNode = channel.getTarget();
+ int sourceParallelism = sourceNode.getParallelism();
+ int targetParallelism = targetNode.getParallelism();
+ TypeComparatorFactory<?> comparator =
Utils.getShipComparator(channel,
this.plan.getOriginalPlan().getExecutionConfig());
+ // 1. Fixed size sample in each partitions.
+ long seed = org.apache.flink.api.java.Utils.RNG.nextLong();
+ int sampleSize = 20 * targetParallelism;
+ SampleInPartition sampleInPartition = new
SampleInPartition(false, sampleSize, seed);
+ TypeInformation<?> sourceOutputType =
sourceNode.getOptimizerNode().getOperator().getOperatorInfo().getOutputType();
+ TypeInformation<IntermediateSampleData> isdTypeInformation =
TypeExtractor.getForClass(IntermediateSampleData.class);
+ UnaryOperatorInformation sipOperatorInformation = new
UnaryOperatorInformation(sourceOutputType, isdTypeInformation);
+ MapPartitionOperatorBase sipOperatorBase = new
MapPartitionOperatorBase(sampleInPartition, sipOperatorInformation, "Sample in
partitions");
+ MapPartitionNode sipNode = new
MapPartitionNode(sipOperatorBase);
+ Channel sipChannel = new Channel(sourceNode, TempMode.NONE);
+ sipChannel.setShipStrategy(ShipStrategyType.FORWARD,
channel.getDataExchangeMode());
+ SingleInputPlanNode sipPlanNode = new
SingleInputPlanNode(sipNode, "SampleInPartition PlanNode", sipChannel,
DriverStrategy.MAP_PARTITION);
+ sipPlanNode.setParallelism(sourceParallelism);
+ sipChannel.setTarget(sipPlanNode);
+ appendOutgoingChannels.add(sipChannel);
+ this.plan.getAllNodes().add(sipPlanNode);
+
+ // 2. Fixed size sample in a single coordinator.
+ SampleInCoordinator sampleInCoordinator = new
SampleInCoordinator(false, sampleSize, seed);
+ UnaryOperatorInformation sicOperatorInformation = new
UnaryOperatorInformation(isdTypeInformation, sourceOutputType);
+ GroupReduceOperatorBase sicOperatorBase = new
GroupReduceOperatorBase(sampleInCoordinator, sicOperatorInformation, "Sample in
coordinator");
+ GroupReduceNode sicNode = new GroupReduceNode(sicOperatorBase);
+ Channel sicChannel = new Channel(sipPlanNode, TempMode.NONE);
+ sicChannel.setShipStrategy(ShipStrategyType.PARTITION_HASH,
channel.getShipStrategyKeys(), channel.getShipStrategySortOrder(), null,
channel.getDataExchangeMode());
+ SingleInputPlanNode sicPlanNode = new
SingleInputPlanNode(sicNode, "SampleInCoordinator PlanNode", sicChannel,
DriverStrategy.ALL_GROUP_REDUCE);
+ sicPlanNode.setParallelism(1);
+ sicChannel.setTarget(sicPlanNode);
+ sipPlanNode.addOutgoingChannel(sicChannel);
+ this.plan.getAllNodes().add(sicPlanNode);
+
+ // 3. Use sampled data to build range boundaries.
+ RangeBoundaryBuilder rangeBoundaryBuilder = new
RangeBoundaryBuilder(comparator, targetParallelism);
+ TypeInformation<CommonRangeBoundaries> rbTypeInformation =
TypeExtractor.getForClass(CommonRangeBoundaries.class);
+ UnaryOperatorInformation rbOperatorInformation = new
UnaryOperatorInformation(sourceOutputType, rbTypeInformation);
+ MapPartitionOperatorBase rbOperatorBase = new
MapPartitionOperatorBase(rangeBoundaryBuilder, rbOperatorInformation,
"RangeBoundaryBuilder");
+ MapPartitionNode rbNode= new MapPartitionNode(rbOperatorBase);
+ Channel rbChannel = new Channel(sicPlanNode, TempMode.NONE);
+ rbChannel.setShipStrategy(ShipStrategyType.FORWARD,
channel.getDataExchangeMode());
+ SingleInputPlanNode rbPlanNode = new
SingleInputPlanNode(rbNode, "RangeBoundary PlanNode", rbChannel,
DriverStrategy.MAP_PARTITION);
+ rbPlanNode.setParallelism(1);
+ 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.
+ AssignRangeIndex assignRangeIndex = new
AssignRangeIndex(comparator);
+ TypeInformation<Tuple2> ariOutputTypeInformation = new
TupleTypeInfo<>(BasicTypeInfo.INT_TYPE_INFO, sourceOutputType);
+ UnaryOperatorInformation ariOperatorInformation = new
UnaryOperatorInformation(sourceOutputType, ariOutputTypeInformation);
+ MapPartitionOperatorBase ariOperatorBase = new
MapPartitionOperatorBase(assignRangeIndex, ariOperatorInformation, "Assign
Range Index");
+ MapPartitionNode ariNode= new MapPartitionNode(ariOperatorBase);
+ Channel ariChannel = new Channel(sourceNode, TempMode.NONE);
--- End diff --
We need to be very careful at this point! We cannot add two channels to the
same node (`sourceNode`) that pipeline the data, where one of the successors
has to wait for the other to complete. In our case this is because of the
broadcast set which implies that the boundaries are first computed and
broadcasted before the idAssigner can start to process the data. Since we need
to see all data from `sourceNode` in order to build the boundaries but have to
wait for the boundaries to be able process the data from `sourceNode`, this
will lead to a deadlock in the data flow.
I have to admit, I am not sure how we can break the pipeline. Previously
that was achieved by the `TempMode` but recently the `DataExchangeMode` was
introduced and I am not quite sure write how both modes are different from each
other and how they interact. I will ask another committer to comment on this.
For now, I only want to raise a flag and get your attention on this issue.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
