Search in sources :

Example 41 with Channel

use of org.apache.flink.optimizer.plan.Channel in project flink by apache.

the class GenericFlatTypePostPass method traverse.

@SuppressWarnings("unchecked")
protected void traverse(PlanNode node, T parentSchema, boolean createUtilities) {
    // distinguish the node types
    if (node instanceof SinkPlanNode) {
        SinkPlanNode sn = (SinkPlanNode) node;
        Channel inchannel = sn.getInput();
        T schema = createEmptySchema();
        sn.postPassHelper = schema;
        // add the sinks information to the schema
        try {
            getSinkSchema(sn, schema);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Conflicting type infomation for the data sink '" + sn.getSinkNode().getOperator().getName() + "'.");
        }
        // descend to the input channel
        try {
            propagateToChannel(schema, inchannel, createUtilities);
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Missing type infomation for the channel that inputs to the data sink '" + sn.getSinkNode().getOperator().getName() + "'.");
        }
    } else if (node instanceof SourcePlanNode) {
        if (createUtilities) {
            ((SourcePlanNode) node).setSerializer(createSerializer(parentSchema, node));
        // nothing else to be done here. the source has no input and no strategy itself
        }
    } else if (node instanceof BulkIterationPlanNode) {
        BulkIterationPlanNode iterationNode = (BulkIterationPlanNode) node;
        // get the nodes current schema
        T schema;
        if (iterationNode.postPassHelper == null) {
            schema = createEmptySchema();
            iterationNode.postPassHelper = schema;
        } else {
            schema = (T) iterationNode.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, iterationNode.getProgramOperator().getName());
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < iterationNode.getOutgoingChannels().size()) {
            return;
        }
        if (iterationNode.getRootOfStepFunction() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile an iteration step function where next partial solution is created by a Union node.");
        }
        // traverse the termination criterion for the first time. create schema only, no utilities. Needed in case of intermediate termination criterion
        if (iterationNode.getRootOfTerminationCriterion() != null) {
            SingleInputPlanNode addMapper = (SingleInputPlanNode) iterationNode.getRootOfTerminationCriterion();
            traverse(addMapper.getInput().getSource(), createEmptySchema(), false);
            try {
                addMapper.getInput().setSerializer(createSerializer(createEmptySchema()));
            } catch (MissingFieldTypeInfoException e) {
                throw new RuntimeException(e);
            }
        }
        // traverse the step function for the first time. create schema only, no utilities
        traverse(iterationNode.getRootOfStepFunction(), schema, false);
        T pss = (T) iterationNode.getPartialSolutionPlanNode().postPassHelper;
        if (pss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Partial solution schema is null after first traversal of the step function.");
        }
        // traverse the step function for the second time, taking the schema of the partial solution
        traverse(iterationNode.getRootOfStepFunction(), pss, createUtilities);
        if (iterationNode.getRootOfTerminationCriterion() != null) {
            SingleInputPlanNode addMapper = (SingleInputPlanNode) iterationNode.getRootOfTerminationCriterion();
            traverse(addMapper.getInput().getSource(), createEmptySchema(), createUtilities);
            try {
                addMapper.getInput().setSerializer(createSerializer(createEmptySchema()));
            } catch (MissingFieldTypeInfoException e) {
                throw new RuntimeException(e);
            }
        }
        // take the schema from the partial solution node and add its fields to the iteration result schema.
        // input and output schema need to be identical, so this is essentially a sanity check
        addSchemaToSchema(pss, schema, iterationNode.getProgramOperator().getName());
        // set the serializer
        if (createUtilities) {
            iterationNode.setSerializerForIterationChannel(createSerializer(pss, iterationNode.getPartialSolutionPlanNode()));
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, iterationNode.getInput(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + iterationNode.getProgramOperator().getName() + "'. Missing type information for key field " + e.getFieldNumber());
        }
    } else if (node instanceof WorksetIterationPlanNode) {
        WorksetIterationPlanNode iterationNode = (WorksetIterationPlanNode) node;
        // get the nodes current schema
        T schema;
        if (iterationNode.postPassHelper == null) {
            schema = createEmptySchema();
            iterationNode.postPassHelper = schema;
        } else {
            schema = (T) iterationNode.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema (which refers to the solution set schema)
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, iterationNode.getProgramOperator().getName());
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < iterationNode.getOutgoingChannels().size()) {
            return;
        }
        if (iterationNode.getNextWorkSetPlanNode() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile a workset iteration step function where the next workset is produced by a Union node.");
        }
        if (iterationNode.getSolutionSetDeltaPlanNode() instanceof NAryUnionPlanNode) {
            throw new CompilerException("Optimizer cannot compile a workset iteration step function where the solution set delta is produced by a Union node.");
        }
        // traverse the step function
