All Downloads are FREE. Search and download functionalities are using the official Maven repository.

org.apache.hadoop.hive.ql.parse.TezCompiler Maven / Gradle / Ivy

/*
 *  Licensed to the Apache Software Foundation (ASF) under one
 *  or more contributor license agreements.  See the NOTICE file
 *  distributed with this work for additional information
 *  regarding copyright ownership.  The ASF licenses this file
 *  to you under the Apache License, Version 2.0 (the
 *  "License"); you may not use this file except in compliance
 *  with the License.  You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 *  Unless required by applicable law or agreed to in writing, software
 *  distributed under the License is distributed on an "AS IS" BASIS,
 *  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 *  See the License for the specific language governing permissions and
 *  limitations under the License.
 */
package org.apache.hadoop.hive.ql.parse;

import java.io.Serializable;
import java.util.ArrayList;
import java.util.Deque;
import java.util.HashMap;
import java.util.HashSet;
import java.util.LinkedHashMap;
import java.util.LinkedHashSet;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import java.util.Set;
import java.util.Stack;
import java.util.concurrent.atomic.AtomicInteger;

import org.apache.hadoop.hive.conf.HiveConf;
import org.apache.hadoop.hive.conf.HiveConf.ConfVars;
import org.apache.hadoop.hive.ql.Context;
import org.apache.hadoop.hive.ql.QueryState;
import org.apache.hadoop.hive.ql.exec.AppMasterEventOperator;
import org.apache.hadoop.hive.ql.exec.CommonMergeJoinOperator;
import org.apache.hadoop.hive.ql.exec.ConditionalTask;
import org.apache.hadoop.hive.ql.exec.DummyStoreOperator;
import org.apache.hadoop.hive.ql.exec.FileSinkOperator;
import org.apache.hadoop.hive.ql.exec.FilterOperator;
import org.apache.hadoop.hive.ql.exec.GroupByOperator;
import org.apache.hadoop.hive.ql.exec.JoinOperator;
import org.apache.hadoop.hive.ql.exec.MapJoinOperator;
import org.apache.hadoop.hive.ql.exec.Operator;
import org.apache.hadoop.hive.ql.exec.OperatorUtils;
import org.apache.hadoop.hive.ql.exec.ReduceSinkOperator;
import org.apache.hadoop.hive.ql.exec.SelectOperator;
import org.apache.hadoop.hive.ql.exec.TableScanOperator;
import org.apache.hadoop.hive.ql.exec.Task;
import org.apache.hadoop.hive.ql.exec.TezDummyStoreOperator;
import org.apache.hadoop.hive.ql.exec.UnionOperator;
import org.apache.hadoop.hive.ql.exec.tez.TezTask;
import org.apache.hadoop.hive.ql.hooks.ReadEntity;
import org.apache.hadoop.hive.ql.hooks.WriteEntity;
import org.apache.hadoop.hive.ql.lib.CompositeProcessor;
import org.apache.hadoop.hive.ql.lib.DefaultRuleDispatcher;
import org.apache.hadoop.hive.ql.lib.Dispatcher;
import org.apache.hadoop.hive.ql.lib.ForwardWalker;
import org.apache.hadoop.hive.ql.lib.GraphWalker;
import org.apache.hadoop.hive.ql.lib.Node;
import org.apache.hadoop.hive.ql.lib.NodeProcessor;
import org.apache.hadoop.hive.ql.lib.NodeProcessorCtx;
import org.apache.hadoop.hive.ql.lib.PreOrderOnceWalker;
import org.apache.hadoop.hive.ql.lib.Rule;
import org.apache.hadoop.hive.ql.lib.RuleRegExp;
import org.apache.hadoop.hive.ql.log.PerfLogger;
import org.apache.hadoop.hive.ql.metadata.Hive;
import org.apache.hadoop.hive.ql.optimizer.ConstantPropagate;
import org.apache.hadoop.hive.ql.optimizer.ConstantPropagateProcCtx.ConstantPropagateOption;
import org.apache.hadoop.hive.ql.optimizer.ConvertJoinMapJoin;
import org.apache.hadoop.hive.ql.optimizer.DynamicPartitionPruningOptimization;
import org.apache.hadoop.hive.ql.optimizer.MergeJoinProc;
import org.apache.hadoop.hive.ql.optimizer.ReduceSinkMapJoinProc;
import org.apache.hadoop.hive.ql.optimizer.RemoveDynamicPruningBySize;
import org.apache.hadoop.hive.ql.optimizer.SetReducerParallelism;
import org.apache.hadoop.hive.ql.optimizer.SharedWorkOptimizer;
import org.apache.hadoop.hive.ql.optimizer.correlation.ReduceSinkJoinDeDuplication;
import org.apache.hadoop.hive.ql.optimizer.metainfo.annotation.AnnotateWithOpTraits;
import org.apache.hadoop.hive.ql.optimizer.physical.AnnotateRunTimeStatsOptimizer;
import org.apache.hadoop.hive.ql.optimizer.physical.CrossProductHandler;
import org.apache.hadoop.hive.ql.optimizer.physical.LlapClusterStateForCompile;
import org.apache.hadoop.hive.ql.optimizer.physical.LlapDecider;
import org.apache.hadoop.hive.ql.optimizer.physical.LlapPreVectorizationPass;
import org.apache.hadoop.hive.ql.optimizer.physical.MemoryDecider;
import org.apache.hadoop.hive.ql.optimizer.physical.MetadataOnlyOptimizer;
import org.apache.hadoop.hive.ql.optimizer.physical.NullScanOptimizer;
import org.apache.hadoop.hive.ql.optimizer.physical.PhysicalContext;
import org.apache.hadoop.hive.ql.optimizer.physical.SerializeFilter;
import org.apache.hadoop.hive.ql.optimizer.physical.StageIDsRearranger;
import org.apache.hadoop.hive.ql.optimizer.physical.Vectorizer;
import org.apache.hadoop.hive.ql.optimizer.stats.annotation.AnnotateWithStatistics;
import org.apache.hadoop.hive.ql.plan.AggregationDesc;
import org.apache.hadoop.hive.ql.plan.BaseWork;
import org.apache.hadoop.hive.ql.plan.ColStatistics;
import org.apache.hadoop.hive.ql.plan.DynamicPruningEventDesc;
import org.apache.hadoop.hive.ql.plan.ExprNodeColumnDesc;
import org.apache.hadoop.hive.ql.plan.ExprNodeDesc;
import org.apache.hadoop.hive.ql.plan.ExprNodeDescUtils;
import org.apache.hadoop.hive.ql.plan.GroupByDesc;
import org.apache.hadoop.hive.ql.plan.MapWork;
import org.apache.hadoop.hive.ql.plan.MoveWork;
import org.apache.hadoop.hive.ql.plan.OperatorDesc;
import org.apache.hadoop.hive.ql.plan.Statistics;
import org.apache.hadoop.hive.ql.plan.TezWork;
import org.apache.hadoop.hive.ql.session.SessionState;
import org.apache.hadoop.hive.ql.session.SessionState.LogHelper;
import org.apache.hadoop.hive.ql.stats.StatsUtils;
import org.apache.hadoop.hive.ql.udf.generic.GenericUDAFBloomFilter.GenericUDAFBloomFilterEvaluator;
import com.facebook.presto.hive.$internal.org.slf4j.Logger;
import com.facebook.presto.hive.$internal.org.slf4j.LoggerFactory;

