Search in sources :

Example 1 with NamedChannel

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

the class WorksetIterationNode method instantiate.

@SuppressWarnings("unchecked")
@Override
protected void instantiate(OperatorDescriptorDual operator, Channel solutionSetIn, Channel worksetIn, List<Set<? extends NamedChannel>> broadcastPlanChannels, List<PlanNode> target, CostEstimator estimator, RequestedGlobalProperties globPropsReqSolutionSet, RequestedGlobalProperties globPropsReqWorkset, RequestedLocalProperties locPropsReqSolutionSet, RequestedLocalProperties locPropsReqWorkset) {
    // check for pipeline breaking using hash join with build on the solution set side
    placePipelineBreakersIfNecessary(DriverStrategy.HYBRIDHASH_BUILD_FIRST, solutionSetIn, worksetIn);
    // NOTES ON THE ENUMERATION OF THE STEP FUNCTION PLANS:
    // Whenever we instantiate the iteration, we enumerate new candidates for the step function.
    // That way, we make sure we have an appropriate plan for each candidate for the initial
    // partial solution,
    // we have a fitting candidate for the step function (often, work is pushed out of the step
    // function).
    // Among the candidates of the step function, we keep only those that meet the requested
    // properties of the
    // current candidate initial partial solution. That makes sure these properties exist at the
    // beginning of
    // every iteration.
    // 1) Because we enumerate multiple times, we may need to clean the cached plans
    // before starting another enumeration
    this.nextWorkset.accept(PlanCacheCleaner.INSTANCE);
    this.solutionSetDelta.accept(PlanCacheCleaner.INSTANCE);
    // 2) Give the partial solution the properties of the current candidate for the initial
    // partial solution
    // This concerns currently only the workset.
    this.worksetNode.setCandidateProperties(worksetIn.getGlobalProperties(), worksetIn.getLocalProperties(), worksetIn);
    this.solutionSetNode.setCandidateProperties(this.partitionedProperties, new LocalProperties(), solutionSetIn);
    final SolutionSetPlanNode sspn = this.solutionSetNode.getCurrentSolutionSetPlanNode();
    final WorksetPlanNode wspn = this.worksetNode.getCurrentWorksetPlanNode();
    // 3) Get the alternative plans
    List<PlanNode> solutionSetDeltaCandidates = this.solutionSetDelta.getAlternativePlans(estimator);
    List<PlanNode> worksetCandidates = this.nextWorkset.getAlternativePlans(estimator);
    // 4) Throw away all that are not compatible with the properties currently requested to the
    // initial partial solution
    // Make sure that the workset candidates fulfill the input requirements
    {
        List<PlanNode> newCandidates = new ArrayList<PlanNode>();
        for (Iterator<PlanNode> planDeleter = worksetCandidates.iterator(); planDeleter.hasNext(); ) {
            PlanNode candidate = planDeleter.next();
            GlobalProperties atEndGlobal = candidate.getGlobalProperties();
            LocalProperties atEndLocal = candidate.getLocalProperties();
            FeedbackPropertiesMeetRequirementsReport report = candidate.checkPartialSolutionPropertiesMet(wspn, atEndGlobal, atEndLocal);
            if (report == FeedbackPropertiesMeetRequirementsReport.NO_PARTIAL_SOLUTION) {
            // depends only through broadcast variable on the workset solution
            } else if (report == FeedbackPropertiesMeetRequirementsReport.NOT_MET) {
                // attach a no-op node through which we create the properties of the original
                // input
                Channel toNoOp = new Channel(candidate);
                globPropsReqWorkset.parameterizeChannel(toNoOp, false, nextWorksetRootConnection.getDataExchangeMode(), false);
                locPropsReqWorkset.parameterizeChannel(toNoOp);
                NoOpUnaryUdfOp noOpUnaryUdfOp = new NoOpUnaryUdfOp<>();
                noOpUnaryUdfOp.setInput(candidate.getProgramOperator());
                UnaryOperatorNode rebuildWorksetPropertiesNode = new UnaryOperatorNode("Rebuild Workset Properties", noOpUnaryUdfOp, true);
                rebuildWorksetPropertiesNode.setParallelism(candidate.getParallelism());
                SingleInputPlanNode rebuildWorksetPropertiesPlanNode = new SingleInputPlanNode(rebuildWorksetPropertiesNode, "Rebuild Workset Properties", toNoOp, DriverStrategy.UNARY_NO_OP);
                rebuildWorksetPropertiesPlanNode.initProperties(toNoOp.getGlobalProperties(), toNoOp.getLocalProperties());
                estimator.costOperator(rebuildWorksetPropertiesPlanNode);
                GlobalProperties atEndGlobalModified = rebuildWorksetPropertiesPlanNode.getGlobalProperties();
                LocalProperties atEndLocalModified = rebuildWorksetPropertiesPlanNode.getLocalProperties();
                if (!(atEndGlobalModified.equals(atEndGlobal) && atEndLocalModified.equals(atEndLocal))) {