        // pass an empty schema to the next workset and the parent schema to the solution set delta
        // these first traversals are schema only
        traverse(iterationNode.getNextWorkSetPlanNode(), createEmptySchema(), false);
        traverse(iterationNode.getSolutionSetDeltaPlanNode(), schema, false);
        T wss = (T) iterationNode.getWorksetPlanNode().postPassHelper;
        T sss = (T) iterationNode.getSolutionSetPlanNode().postPassHelper;
        if (wss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Workset schema is null after first traversal of the step function.");
        }
        if (sss == null) {
            throw new CompilerException("Error in Optimizer Post Pass: Solution set schema is null after first traversal of the step function.");
        }
        // make the second pass and instantiate the utilities
        traverse(iterationNode.getNextWorkSetPlanNode(), wss, createUtilities);
        traverse(iterationNode.getSolutionSetDeltaPlanNode(), sss, createUtilities);
        // the solution set input and the result must have the same schema, this acts as a sanity check.
        try {
            for (Map.Entry<Integer, X> entry : sss) {
                Integer pos = entry.getKey();
                schema.addType(pos, entry.getValue());
            }
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Conflicting type information for field " + e.getFieldNumber() + " in node '" + iterationNode.getProgramOperator().getName() + "'. Contradicting types between the " + "result of the iteration and the solution set schema: " + e.getPreviousType() + " and " + e.getNewType() + ". Most probable cause: Invalid constant field annotations.");
        }
        // set the serializers and comparators
        if (createUtilities) {
            WorksetIterationNode optNode = iterationNode.getIterationNode();
            iterationNode.setWorksetSerializer(createSerializer(wss, iterationNode.getWorksetPlanNode()));
            iterationNode.setSolutionSetSerializer(createSerializer(sss, iterationNode.getSolutionSetPlanNode()));
            try {
                iterationNode.setSolutionSetComparator(createComparator(optNode.getSolutionSetKeyFields(), null, sss));
            } catch (MissingFieldTypeInfoException ex) {
                throw new CompilerPostPassException("Could not set up the solution set for workset iteration '" + optNode.getOperator().getName() + "'. Missing type information for key field " + ex.getFieldNumber() + '.');
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, iterationNode.getInitialSolutionSetInput(), createUtilities);
            propagateToChannel(wss, iterationNode.getInitialWorksetInput(), createUtilities);
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + iterationNode.getProgramOperator().getName() + "'. Missing type information for key field " + ex.getFieldNumber());
        }
    } else if (node instanceof SingleInputPlanNode) {
        SingleInputPlanNode sn = (SingleInputPlanNode) node;
        // get the nodes current schema
        T schema;
        if (sn.postPassHelper == null) {
            schema = createEmptySchema();
            sn.postPassHelper = schema;
        } else {
            schema = (T) sn.postPassHelper;
        }
        schema.increaseNumConnectionsThatContributed();
        SingleInputNode optNode = sn.getSingleInputNode();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema, optNode, 0);
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema.getNumConnectionsThatContributed() < sn.getOutgoingChannels().size()) {
            return;
        }
        // add the nodes local information
        try {
            getSingleInputNodeSchema(sn, schema);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException(getConflictingTypeErrorMessage(e, optNode.getOperator().getName()));
        }
        if (createUtilities) {
            // parameterize the node's driver strategy
            for (int i = 0; i < sn.getDriverStrategy().getNumRequiredComparators(); i++) {
                try {
                    sn.setComparator(createComparator(sn.getKeys(i), sn.getSortOrders(i), schema), i);
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for key field " + e.getFieldNumber());
                }
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema, sn.getInput(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        // don't forget the broadcast inputs
        for (Channel c : sn.getBroadcastInputs()) {
            try {
                propagateToChannel(createEmptySchema(), c, createUtilities);
            } catch (MissingFieldTypeInfoException e) {
                throw new CompilerPostPassException("Could not set up runtime strategy for broadcast channel in node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
            }
        }
    } else if (node instanceof DualInputPlanNode) {
        DualInputPlanNode dn = (DualInputPlanNode) node;
        // get the nodes current schema
        T schema1;
        T schema2;
        if (dn.postPassHelper1 == null) {
            schema1 = createEmptySchema();
            schema2 = createEmptySchema();
            dn.postPassHelper1 = schema1;
            dn.postPassHelper2 = schema2;
        } else {
            schema1 = (T) dn.postPassHelper1;
            schema2 = (T) dn.postPassHelper2;
        }
        schema1.increaseNumConnectionsThatContributed();
        schema2.increaseNumConnectionsThatContributed();
        TwoInputNode optNode = dn.getTwoInputNode();
        // add the parent schema to the schema
        if (propagateParentSchemaDown) {
            addSchemaToSchema(parentSchema, schema1, optNode, 0);
            addSchemaToSchema(parentSchema, schema2, optNode, 1);
        }
        // check whether all outgoing channels have not yet contributed. come back later if not.
        if (schema1.getNumConnectionsThatContributed() < dn.getOutgoingChannels().size()) {
            return;
        }
        // add the nodes local information