/**
 * TezCompiler translates the operator plan into TezTasks.
 */
public class TezCompiler extends TaskCompiler {

  protected static final Logger LOG = LoggerFactory.getLogger(TezCompiler.class);

  public TezCompiler() {
  }

  @Override
  public void init(QueryState queryState, LogHelper console, Hive db) {
    super.init(queryState, console, db);

    // Tez requires us to use RPC for the query plan
    HiveConf.setBoolVar(conf, ConfVars.HIVE_RPC_QUERY_PLAN, true);

    // We require the use of recursive input dirs for union processing
    conf.setBoolean("mapred.input.dir.recursive", true);
  }

  @Override
  protected void optimizeOperatorPlan(ParseContext pCtx, Set inputs,
      Set outputs) throws SemanticException {
    PerfLogger perfLogger = SessionState.getPerfLogger();
    // Create the context for the walker
    OptimizeTezProcContext procCtx = new OptimizeTezProcContext(conf, pCtx, inputs, outputs);

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // setup dynamic partition pruning where possible
    runDynamicPartitionPruning(procCtx, inputs, outputs);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Setup dynamic partition pruning");

    // need to run this; to get consistent filterop conditions(for operator tree matching)
    if (procCtx.conf.getBoolVar(ConfVars.HIVEOPTCONSTANTPROPAGATION)) {
      new ConstantPropagate(ConstantPropagateOption.SHORTCUT).transform(procCtx.parseContext);
    }

