Search in sources :

Example 1 with OperatorDescriptorDual

use of org.apache.flink.optimizer.operators.OperatorDescriptorDual in project flink by apache.

the class TwoInputNode method computeInterestingPropertiesForInputs.

@Override
public void computeInterestingPropertiesForInputs(CostEstimator estimator) {
    // get what we inherit and what is preserved by our user code
    final InterestingProperties props1 = getInterestingProperties().filterByCodeAnnotations(this, 0);
    final InterestingProperties props2 = getInterestingProperties().filterByCodeAnnotations(this, 1);
    // add all properties relevant to this node
    for (OperatorDescriptorDual dpd : getProperties()) {
        for (GlobalPropertiesPair gp : dpd.getPossibleGlobalProperties()) {
            // input 1
            props1.addGlobalProperties(gp.getProperties1());
            // input 2
            props2.addGlobalProperties(gp.getProperties2());
        }
        for (LocalPropertiesPair lp : dpd.getPossibleLocalProperties()) {
            // input 1
            props1.addLocalProperties(lp.getProperties1());
            // input 2
            props2.addLocalProperties(lp.getProperties2());
        }
    }
    this.input1.setInterestingProperties(props1);
    this.input2.setInterestingProperties(props2);
    for (DagConnection conn : getBroadcastConnections()) {
        conn.setInterestingProperties(new InterestingProperties());
    }
}
Also used : InterestingProperties(org.apache.flink.optimizer.dataproperties.InterestingProperties) GlobalPropertiesPair(org.apache.flink.optimizer.operators.OperatorDescriptorDual.GlobalPropertiesPair) OperatorDescriptorDual(org.apache.flink.optimizer.operators.OperatorDescriptorDual) LocalPropertiesPair(org.apache.flink.optimizer.operators.OperatorDescriptorDual.LocalPropertiesPair)

Example 2 with OperatorDescriptorDual

use of org.apache.flink.optimizer.operators.OperatorDescriptorDual in project flink by apache.

the class JoinNode method makeJoinWithSolutionSet.

public void makeJoinWithSolutionSet(int solutionsetInputIndex) {
    OperatorDescriptorDual op;
    if (solutionsetInputIndex == 0) {
        op = new HashJoinBuildFirstProperties(this.keys1, this.keys2);
    } else if (solutionsetInputIndex == 1) {
        op = new HashJoinBuildSecondProperties(this.keys1, this.keys2);
    } else {
        throw new IllegalArgumentException();
    }
    this.dataProperties = Collections.singletonList(op);
}
Also used : HashJoinBuildFirstProperties(org.apache.flink.optimizer.operators.HashJoinBuildFirstProperties) HashJoinBuildSecondProperties(org.apache.flink.optimizer.operators.HashJoinBuildSecondProperties) OperatorDescriptorDual(org.apache.flink.optimizer.operators.OperatorDescriptorDual)

Example 3 with OperatorDescriptorDual

use of org.apache.flink.optimizer.operators.OperatorDescriptorDual in project flink by apache.

the class OuterJoinNode method getDataProperties.

private List<OperatorDescriptorDual> getDataProperties() {
    OuterJoinOperatorBase<?, ?, ?, ?> operator = getOperator();
    OuterJoinType type = operator.getOuterJoinType();
    JoinHint joinHint = operator.getJoinHint();
    joinHint = joinHint == null ? JoinHint.OPTIMIZER_CHOOSES : joinHint;
    List<OperatorDescriptorDual> list;
    switch(type) {
        case LEFT:
            list = createLeftOuterJoinDescriptors(joinHint);
            break;
        case RIGHT:
            list = createRightOuterJoinDescriptors(joinHint);
            break;
        case FULL:
            list = createFullOuterJoinDescriptors(joinHint);
            break;
        default:
            throw new CompilerException("Unknown outer join type: " + type);
    }
    Partitioner<?> customPartitioner = operator.getCustomPartitioner();
    if (customPartitioner != null) {
        for (OperatorDescriptorDual desc : list) {
            ((AbstractJoinDescriptor) desc).setCustomPartitioner(customPartitioner);
        }
    }
    return list;
}
Also used : JoinHint(org.apache.flink.api.common.operators.base.JoinOperatorBase.JoinHint) CompilerException(org.apache.flink.optimizer.CompilerException) OperatorDescriptorDual(org.apache.flink.optimizer.operators.OperatorDescriptorDual) AbstractJoinDescriptor(org.apache.flink.optimizer.operators.AbstractJoinDescriptor) OuterJoinType(org.apache.flink.api.common.operators.base.OuterJoinOperatorBase.OuterJoinType)

Example 4 with OperatorDescriptorDual

use of org.apache.flink.optimizer.operators.OperatorDescriptorDual in project flink by apache.

the class TwoInputNode method getAlternativePlans.