        try {
            getDualInputNodeSchema(dn, schema1, schema2);
        } catch (ConflictingFieldTypeInfoException e) {
            throw new CompilerPostPassException(getConflictingTypeErrorMessage(e, optNode.getOperator().getName()));
        }
        // parameterize the node's driver strategy
        if (createUtilities) {
            if (dn.getDriverStrategy().getNumRequiredComparators() > 0) {
                // set the individual comparators
                try {
                    dn.setComparator1(createComparator(dn.getKeysForInput1(), dn.getSortOrders(), schema1));
                    dn.setComparator2(createComparator(dn.getKeysForInput2(), dn.getSortOrders(), schema2));
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
                }
                // set the pair comparator
                try {
                    dn.setPairComparator(createPairComparator(dn.getKeysForInput1(), dn.getKeysForInput2(), dn.getSortOrders(), schema1, schema2));
                } catch (MissingFieldTypeInfoException e) {
                    throw new CompilerPostPassException("Could not set up runtime strategy for node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
                }
            }
        }
        // done, we can now propagate our info down
        try {
            propagateToChannel(schema1, dn.getInput1(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the first input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        try {
            propagateToChannel(schema2, dn.getInput2(), createUtilities);
        } catch (MissingFieldTypeInfoException e) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the second input channel to node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
        }
        // don't forget the broadcast inputs
        for (Channel c : dn.getBroadcastInputs()) {
            try {
                propagateToChannel(createEmptySchema(), c, createUtilities);
            } catch (MissingFieldTypeInfoException e) {
                throw new CompilerPostPassException("Could not set up runtime strategy for broadcast channel in node '" + optNode.getOperator().getName() + "'. Missing type information for field " + e.getFieldNumber());
            }
        }
    } else if (node instanceof NAryUnionPlanNode) {
        // only propagate the info down
        try {
            for (Channel channel : node.getInputs()) {
                propagateToChannel(parentSchema, channel, createUtilities);
            }
        } catch (MissingFieldTypeInfoException ex) {
            throw new CompilerPostPassException("Could not set up runtime strategy for the input channel to " + " a union node. Missing type information for field " + ex.getFieldNumber());
        }
    } else // catch the sources of the iterative step functions
    if (node instanceof BulkPartialSolutionPlanNode || node instanceof SolutionSetPlanNode || node instanceof WorksetPlanNode) {
        // get the nodes current schema
        T schema;
        String name;
        if (node instanceof BulkPartialSolutionPlanNode) {
            BulkPartialSolutionPlanNode psn = (BulkPartialSolutionPlanNode) node;
            if (psn.postPassHelper == null) {
                schema = createEmptySchema();
                psn.postPassHelper = schema;
            } else {
                schema = (T) psn.postPassHelper;
            }
            name = "partial solution of bulk iteration '" + psn.getPartialSolutionNode().getIterationNode().getOperator().getName() + "'";
        } else if (node instanceof SolutionSetPlanNode) {
            SolutionSetPlanNode ssn = (SolutionSetPlanNode) node;
            if (ssn.postPassHelper == null) {
                schema = createEmptySchema();
                ssn.postPassHelper = schema;
            } else {
                schema = (T) ssn.postPassHelper;
            }
            name = "solution set of workset iteration '" + ssn.getSolutionSetNode().getIterationNode().getOperator().getName() + "'";
        } else if (node instanceof WorksetPlanNode) {
            WorksetPlanNode wsn = (WorksetPlanNode) node;
            if (wsn.postPassHelper == null) {
                schema = createEmptySchema();
                wsn.postPassHelper = schema;
            } else {
                schema = (T) wsn.postPassHelper;
            }
            name = "workset of workset iteration '" + wsn.getWorksetNode().getIterationNode().getOperator().getName() + "'";
        } else {
            throw new CompilerException();
        }
        schema.increaseNumConnectionsThatContributed();
        // add the parent schema to the schema
        addSchemaToSchema(parentSchema, schema, name);
    } else {
        throw new CompilerPostPassException("Unknown node type encountered: " + node.getClass().getName());
    }
}
Also used : SingleInputNode(org.apache.flink.optimizer.dag.SingleInputNode) SolutionSetPlanNode(org.apache.flink.optimizer.plan.SolutionSetPlanNode) WorksetIterationPlanNode(org.apache.flink.optimizer.plan.WorksetIterationPlanNode) BulkPartialSolutionPlanNode(org.apache.flink.optimizer.plan.BulkPartialSolutionPlanNode) Channel(org.apache.flink.optimizer.plan.Channel) NAryUnionPlanNode(org.apache.flink.optimizer.plan.NAryUnionPlanNode) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) DualInputPlanNode(org.apache.flink.optimizer.plan.DualInputPlanNode) WorksetIterationNode(org.apache.flink.optimizer.dag.WorksetIterationNode) SinkPlanNode(org.apache.flink.optimizer.plan.SinkPlanNode) SourcePlanNode(org.apache.flink.optimizer.plan.SourcePlanNode) CompilerException(org.apache.flink.optimizer.CompilerException) WorksetPlanNode(org.apache.flink.optimizer.plan.WorksetPlanNode) CompilerPostPassException(org.apache.flink.optimizer.CompilerPostPassException) BulkIterationPlanNode(org.apache.flink.optimizer.plan.BulkIterationPlanNode) TwoInputNode(org.apache.flink.optimizer.dag.TwoInputNode)