                    FeedbackPropertiesMeetRequirementsReport report2 = candidate.checkPartialSolutionPropertiesMet(wspn, atEndGlobalModified, atEndLocalModified);
                    if (report2 != FeedbackPropertiesMeetRequirementsReport.NOT_MET) {
                        newCandidates.add(rebuildWorksetPropertiesPlanNode);
                    }
                }
                // remove the original operator and add the modified candidate
                planDeleter.remove();
            }
        }
        worksetCandidates.addAll(newCandidates);
    }
    if (worksetCandidates.isEmpty()) {
        return;
    }
    // sanity check the solution set delta
    for (PlanNode solutionSetDeltaCandidate : solutionSetDeltaCandidates) {
        SingleInputPlanNode candidate = (SingleInputPlanNode) solutionSetDeltaCandidate;
        GlobalProperties gp = candidate.getGlobalProperties();
        if (gp.getPartitioning() != PartitioningProperty.HASH_PARTITIONED || gp.getPartitioningFields() == null || !gp.getPartitioningFields().equals(this.solutionSetKeyFields)) {
            throw new CompilerException("Bug: The solution set delta is not partitioned.");
        }
    }
    // 5) Create a candidate for the Iteration Node for every remaining plan of the step
    // function.
    final GlobalProperties gp = new GlobalProperties();
    gp.setHashPartitioned(this.solutionSetKeyFields);
    gp.addUniqueFieldCombination(this.solutionSetKeyFields);
    LocalProperties lp = LocalProperties.EMPTY.addUniqueFields(this.solutionSetKeyFields);
    // take all combinations of solution set delta and workset plans
    for (PlanNode worksetCandidate : worksetCandidates) {
        for (PlanNode solutionSetCandidate : solutionSetDeltaCandidates) {
            // check whether they have the same operator at their latest branching point
            if (this.singleRoot.areBranchCompatible(solutionSetCandidate, worksetCandidate)) {
                SingleInputPlanNode siSolutionDeltaCandidate = (SingleInputPlanNode) solutionSetCandidate;
                boolean immediateDeltaUpdate;
                // can update on the fly
                if (siSolutionDeltaCandidate.getInput().getShipStrategy() == ShipStrategyType.FORWARD && this.solutionDeltaImmediatelyAfterSolutionJoin) {
                    // sanity check the node and connection
                    if (siSolutionDeltaCandidate.getDriverStrategy() != DriverStrategy.UNARY_NO_OP || siSolutionDeltaCandidate.getInput().getLocalStrategy() != LocalStrategy.NONE) {
                        throw new CompilerException("Invalid Solution set delta node.");
                    }
                    solutionSetCandidate = siSolutionDeltaCandidate.getInput().getSource();
                    immediateDeltaUpdate = true;
                } else {
                    // was not partitioned, we need to keep this node.
                    // mark that we materialize the input
                    siSolutionDeltaCandidate.getInput().setTempMode(TempMode.PIPELINE_BREAKER);
                    immediateDeltaUpdate = false;
                }
                WorksetIterationPlanNode wsNode = new WorksetIterationPlanNode(this, this.getOperator().getName(), solutionSetIn, worksetIn, sspn, wspn, worksetCandidate, solutionSetCandidate);
                wsNode.setImmediateSolutionSetUpdate(immediateDeltaUpdate);
                wsNode.initProperties(gp, lp);
                target.add(wsNode);
            }
        }
    }
}
Also used : SolutionSetPlanNode(org.apache.flink.optimizer.plan.SolutionSetPlanNode) FeedbackPropertiesMeetRequirementsReport(org.apache.flink.optimizer.plan.PlanNode.FeedbackPropertiesMeetRequirementsReport) WorksetIterationPlanNode(org.apache.flink.optimizer.plan.WorksetIterationPlanNode) Channel(org.apache.flink.optimizer.plan.Channel) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) WorksetIterationPlanNode(org.apache.flink.optimizer.plan.WorksetIterationPlanNode) SolutionSetPlanNode(org.apache.flink.optimizer.plan.SolutionSetPlanNode) WorksetPlanNode(org.apache.flink.optimizer.plan.WorksetPlanNode) DualInputPlanNode(org.apache.flink.optimizer.plan.DualInputPlanNode) PlanNode(org.apache.flink.optimizer.plan.PlanNode) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) RequestedGlobalProperties(org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties) GlobalProperties(org.apache.flink.optimizer.dataproperties.GlobalProperties) NoOpUnaryUdfOp(org.apache.flink.optimizer.util.NoOpUnaryUdfOp) Iterator(java.util.Iterator) WorksetPlanNode(org.apache.flink.optimizer.plan.WorksetPlanNode) CompilerException(org.apache.flink.optimizer.CompilerException) ArrayList(java.util.ArrayList) FieldList(org.apache.flink.api.common.operators.util.FieldList) List(java.util.List) RequestedLocalProperties(org.apache.flink.optimizer.dataproperties.RequestedLocalProperties) LocalProperties(org.apache.flink.optimizer.dataproperties.LocalProperties)

