in src/org/apache/pig/backend/hadoop/executionengine/tez/plan/optimizer/MultiQueryOptimizerTez.java [58:307]
public void visitTezOp(TezOperator tezOp) throws VisitorException {
try {
if (!tezOp.isSplitter()) {
return;
}
// Using a list instead of set to have consistently ordered plans
List<TezOperator> splittees = new ArrayList<TezOperator>();
Set<TezOperator> mergedNonPackageInputSuccessors = new HashSet<TezOperator>();
// When there is a union successor with unsupported storefunc, those splittees
// can only be merged into the split if all the union members will be from the split
// This is to ensure that there are no vertex groups created with unsupported storefunc.
Map<TezOperator, Set<OperatorKey>> tentativeMergeUnionMembers = new HashMap<TezOperator, Set<OperatorKey>>();
List<TezOperator> successors = getPlan().getSuccessors(tezOp);
Set<OperatorKey> splitterAndSuccessorKeys = new HashSet<OperatorKey>();
splitterAndSuccessorKeys.add(tezOp.getOperatorKey());
for (TezOperator successor : successors) {
splitterAndSuccessorKeys.add(successor.getOperatorKey());
}
for (TezOperator successor : successors) {
HashSet<TezOperator> predecessors = new HashSet<TezOperator>(getPlan().getPredecessors(successor));
predecessors.remove(tezOp);
if (!predecessors.isEmpty()) {
// If has other dependency that conflicts with other splittees, don't merge into split
// For eg: self replicate join/skewed join
// But if replicate input is from a different operator allow it, but ensure
// that we don't have more than one input coming from that operator into the split
// Check if other splittees or its predecessors (till the root) are not present in
// the predecessors (till the root) of this splittee.
// Need to check the whole predecessors hierarchy till root as the conflict
// could be multiple levels up
for (TezOperator predecessor : getPlan().getPredecessors(successor)) {
if (predecessor != tezOp) {
predecessors.add(predecessor);
TezCompilerUtil.addAllPredecessors(getPlan(), predecessor, predecessors);
}
}
Set<TezOperator> toMergeSuccPredecessors = new HashSet<TezOperator>(successors);
toMergeSuccPredecessors.remove(successor);
for (TezOperator splittee : splittees) {
for (TezOperator spliteePred : getPlan().getPredecessors(splittee)) {
if (spliteePred != tezOp) {
toMergeSuccPredecessors.add(spliteePred);
TezCompilerUtil.addAllPredecessors(getPlan(), spliteePred, toMergeSuccPredecessors);
}
}
}
if (predecessors.removeAll(toMergeSuccPredecessors)) {
continue;
}
}
// Split contains right input of different skewed joins
if (successor.getSampleOperator() != null
&& tezOp.getSampleOperator() != null
&& !successor.getSampleOperator().equals(
tezOp.getSampleOperator())) {
continue;
}
// Detect diamond shape into successor operator, we cannot merge it into split,
// since Tez does not handle double edge between vertexes
// Successor could be
// - union operator (if no union optimizer changing it to vertex group which supports multiple edges)
// - self replicate join, self skewed join or scalar
// - POPackage (Self hash joins can write to same output edge and is handled by POShuffleTezLoad)
Set<TezOperator> mergedSuccessors = new HashSet<TezOperator>();
// These successors should not be merged due to diamond shape
Set<TezOperator> toNotMergeSuccessors = new HashSet<TezOperator>();
// These successors can be merged
Set<TezOperator> toMergeSuccessors = new HashSet<TezOperator>();
// These successors (Scalar, POFRJoinTez) can be merged if they are the only input.
// Only in case of POPackage(POShuffleTezLoad) multiple inputs can be handled from a Split
Set<TezOperator> nonPackageInputSuccessors = new HashSet<TezOperator>();
boolean canMerge = true;
Set<TezOperator> successorUnsupportedStoreUnions = new HashSet<TezOperator>();
mergedSuccessors.addAll(successors);
for (TezOperator splittee : splittees) {
if (getPlan().getSuccessors(splittee) != null) {
mergedSuccessors.addAll(getPlan().getSuccessors(splittee));
}
}
if (getPlan().getSuccessors(successor) != null) {
nonPackageInputSuccessors.clear();
toMergeSuccessors.clear();
for (TezOperator succSuccessor : getPlan().getSuccessors(successor)) {
if (succSuccessor.isUnion()) {
if (!(unionOptimizerOn &&
UnionOptimizer.isOptimizable(succSuccessor))) {
toNotMergeSuccessors.add(succSuccessor);
} else {
if (unionOptimizerOn && !UnionOptimizer.isOptimizableStoreFunc(succSuccessor,unionSupportedStoreFuncs,unionUnsupportedStoreFuncs)) {
// This optimization of using UnionOptimizer for unsupported storefunc
// is only good for one level of split and does not handle multiple level of split.
Set<OperatorKey> unionMembers = new HashSet<OperatorKey>(succSuccessor.getUnionMembers());
unionMembers.removeAll(splitterAndSuccessorKeys);
if(unionMembers.isEmpty()) {
successorUnsupportedStoreUnions.add(succSuccessor);
} else {
toNotMergeSuccessors.add(succSuccessor);
continue;
}
}
if (TezCompilerUtil.isNonPackageInput(successor.getOperatorKey().toString(), succSuccessor)) {
// Output goes to scalar or POFRJoinTez in the union operator
// We need to ensure it is the only one to avoid parallel edges
canMerge = canMerge ? nonPackageInputSuccessors.add(succSuccessor) : false;
} else {
toMergeSuccessors.add(succSuccessor);
}
List<TezOperator> unionSuccessors = getPlan().getSuccessors(succSuccessor);
if (unionSuccessors != null) {
for (TezOperator unionSuccessor : unionSuccessors) {
if (TezCompilerUtil.isNonPackageInput(succSuccessor.getOperatorKey().toString(), unionSuccessor)) {
canMerge = canMerge ? nonPackageInputSuccessors.add(unionSuccessor) : false;
} else {
toMergeSuccessors.add(unionSuccessor);
}
}
}
}
} else if (TezCompilerUtil.isNonPackageInput(successor.getOperatorKey().toString(), succSuccessor)) {
// Output goes to scalar or POFRJoinTez instead of POPackage
// POPackage/POShuffleTezLoad can handle multiple inputs from a Split.