Example 42 with Channel

use of org.apache.flink.optimizer.plan.Channel in project flink by apache.

the class SingleInputNode method addLocalCandidates.

protected void addLocalCandidates(Channel template, List<Set<? extends NamedChannel>> broadcastPlanChannels, RequestedGlobalProperties rgps, List<PlanNode> target, CostEstimator estimator) {
    for (RequestedLocalProperties ilp : this.inConn.getInterestingProperties().getLocalProperties()) {
        final Channel in = template.clone();
        ilp.parameterizeChannel(in);
        // instantiate a candidate, if the instantiated local properties meet one possible local property set
        outer: for (OperatorDescriptorSingle dps : getPossibleProperties()) {
            for (RequestedLocalProperties ilps : dps.getPossibleLocalProperties()) {
                if (ilps.isMetBy(in.getLocalProperties())) {
                    in.setRequiredLocalProps(ilps);
                    instantiateCandidate(dps, in, broadcastPlanChannels, target, estimator, rgps, ilp);
                    break outer;
                }
            }
        }
    }
}
Also used : RequestedLocalProperties(org.apache.flink.optimizer.dataproperties.RequestedLocalProperties) OperatorDescriptorSingle(org.apache.flink.optimizer.operators.OperatorDescriptorSingle) Channel(org.apache.flink.optimizer.plan.Channel) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel)

Example 43 with Channel

use of org.apache.flink.optimizer.plan.Channel in project flink by apache.

the class AllGroupWithPartialPreGroupProperties method instantiate.