Example 2 with NamedChannel

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

the class SingleInputNode 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;
    // calculate alternative sub-plans for predecessor
    final List<? extends PlanNode> subPlans = getPredecessorNode().getAlternativePlans(estimator);
    final Set<RequestedGlobalProperties> intGlobal = this.inConn.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) {
            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 RequestedGlobalProperties[] allValidGlobals;
    {
        Set<RequestedGlobalProperties> pairs = new HashSet<RequestedGlobalProperties>();
        for (OperatorDescriptorSingle ods : getPossibleProperties()) {
            pairs.addAll(ods.getPossibleGlobalProperties());
        }
        allValidGlobals = pairs.toArray(new RequestedGlobalProperties[pairs.size()]);
    }
    final ArrayList<PlanNode> outputPlans = new ArrayList<PlanNode>();
    final ExecutionMode executionMode = this.inConn.getDataExchangeMode();
    final int parallelism = getParallelism();
    final int inParallelism = getPredecessorNode().getParallelism();
    final boolean parallelismChange = inParallelism != parallelism;
    final boolean breaksPipeline = this.inConn.isBreakingPipeline();
    // create all candidates
    for (PlanNode child : subPlans) {
        if (child.getGlobalProperties().isFullyReplicated()) {
            // changed
            if (parallelismChange) {
                // can not continue with this child
                childrenSkippedDueToReplicatedInput = true;
                continue;
            } else {
                this.inConn.setShipStrategy(ShipStrategyType.FORWARD);
            }
        }
        if (this.inConn.getShipStrategy() == null) {
            // pick the strategy ourselves
            for (RequestedGlobalProperties igps : intGlobal) {
                final Channel c = new Channel(child, this.inConn.getMaterializationMode());
                igps.parameterizeChannel(c, parallelismChange, executionMode, breaksPipeline);
                // ship strategy preserves/establishes them even under changing parallelisms
                if (parallelismChange && !c.getShipStrategy().isNetworkStrategy()) {
                    c.getGlobalProperties().reset();
                }
                // requested properties
                for (RequestedGlobalProperties rgps : allValidGlobals) {
                    if (rgps.isMetBy(c.getGlobalProperties())) {
                        c.setRequiredGlobalProps(rgps);
                        addLocalCandidates(c, broadcastPlanChannels, igps, outputPlans, estimator);
                        break;
                    }
                }
            }
        } else {
            // hint fixed the strategy
            final Channel c = new Channel(child, this.inConn.getMaterializationMode());
            final ShipStrategyType shipStrategy = this.inConn.getShipStrategy();
            final DataExchangeMode exMode = DataExchangeMode.select(executionMode, shipStrategy, breaksPipeline);
            if (this.keys != null) {
                c.setShipStrategy(shipStrategy, this.keys.toFieldList(), exMode);
            } else {
                c.setShipStrategy(shipStrategy, exMode);
            }
            if (parallelismChange) {
                c.adjustGlobalPropertiesForFullParallelismChange();
            }
            // check whether we meet any of the accepted properties
            for (RequestedGlobalProperties rgps : allValidGlobals) {
                if (rgps.isMetBy(c.getGlobalProperties())) {
                    addLocalCandidates(c, broadcastPlanChannels, rgps, outputPlans, estimator);
                    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) FieldSet(org.apache.flink.api.common.operators.util.FieldSet) ArrayList(java.util.ArrayList) ShipStrategyType(org.apache.flink.runtime.operators.shipping.ShipStrategyType) OperatorDescriptorSingle(org.apache.flink.optimizer.operators.OperatorDescriptorSingle) PlanNode(org.apache.flink.optimizer.plan.PlanNode) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) 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 3 with NamedChannel