@Override
public List<PlanNode> getAlternativePlans(CostEstimator estimator) {
    // check if we have a cached version
    if (this.cachedPlans != null) {
        return this.cachedPlans;
    }
    boolean childrenSkippedDueToReplicatedInput = false;
    // step down to all producer nodes and calculate alternative plans
    final List<? extends PlanNode> subPlans1 = getFirstPredecessorNode().getAlternativePlans(estimator);
    final List<? extends PlanNode> subPlans2 = getSecondPredecessorNode().getAlternativePlans(estimator);
    // calculate alternative sub-plans for predecessor
    final Set<RequestedGlobalProperties> intGlobal1 = this.input1.getInterestingProperties().getGlobalProperties();
    final Set<RequestedGlobalProperties> intGlobal2 = this.input2.getInterestingProperties().getGlobalProperties();
    // calculate alternative sub-plans for broadcast inputs
    final List<Set<? extends NamedChannel>> broadcastPlanChannels = new ArrayList<Set<? extends NamedChannel>>();
    List<DagConnection> broadcastConnections = getBroadcastConnections();
    List<String> broadcastConnectionNames = getBroadcastConnectionNames();
    for (int i = 0; i < broadcastConnections.size(); i++) {
        DagConnection broadcastConnection = broadcastConnections.get(i);
        String broadcastConnectionName = broadcastConnectionNames.get(i);
        List<PlanNode> broadcastPlanCandidates = broadcastConnection.getSource().getAlternativePlans(estimator);
        // wrap the plan candidates in named channels
        HashSet<NamedChannel> broadcastChannels = new HashSet<NamedChannel>(broadcastPlanCandidates.size());
        for (PlanNode plan : broadcastPlanCandidates) {
            final NamedChannel c = new NamedChannel(broadcastConnectionName, plan);
            DataExchangeMode exMode = DataExchangeMode.select(broadcastConnection.getDataExchangeMode(), ShipStrategyType.BROADCAST, broadcastConnection.isBreakingPipeline());
            c.setShipStrategy(ShipStrategyType.BROADCAST, exMode);
            broadcastChannels.add(c);
        }
        broadcastPlanChannels.add(broadcastChannels);
    }
    final GlobalPropertiesPair[] allGlobalPairs;
    final LocalPropertiesPair[] allLocalPairs;
    {
        Set<GlobalPropertiesPair> pairsGlob = new HashSet<GlobalPropertiesPair>();
        Set<LocalPropertiesPair> pairsLoc = new HashSet<LocalPropertiesPair>();
        for (OperatorDescriptorDual ods : getProperties()) {
            pairsGlob.addAll(ods.getPossibleGlobalProperties());
            pairsLoc.addAll(ods.getPossibleLocalProperties());
        }
        allGlobalPairs = pairsGlob.toArray(new GlobalPropertiesPair[pairsGlob.size()]);
        allLocalPairs = pairsLoc.toArray(new LocalPropertiesPair[pairsLoc.size()]);
    }
    final ArrayList<PlanNode> outputPlans = new ArrayList<PlanNode>();
    final ExecutionMode input1Mode = this.input1.getDataExchangeMode();
    final ExecutionMode input2Mode = this.input2.getDataExchangeMode();
    final int parallelism = getParallelism();
    final int inParallelism1 = getFirstPredecessorNode().getParallelism();
    final int inParallelism2 = getSecondPredecessorNode().getParallelism();
    final boolean dopChange1 = parallelism != inParallelism1;
    final boolean dopChange2 = parallelism != inParallelism2;
    final boolean input1breaksPipeline = this.input1.isBreakingPipeline();
    final boolean input2breaksPipeline = this.input2.isBreakingPipeline();
    // create all candidates
    for (PlanNode child1 : subPlans1) {
        if (child1.getGlobalProperties().isFullyReplicated()) {
            // fully replicated input is always locally forwarded if parallelism is not changed
            if (dopChange1) {
                // can not continue with this child
                childrenSkippedDueToReplicatedInput = true;
                continue;
            } else {
                this.input1.setShipStrategy(ShipStrategyType.FORWARD);
            }
        }
        for (PlanNode child2 : subPlans2) {
            if (child2.getGlobalProperties().isFullyReplicated()) {
                // changed
                if (dopChange2) {
                    // can not continue with this child
                    childrenSkippedDueToReplicatedInput = true;
                    continue;
                } else {
                    this.input2.setShipStrategy(ShipStrategyType.FORWARD);
                }
            }
            // candidate at the joined branch plan.
            if (!areBranchCompatible(child1, child2)) {
                continue;
            }
            for (RequestedGlobalProperties igps1 : intGlobal1) {
                // create a candidate channel for the first input. mark it cached, if the
                // connection says so
                final Channel c1 = new Channel(child1, this.input1.getMaterializationMode());
                if (this.input1.getShipStrategy() == null) {
                    // free to choose the ship strategy
                    igps1.parameterizeChannel(c1, dopChange1, input1Mode, input1breaksPipeline);
                    // ship strategy preserves/establishes them even under changing parallelisms
                    if (dopChange1 && !c1.getShipStrategy().isNetworkStrategy()) {
                        c1.getGlobalProperties().reset();
                    }
                } else {
                    // ship strategy fixed by compiler hint