@Override
public SingleInputPlanNode instantiate(Channel in, SingleInputNode node) {
    if (in.getShipStrategy() == ShipStrategyType.FORWARD) {
        // locally connected, directly instantiate
        return new SingleInputPlanNode(node, "GroupReduce (" + node.getOperator().getName() + ")", in, DriverStrategy.ALL_GROUP_REDUCE);
    } else {
        // non forward case.plug in a combiner
        Channel toCombiner = new Channel(in.getSource());
        toCombiner.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
        // create an input node for combine with same parallelism as input node
        GroupReduceNode combinerNode = ((GroupReduceNode) node).getCombinerUtilityNode();
        combinerNode.setParallelism(in.getSource().getParallelism());
        SingleInputPlanNode combiner = new SingleInputPlanNode(combinerNode, "Combine (" + node.getOperator().getName() + ")", toCombiner, DriverStrategy.ALL_GROUP_REDUCE_COMBINE);
        combiner.setCosts(new Costs(0, 0));
        combiner.initProperties(toCombiner.getGlobalProperties(), toCombiner.getLocalProperties());
        Channel toReducer = new Channel(combiner);
        toReducer.setShipStrategy(in.getShipStrategy(), in.getShipStrategyKeys(), in.getShipStrategySortOrder(), in.getDataExchangeMode());
        toReducer.setLocalStrategy(in.getLocalStrategy(), in.getLocalStrategyKeys(), in.getLocalStrategySortOrder());
        return new SingleInputPlanNode(node, "GroupReduce (" + node.getOperator().getName() + ")", toReducer, DriverStrategy.ALL_GROUP_REDUCE);
    }
}
Also used : SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) GroupReduceNode(org.apache.flink.optimizer.dag.GroupReduceNode) Costs(org.apache.flink.optimizer.costs.Costs) Channel(org.apache.flink.optimizer.plan.Channel)

Example 44 with Channel

use of org.apache.flink.optimizer.plan.Channel in project flink by apache.

the class ReduceProperties method instantiate.

@Override
public SingleInputPlanNode instantiate(Channel in, SingleInputNode node) {
    if (in.getShipStrategy() == ShipStrategyType.FORWARD || (node.getBroadcastConnections() != null && !node.getBroadcastConnections().isEmpty())) {
        if (in.getSource().getOptimizerNode() instanceof PartitionNode) {
            LOG.warn("Cannot automatically inject combiner for ReduceFunction. Please add an explicit combiner with combineGroup() in front of the partition operator.");
        }
        return new SingleInputPlanNode(node, "Reduce (" + node.getOperator().getName() + ")", in, DriverStrategy.SORTED_REDUCE, this.keyList);
    } else {
        // non forward case. all local properties are killed anyways, so we can safely plug in a combiner
        Channel toCombiner = new Channel(in.getSource());
        toCombiner.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
        // create an input node for combine with same parallelism as input node
        ReduceNode combinerNode = ((ReduceNode) node).getCombinerUtilityNode();
        combinerNode.setParallelism(in.getSource().getParallelism());
        SingleInputPlanNode combiner = new SingleInputPlanNode(combinerNode, "Combine (" + node.getOperator().getName() + ")", toCombiner, this.combinerStrategy, this.keyList);
        combiner.setCosts(new Costs(0, 0));
        combiner.initProperties(toCombiner.getGlobalProperties(), toCombiner.getLocalProperties());
        Channel toReducer = new Channel(combiner);
        toReducer.setShipStrategy(in.getShipStrategy(), in.getShipStrategyKeys(), in.getShipStrategySortOrder(), in.getDataExchangeMode());
        toReducer.setLocalStrategy(LocalStrategy.SORT, in.getLocalStrategyKeys(), in.getLocalStrategySortOrder());
        return new SingleInputPlanNode(node, "Reduce(" + node.getOperator().getName() + ")", toReducer, DriverStrategy.SORTED_REDUCE, this.keyList);
    }
}
Also used : SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) ReduceNode(org.apache.flink.optimizer.dag.ReduceNode) Costs(org.apache.flink.optimizer.costs.Costs) Channel(org.apache.flink.optimizer.plan.Channel) PartitionNode(org.apache.flink.optimizer.dag.PartitionNode)

Example 45 with Channel

use of org.apache.flink.optimizer.plan.Channel in project flink by apache.

the class GroupReduceWithCombineProperties method instantiate.