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

the class SingleInputNode method instantiateCandidate.

protected void instantiateCandidate(OperatorDescriptorSingle dps, Channel in, List<Set<? extends NamedChannel>> broadcastPlanChannels, List<PlanNode> target, CostEstimator estimator, RequestedGlobalProperties globPropsReq, RequestedLocalProperties locPropsReq) {
    final PlanNode inputSource = in.getSource();
    for (List<NamedChannel> broadcastChannelsCombination : Sets.cartesianProduct(broadcastPlanChannels)) {
        boolean validCombination = true;
        boolean requiresPipelinebreaker = false;
        // check whether the broadcast inputs use the same plan candidate at the branching point
        for (int i = 0; i < broadcastChannelsCombination.size(); i++) {
            NamedChannel nc = broadcastChannelsCombination.get(i);
            PlanNode bcSource = nc.getSource();
            // check branch compatibility against input
            if (!areBranchCompatible(bcSource, inputSource)) {
                validCombination = false;
                break;
            }
            // check branch compatibility against all other broadcast variables
            for (int k = 0; k < i; k++) {
                PlanNode otherBcSource = broadcastChannelsCombination.get(k).getSource();
                if (!areBranchCompatible(bcSource, otherBcSource)) {
                    validCombination = false;
                    break;
                }
            }
            // all common predecessors
            if (in.isOnDynamicPath() && this.hereJoinedBranches != null) {
                for (OptimizerNode brancher : this.hereJoinedBranches) {
                    PlanNode candAtBrancher = in.getSource().getCandidateAtBranchPoint(brancher);
                    if (candAtBrancher == null) {
                        // closed branch between two broadcast variables
                        continue;
                    }
                    SourceAndDamReport res = in.getSource().hasDamOnPathDownTo(candAtBrancher);
                    if (res == NOT_FOUND) {
                        throw new CompilerException("Bug: Tracing dams for deadlock detection is broken.");
                    } else if (res == FOUND_SOURCE) {
                        requiresPipelinebreaker = true;
                        break;
                    } else if (res == FOUND_SOURCE_AND_DAM) {
                    // good
                    } else {
                        throw new CompilerException();
                    }
                }
            }
        }
        if (!validCombination) {
            continue;
        }
        if (requiresPipelinebreaker) {
            in.setTempMode(in.getTempMode().makePipelineBreaker());
        }
        final SingleInputPlanNode node = dps.instantiate(in, this);
        node.setBroadcastInputs(broadcastChannelsCombination);
        // compute how the strategy affects the properties
        GlobalProperties gProps = in.getGlobalProperties().clone();
        LocalProperties lProps = in.getLocalProperties().clone();
        gProps = dps.computeGlobalProperties(gProps);
        lProps = dps.computeLocalProperties(lProps);
        // filter by the user code field copies
        gProps = gProps.filterBySemanticProperties(getSemanticPropertiesForGlobalPropertyFiltering(), 0);
        lProps = lProps.filterBySemanticProperties(getSemanticPropertiesForLocalPropertyFiltering(), 0);
        // apply
        node.initProperties(gProps, lProps);
        node.updatePropertiesWithUniqueSets(getUniqueFields());
        target.add(node);
    }
}
Also used : SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) PlanNode(org.apache.flink.optimizer.plan.PlanNode) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) RequestedGlobalProperties(org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties) GlobalProperties(org.apache.flink.optimizer.dataproperties.GlobalProperties) SourceAndDamReport(org.apache.flink.optimizer.plan.PlanNode.SourceAndDamReport) CompilerException(org.apache.flink.optimizer.CompilerException) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel) RequestedLocalProperties(org.apache.flink.optimizer.dataproperties.RequestedLocalProperties) LocalProperties(org.apache.flink.optimizer.dataproperties.LocalProperties)

Example 4 with NamedChannel

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

the class BulkIterationNode method instantiateCandidate.