                    ShipStrategyType shipType = this.input1.getShipStrategy();
                    DataExchangeMode exMode = DataExchangeMode.select(input1Mode, shipType, input1breaksPipeline);
                    if (this.keys1 != null) {
                        c1.setShipStrategy(shipType, this.keys1.toFieldList(), exMode);
                    } else {
                        c1.setShipStrategy(shipType, exMode);
                    }
                    if (dopChange1) {
                        c1.adjustGlobalPropertiesForFullParallelismChange();
                    }
                }
                for (RequestedGlobalProperties igps2 : intGlobal2) {
                    // create a candidate channel for the second input. mark it cached, if the
                    // connection says so
                    final Channel c2 = new Channel(child2, this.input2.getMaterializationMode());
                    if (this.input2.getShipStrategy() == null) {
                        // free to choose the ship strategy
                        igps2.parameterizeChannel(c2, dopChange2, input2Mode, input2breaksPipeline);
                        // parallelisms
                        if (dopChange2 && !c2.getShipStrategy().isNetworkStrategy()) {
                            c2.getGlobalProperties().reset();
                        }
                    } else {
                        // ship strategy fixed by compiler hint
                        ShipStrategyType shipType = this.input2.getShipStrategy();
                        DataExchangeMode exMode = DataExchangeMode.select(input2Mode, shipType, input2breaksPipeline);
                        if (this.keys2 != null) {
                            c2.setShipStrategy(shipType, this.keys2.toFieldList(), exMode);
                        } else {
                            c2.setShipStrategy(shipType, exMode);
                        }
                        if (dopChange2) {
                            c2.adjustGlobalPropertiesForFullParallelismChange();
                        }
                    }
                    outer: for (GlobalPropertiesPair gpp : allGlobalPairs) {
                        if (gpp.getProperties1().isMetBy(c1.getGlobalProperties()) && gpp.getProperties2().isMetBy(c2.getGlobalProperties())) {
                            for (OperatorDescriptorDual desc : getProperties()) {
                                if (desc.areCompatible(gpp.getProperties1(), gpp.getProperties2(), c1.getGlobalProperties(), c2.getGlobalProperties())) {
                                    Channel c1Clone = c1.clone();
                                    c1Clone.setRequiredGlobalProps(gpp.getProperties1());
                                    c2.setRequiredGlobalProps(gpp.getProperties2());
                                    // we form a valid combination, so create the local
                                    // candidates
                                    // for this
                                    addLocalCandidates(c1Clone, c2, broadcastPlanChannels, igps1, igps2, outputPlans, allLocalPairs, estimator);
                                    break outer;
                                }
                            }
                        }
                    }
                    // so we can stop after the first
                    if (this.input2.getShipStrategy() != null) {
                        break;
                    }
                }
                // so we can stop after the first
                if (this.input1.getShipStrategy() != null) {
                    break;
                }
            }
        }
    }
    if (outputPlans.isEmpty()) {
        if (childrenSkippedDueToReplicatedInput) {
            throw new CompilerException("No plan meeting the requirements could be created @ " + this + ". Most likely reason: Invalid use of replicated input.");
        } else {
            throw new CompilerException("No plan meeting the requirements could be created @ " + this + ". Most likely reason: Too restrictive plan hints.");
        }
    }
    // cost and prune the plans
    for (PlanNode node : outputPlans) {
        estimator.costOperator(node);
    }
    prunePlanAlternatives(outputPlans);
    outputPlans.trimToSize();
    this.cachedPlans = outputPlans;
    return outputPlans;
}
Also used : HashSet(java.util.HashSet) Set(java.util.Set) ArrayList(java.util.ArrayList) GlobalPropertiesPair(org.apache.flink.optimizer.operators.OperatorDescriptorDual.GlobalPropertiesPair) OperatorDescriptorDual(org.apache.flink.optimizer.operators.OperatorDescriptorDual) LocalPropertiesPair(org.apache.flink.optimizer.operators.OperatorDescriptorDual.LocalPropertiesPair) ShipStrategyType(org.apache.flink.runtime.operators.shipping.ShipStrategyType) DualInputPlanNode(org.apache.flink.optimizer.plan.DualInputPlanNode) PlanNode(org.apache.flink.optimizer.plan.PlanNode) DataExchangeMode(org.apache.flink.runtime.io.network.DataExchangeMode) CompilerException(org.apache.flink.optimizer.CompilerException) HashSet(java.util.HashSet) RequestedGlobalProperties(org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties) Channel(org.apache.flink.optimizer.plan.Channel) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel) ExecutionMode(org.apache.flink.api.common.ExecutionMode) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel)