@Override
public SingleInputPlanNode instantiate(Channel in, SingleInputNode node) {
    if (in.getShipStrategy() == ShipStrategyType.FORWARD) {
        if (in.getSource().getOptimizerNode() instanceof PartitionNode) {
            LOG.warn("Cannot automatically inject combiner for GroupReduceFunction. Please add an explicit combiner with combineGroup() in front of the partition operator.");
        }
        // adjust a sort (changes grouping, so it must be for this driver to combining sort
        if (in.getLocalStrategy() == LocalStrategy.SORT) {
            if (!in.getLocalStrategyKeys().isValidUnorderedPrefix(this.keys)) {
                throw new RuntimeException("Bug: Inconsistent sort for group strategy.");
            }
            in.setLocalStrategy(LocalStrategy.COMBININGSORT, in.getLocalStrategyKeys(), in.getLocalStrategySortOrder());
        }
        return new SingleInputPlanNode(node, "Reduce(" + node.getOperator().getName() + ")", in, DriverStrategy.SORTED_GROUP_REDUCE, this.keyList);
    } else {
        // non forward case. all local properties are killed anyways, so we can safely plug in a combiner
        Channel toCombiner = new Channel(in.getSource());
        toCombiner.setShipStrategy(ShipStrategyType.FORWARD, DataExchangeMode.PIPELINED);
        // create an input node for combine with same parallelism as input node
        GroupReduceNode combinerNode = ((GroupReduceNode) node).getCombinerUtilityNode();
        combinerNode.setParallelism(in.getSource().getParallelism());
        SingleInputPlanNode combiner = new SingleInputPlanNode(combinerNode, "Combine(" + node.getOperator().getName() + ")", toCombiner, DriverStrategy.SORTED_GROUP_COMBINE);
        combiner.setCosts(new Costs(0, 0));
        combiner.initProperties(toCombiner.getGlobalProperties(), toCombiner.getLocalProperties());
        // set sorting comparator key info
        combiner.setDriverKeyInfo(in.getLocalStrategyKeys(), in.getLocalStrategySortOrder(), 0);
        // set grouping comparator key info
        combiner.setDriverKeyInfo(this.keyList, 1);
        Channel toReducer = new Channel(combiner);
        toReducer.setShipStrategy(in.getShipStrategy(), in.getShipStrategyKeys(), in.getShipStrategySortOrder(), in.getDataExchangeMode());
        if (in.getShipStrategy() == ShipStrategyType.PARTITION_RANGE) {
            toReducer.setDataDistribution(in.getDataDistribution());
        }
        toReducer.setLocalStrategy(LocalStrategy.COMBININGSORT, in.getLocalStrategyKeys(), in.getLocalStrategySortOrder());
        return new SingleInputPlanNode(node, "Reduce (" + node.getOperator().getName() + ")", toReducer, DriverStrategy.SORTED_GROUP_REDUCE, this.keyList);
    }
}
Also used : SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) GroupReduceNode(org.apache.flink.optimizer.dag.GroupReduceNode) Costs(org.apache.flink.optimizer.costs.Costs) Channel(org.apache.flink.optimizer.plan.Channel) PartitionNode(org.apache.flink.optimizer.dag.PartitionNode)

Aggregations

Channel (org.apache.flink.optimizer.plan.Channel)60 SingleInputPlanNode (org.apache.flink.optimizer.plan.SingleInputPlanNode)41 Test (org.junit.Test)30 OptimizedPlan (org.apache.flink.optimizer.plan.OptimizedPlan)26 ExecutionEnvironment (org.apache.flink.api.java.ExecutionEnvironment)24 PlanNode (org.apache.flink.optimizer.plan.PlanNode)24 DualInputPlanNode (org.apache.flink.optimizer.plan.DualInputPlanNode)23 SourcePlanNode (org.apache.flink.optimizer.plan.SourcePlanNode)23 SinkPlanNode (org.apache.flink.optimizer.plan.SinkPlanNode)20 NAryUnionPlanNode (org.apache.flink.optimizer.plan.NAryUnionPlanNode)19 Plan (org.apache.flink.api.common.Plan)18 BulkIterationPlanNode (org.apache.flink.optimizer.plan.BulkIterationPlanNode)18 Tuple2 (org.apache.flink.api.java.tuple.Tuple2)16 NamedChannel (org.apache.flink.optimizer.plan.NamedChannel)16 WorksetIterationPlanNode (org.apache.flink.optimizer.plan.WorksetIterationPlanNode)15 CompilerException (org.apache.flink.optimizer.CompilerException)14 GlobalProperties (org.apache.flink.optimizer.dataproperties.GlobalProperties)13 RequestedGlobalProperties (org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties)13 RequestedLocalProperties (org.apache.flink.optimizer.dataproperties.RequestedLocalProperties)13 FieldList (org.apache.flink.api.common.operators.util.FieldList)12