@SuppressWarnings("unchecked")
@Override
protected void instantiateCandidate(OperatorDescriptorSingle dps, Channel in, List<Set<? extends NamedChannel>> broadcastPlanChannels, List<PlanNode> target, CostEstimator estimator, RequestedGlobalProperties globPropsReq, RequestedLocalProperties locPropsReq) {
    // NOTES ON THE ENUMERATION OF THE STEP FUNCTION PLANS:
    // Whenever we instantiate the iteration, we enumerate new candidates for the step function.
    // That way, we make sure we have an appropriate plan for each candidate for the initial
    // partial solution,
    // we have a fitting candidate for the step function (often, work is pushed out of the step
    // function).
    // Among the candidates of the step function, we keep only those that meet the requested
    // properties of the
    // current candidate initial partial solution. That makes sure these properties exist at the
    // beginning of
    // the successive iteration.
    // 1) Because we enumerate multiple times, we may need to clean the cached plans
    // before starting another enumeration
    this.nextPartialSolution.accept(PlanCacheCleaner.INSTANCE);
    if (this.terminationCriterion != null) {
        this.terminationCriterion.accept(PlanCacheCleaner.INSTANCE);
    }
    // 2) Give the partial solution the properties of the current candidate for the initial
    // partial solution
    this.partialSolution.setCandidateProperties(in.getGlobalProperties(), in.getLocalProperties(), in);
    final BulkPartialSolutionPlanNode pspn = this.partialSolution.getCurrentPartialSolutionPlanNode();
    // 3) Get the alternative plans
    List<PlanNode> candidates = this.nextPartialSolution.getAlternativePlans(estimator);
    // 4) Make sure that the beginning of the step function does not assume properties that
    // are not also produced by the end of the step function.
    {
        List<PlanNode> newCandidates = new ArrayList<PlanNode>();
        for (Iterator<PlanNode> planDeleter = candidates.iterator(); planDeleter.hasNext(); ) {
            PlanNode candidate = planDeleter.next();
            GlobalProperties atEndGlobal = candidate.getGlobalProperties();
            LocalProperties atEndLocal = candidate.getLocalProperties();
            FeedbackPropertiesMeetRequirementsReport report = candidate.checkPartialSolutionPropertiesMet(pspn, atEndGlobal, atEndLocal);
            if (report == FeedbackPropertiesMeetRequirementsReport.NO_PARTIAL_SOLUTION) {
            // depends only through broadcast variable on the partial solution
            } else if (report == FeedbackPropertiesMeetRequirementsReport.NOT_MET) {
                // attach a no-op node through which we create the properties of the original
                // input
                Channel toNoOp = new Channel(candidate);
                globPropsReq.parameterizeChannel(toNoOp, false, rootConnection.getDataExchangeMode(), false);
                locPropsReq.parameterizeChannel(toNoOp);
                NoOpUnaryUdfOp noOpUnaryUdfOp = new NoOpUnaryUdfOp<>();
                noOpUnaryUdfOp.setInput(candidate.getProgramOperator());
                UnaryOperatorNode rebuildPropertiesNode = new UnaryOperatorNode("Rebuild Partial Solution Properties", noOpUnaryUdfOp, true);
                rebuildPropertiesNode.setParallelism(candidate.getParallelism());
                SingleInputPlanNode rebuildPropertiesPlanNode = new SingleInputPlanNode(rebuildPropertiesNode, "Rebuild Partial Solution Properties", toNoOp, DriverStrategy.UNARY_NO_OP);
                rebuildPropertiesPlanNode.initProperties(toNoOp.getGlobalProperties(), toNoOp.getLocalProperties());
                estimator.costOperator(rebuildPropertiesPlanNode);
                GlobalProperties atEndGlobalModified = rebuildPropertiesPlanNode.getGlobalProperties();
                LocalProperties atEndLocalModified = rebuildPropertiesPlanNode.getLocalProperties();
                if (!(atEndGlobalModified.equals(atEndGlobal) && atEndLocalModified.equals(atEndLocal))) {
                    FeedbackPropertiesMeetRequirementsReport report2 = candidate.checkPartialSolutionPropertiesMet(pspn, atEndGlobalModified, atEndLocalModified);
                    if (report2 != FeedbackPropertiesMeetRequirementsReport.NOT_MET) {
                        newCandidates.add(rebuildPropertiesPlanNode);
                    }
                }
                planDeleter.remove();
            }
        }
        candidates.addAll(newCandidates);
    }
    if (candidates.isEmpty()) {
        return;
    }
    // function.
    if (terminationCriterion == null) {
        for (PlanNode candidate : candidates) {
            BulkIterationPlanNode node = new BulkIterationPlanNode(this, this.getOperator().getName(), in, pspn, candidate);
            GlobalProperties gProps = candidate.getGlobalProperties().clone();
            LocalProperties lProps = candidate.getLocalProperties().clone();
            node.initProperties(gProps, lProps);
            target.add(node);
        }
    } else if (candidates.size() > 0) {
        List<PlanNode> terminationCriterionCandidates = this.terminationCriterion.getAlternativePlans(estimator);
        SingleRootJoiner singleRoot = (SingleRootJoiner) this.singleRoot;
        for (PlanNode candidate : candidates) {
            for (PlanNode terminationCandidate : terminationCriterionCandidates) {
                if (singleRoot.areBranchCompatible(candidate, terminationCandidate)) {
                    BulkIterationPlanNode node = new BulkIterationPlanNode(this, "BulkIteration (" + this.getOperator().getName() + ")", in, pspn, candidate, terminationCandidate);
                    GlobalProperties gProps = candidate.getGlobalProperties().clone();
                    LocalProperties lProps = candidate.getLocalProperties().clone();
                    node.initProperties(gProps, lProps);
                    target.add(node);
                }
            }
        }
    }
}
Also used : FeedbackPropertiesMeetRequirementsReport(org.apache.flink.optimizer.plan.PlanNode.FeedbackPropertiesMeetRequirementsReport) BulkPartialSolutionPlanNode(org.apache.flink.optimizer.plan.BulkPartialSolutionPlanNode) Channel(org.apache.flink.optimizer.plan.Channel) NamedChannel(org.apache.flink.optimizer.plan.NamedChannel) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) BulkIterationPlanNode(org.apache.flink.optimizer.plan.BulkIterationPlanNode) BulkPartialSolutionPlanNode(org.apache.flink.optimizer.plan.BulkPartialSolutionPlanNode) PlanNode(org.apache.flink.optimizer.plan.PlanNode) SingleInputPlanNode(org.apache.flink.optimizer.plan.SingleInputPlanNode) RequestedGlobalProperties(org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties) GlobalProperties(org.apache.flink.optimizer.dataproperties.GlobalProperties) SingleRootJoiner(org.apache.flink.optimizer.dag.WorksetIterationNode.SingleRootJoiner) NoOpUnaryUdfOp(org.apache.flink.optimizer.util.NoOpUnaryUdfOp) Iterator(java.util.Iterator) ArrayList(java.util.ArrayList) List(java.util.List) RequestedLocalProperties(org.apache.flink.optimizer.dataproperties.RequestedLocalProperties) LocalProperties(org.apache.flink.optimizer.dataproperties.LocalProperties) BulkIterationPlanNode(org.apache.flink.optimizer.plan.BulkIterationPlanNode)