Example 5 with OperatorDescriptorDual

use of org.apache.flink.optimizer.operators.OperatorDescriptorDual in project flink by apache.

the class TwoInputNode method addLocalCandidates.

protected void addLocalCandidates(Channel template1, Channel template2, List<Set<? extends NamedChannel>> broadcastPlanChannels, RequestedGlobalProperties rgps1, RequestedGlobalProperties rgps2, List<PlanNode> target, LocalPropertiesPair[] validLocalCombinations, CostEstimator estimator) {
    for (RequestedLocalProperties ilp1 : this.input1.getInterestingProperties().getLocalProperties()) {
        final Channel in1 = template1.clone();
        ilp1.parameterizeChannel(in1);
        for (RequestedLocalProperties ilp2 : this.input2.getInterestingProperties().getLocalProperties()) {
            final Channel in2 = template2.clone();
            ilp2.parameterizeChannel(in2);
            for (OperatorDescriptorDual dps : getProperties()) {
                for (LocalPropertiesPair lpp : dps.getPossibleLocalProperties()) {
                    if (lpp.getProperties1().isMetBy(in1.getLocalProperties()) && lpp.getProperties2().isMetBy(in2.getLocalProperties())) {
                        // sort order
                        if (dps.areCoFulfilled(lpp.getProperties1(), lpp.getProperties2(), in1.getLocalProperties(), in2.getLocalProperties())) {
                            // copy, because setting required properties and instantiation may
                            // change the channels and should not affect prior candidates
                            Channel in1Copy = in1.clone();
                            in1Copy.setRequiredLocalProps(lpp.getProperties1());
                            Channel in2Copy = in2.clone();
                            in2Copy.setRequiredLocalProps(lpp.getProperties2());
                            // all right, co compatible
                            instantiate(dps, in1Copy, in2Copy, broadcastPlanChannels, target, estimator, rgps1, rgps2, ilp1, ilp2);
                            break;
                        }
                    // else cannot use this pair, fall through the loop and try the next one
                    }
                }
            }
        }
    }
}
Also used : RequestedLocalProperties(org.apache.flink.optimizer.dataproperties.RequestedLocalProperties) Channel(org.apache.flink.optimizer.plan.Channel) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel) OperatorDescriptorDual(org.apache.flink.optimizer.operators.OperatorDescriptorDual) LocalPropertiesPair(org.apache.flink.optimizer.operators.OperatorDescriptorDual.LocalPropertiesPair)

Aggregations

OperatorDescriptorDual (org.apache.flink.optimizer.operators.OperatorDescriptorDual)11 CompilerException (org.apache.flink.optimizer.CompilerException)5 ArrayList (java.util.ArrayList)4 LocalPropertiesPair (org.apache.flink.optimizer.operators.OperatorDescriptorDual.LocalPropertiesPair)3 Ordering (org.apache.flink.api.common.operators.Ordering)2 AbstractJoinDescriptor (org.apache.flink.optimizer.operators.AbstractJoinDescriptor)2 HashJoinBuildFirstProperties (org.apache.flink.optimizer.operators.HashJoinBuildFirstProperties)2 HashJoinBuildSecondProperties (org.apache.flink.optimizer.operators.HashJoinBuildSecondProperties)2 GlobalPropertiesPair (org.apache.flink.optimizer.operators.OperatorDescriptorDual.GlobalPropertiesPair)2 Channel (org.apache.flink.optimizer.plan.Channel)2 NamedChannel (org.apache.flink.optimizer.plan.NamedChannel)2 HashSet (java.util.HashSet)1 Set (java.util.Set)1 ExecutionMode (org.apache.flink.api.common.ExecutionMode)1 JoinHint (org.apache.flink.api.common.operators.base.JoinOperatorBase.JoinHint)1 OuterJoinType (org.apache.flink.api.common.operators.base.OuterJoinOperatorBase.OuterJoinType)1 Configuration (org.apache.flink.configuration.Configuration)1 InterestingProperties (org.apache.flink.optimizer.dataproperties.InterestingProperties)1 RequestedGlobalProperties (org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties)1 RequestedLocalProperties (org.apache.flink.optimizer.dataproperties.RequestedLocalProperties)1