// But if input is sent to any other operator like
// scalar, POFRJoinTez then we need to ensure it is the only one.
canMerge = canMerge ? nonPackageInputSuccessors.add(succSuccessor) : false;
} else {
toMergeSuccessors.add(succSuccessor);
}
}
}
if (canMerge) {
if (!nonPackageInputSuccessors.isEmpty() || !mergedNonPackageInputSuccessors.isEmpty()) {
// If a non-POPackage input successor is already merged or
// if there is a POPackage and non-POPackage to be merged,
// then skip as it will become diamond shape
// For eg: POFRJoinTez+Scalar, POFRJoinTez/Scalar+POPackage
if (nonPackageInputSuccessors.removeAll(mergedSuccessors)
|| toMergeSuccessors.removeAll(mergedNonPackageInputSuccessors)
|| toMergeSuccessors.removeAll(nonPackageInputSuccessors)) {
continue;
}
}
} else {
continue;
}
mergedSuccessors.retainAll(toNotMergeSuccessors);
if (mergedSuccessors.isEmpty()) { // no shared edge after merge
mergedNonPackageInputSuccessors.addAll(nonPackageInputSuccessors);
if (successorUnsupportedStoreUnions.isEmpty()) {
splittees.add(successor);
} else {
// If all other conditions were satisfied, but it had a successor union
// with unsupported storefunc keep it in the tentative list
for (TezOperator unionOp : successorUnsupportedStoreUnions) {
Set<OperatorKey> tentativeSuccessors = tentativeMergeUnionMembers.get(unionOp);
if (tentativeSuccessors == null) {
tentativeSuccessors = new HashSet<OperatorKey>();
tentativeMergeUnionMembers.put(unionOp, tentativeSuccessors);
}
tentativeSuccessors.add(successor.getOperatorKey());
}
}
}
}
Set<TezOperator> spliteesToRemove = new HashSet<TezOperator>();
for (Entry<TezOperator, Set<OperatorKey>> entry : tentativeMergeUnionMembers.entrySet()) {
Set<OperatorKey> unionMembers = new HashSet<OperatorKey>(entry.getKey().getUnionMembers());
if (entry.getValue().containsAll(unionMembers)) {
// If all the union members were tentative splittees then add them
for (OperatorKey key : entry.getValue()) {
TezOperator splittee = getPlan().getOperator(key);
if (!splittees.contains(splittee)) {
splittees.add(splittee);
}
}
} else {
for (OperatorKey key : entry.getValue()) {
spliteesToRemove.add(getPlan().getOperator(key));
}
}
}
for (TezOperator op : spliteesToRemove) {
splittees.remove(op);
}
if (splittees.size() == 0) {
return;
}
if (splittees.size()==1 && successors.size()==1) {
// We don't need a POSplit here, we can merge the splittee into spliter
PhysicalOperator firstNodeLeaf = tezOp.plan.getLeaves().get(0);
PhysicalOperator firstNodeLeafPred = tezOp.plan.getPredecessors(firstNodeLeaf).get(0);
TezOperator singleSplitee = splittees.get(0);
PhysicalOperator secondNodeRoot = singleSplitee.plan.getRoots().get(0);
PhysicalOperator secondNodeSucc = singleSplitee.plan.getSuccessors(secondNodeRoot).get(0);
tezOp.plan.remove(firstNodeLeaf);
singleSplitee.plan.remove(secondNodeRoot);
tezOp.plan.merge(singleSplitee.plan);
tezOp.plan.connect(firstNodeLeafPred, secondNodeSucc);
addSubPlanPropertiesToParent(tezOp, singleSplitee);
removeSplittee(getPlan(), tezOp, singleSplitee);
} else {
POValueOutputTez valueOutput = (POValueOutputTez)tezOp.plan.getLeaves().get(0);
POSplit split = new POSplit(OperatorKey.genOpKey(valueOutput.getOperatorKey().getScope()));
split.copyAliasFrom(valueOutput);
for (TezOperator splitee : splittees) {
PhysicalOperator spliteeRoot = splitee.plan.getRoots().get(0);
splitee.plan.remove(spliteeRoot);
split.addPlan(splitee.plan);
addSubPlanPropertiesToParent(tezOp, splitee);
removeSplittee(getPlan(), tezOp, splitee);
valueOutput.removeOutputKey(splitee.getOperatorKey().toString());
}
if (valueOutput.getTezOutputs().length > 0) {
// We still need valueOutput
PhysicalPlan phyPlan = new PhysicalPlan();
phyPlan.addAsLeaf(valueOutput);
split.addPlan(phyPlan);
}
PhysicalOperator pred = tezOp.plan.getPredecessors(valueOutput).get(0);
tezOp.plan.disconnect(pred, valueOutput);
tezOp.plan.remove(valueOutput);
tezOp.plan.add(split);
tezOp.plan.connect(pred, split);
}
} catch (PlanException e) {
throw new VisitorException(e);
}
}