Example 5 with NamedChannel

use of org.apache.flink.optimizer.plan.NamedChannel 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)

Aggregations

NamedChannel (org.apache.flink.optimizer.plan.NamedChannel)11 Channel (org.apache.flink.optimizer.plan.Channel)9 PlanNode (org.apache.flink.optimizer.plan.PlanNode)9 RequestedGlobalProperties (org.apache.flink.optimizer.dataproperties.RequestedGlobalProperties)7 RequestedLocalProperties (org.apache.flink.optimizer.dataproperties.RequestedLocalProperties)7 ArrayList (java.util.ArrayList)6 CompilerException (org.apache.flink.optimizer.CompilerException)6 GlobalProperties (org.apache.flink.optimizer.dataproperties.GlobalProperties)6 SingleInputPlanNode (org.apache.flink.optimizer.plan.SingleInputPlanNode)6 LocalProperties (org.apache.flink.optimizer.dataproperties.LocalProperties)5 DualInputPlanNode (org.apache.flink.optimizer.plan.DualInputPlanNode)4 Set (java.util.Set)3 ExecutionMode (org.apache.flink.api.common.ExecutionMode)3 DataExchangeMode (org.apache.flink.runtime.io.network.DataExchangeMode)3 ShipStrategyType (org.apache.flink.runtime.operators.shipping.ShipStrategyType)3 HashSet (java.util.HashSet)2 Iterator (java.util.Iterator)2 List (java.util.List)2 FieldList (org.apache.flink.api.common.operators.util.FieldList)2 FieldSet (org.apache.flink.api.common.operators.util.FieldSet)2