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // setup stats in the operator plan
    runStatsAnnotation(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Setup stats in the operator plan");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // run the optimizations that use stats for optimization
    runStatsDependentOptimizations(procCtx, inputs, outputs);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Run the optimizations that use stats for optimization");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    if(procCtx.conf.getBoolVar(ConfVars.HIVEOPTJOINREDUCEDEDUPLICATION)) {
      new ReduceSinkJoinDeDuplication().transform(procCtx.parseContext);
    }
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Run reduce sink after join algorithm selection");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    runRemoveDynamicPruningOptimization(procCtx, inputs, outputs);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Run remove dynamic pruning by size");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    markSemiJoinForDPP(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Mark certain semijoin edges important based ");

    // Removing semijoin optimization when it may not be beneficial
    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    removeSemijoinOptimizationByBenefit(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Remove Semijoins based on cost benefits");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // Remove any parallel edge between semijoin and mapjoin.
    removeSemijoinsParallelToMapJoin(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Run the optimizations that use stats for optimization");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // Remove semijoin optimization if it creates a cycle with mapside joins
    removeSemiJoinCyclesDueToMapsideJoins(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Remove semijoin optimizations if it creates a cycle with mapside join");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // Remove semijoin optimization if SMB join is created.
    removeSemijoinOptimizationFromSMBJoins(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Remove semijoin optimizations if needed");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // Remove bloomfilter if no stats generated
    removeSemiJoinIfNoStats(procCtx);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Remove bloom filter optimizations if needed");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    // after the stats phase we might have some cyclic dependencies that we need
    // to take care of.
    runCycleAnalysisForPartitionPruning(procCtx, inputs, outputs);
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Run cycle analysis for partition pruning");

    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    if(procCtx.conf.getBoolVar(ConfVars.HIVE_SHARED_WORK_OPTIMIZATION)) {
      new SharedWorkOptimizer().transform(procCtx.parseContext);
    }
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "Shared scans optimization");

    // need a new run of the constant folding because we might have created lots
    // of "and true and true" conditions.
    // Rather than run the full constant folding just need to shortcut AND/OR expressions
    // involving constant true/false values.
    if(procCtx.conf.getBoolVar(ConfVars.HIVEOPTCONSTANTPROPAGATION)) {
      new ConstantPropagate(ConstantPropagateOption.SHORTCUT).transform(procCtx.parseContext);
    }

  }

  private void runCycleAnalysisForPartitionPruning(OptimizeTezProcContext procCtx,
      Set inputs, Set outputs) throws SemanticException {

    if (!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_PARTITION_PRUNING)) {
      return;
    }

    boolean cycleFree = false;
    while (!cycleFree) {
      cycleFree = true;
      Set>> components = getComponents(procCtx);
      for (Set> component : components) {
        if (LOG.isDebugEnabled()) {
          LOG.debug("Component: ");
          for (Operator co : component) {
            LOG.debug("Operator: " + co.getName() + ", " + co.getIdentifier());
          }
        }
        if (component.size() != 1) {
          LOG.info("Found cycle in operator plan...");
          cycleFree = false;
          removeCycleOperator(component, procCtx);
          break;
        }
      }
      LOG.info("Cycle free: " + cycleFree);
    }
  }

  private void removeCycleOperator(Set> component, OptimizeTezProcContext context) throws SemanticException {
    AppMasterEventOperator victimAM = null;
    TableScanOperator victimTS = null;
    ReduceSinkOperator victimRS = null;

    // If there is a hint and no operator is removed then throw error
    boolean hasHint = false;
    boolean removed = false;
    for (Operator o : component) {
      // Look for AppMasterEventOperator or ReduceSinkOperator
      if (o instanceof AppMasterEventOperator) {
        if (victimAM == null
                || o.getStatistics().getDataSize() < victimAM.getStatistics()
                .getDataSize()) {
          victimAM = (AppMasterEventOperator) o;
          removed = true;
        }
      } else if (o instanceof ReduceSinkOperator) {

        SemiJoinBranchInfo sjInfo =
                context.parseContext.getRsToSemiJoinBranchInfo().get(o);
        if (sjInfo == null ) {
          continue;
        }
        if (sjInfo.getIsHint()) {
          // Skipping because of hint. Mark this info,
          hasHint = true;
          continue;
        }

        TableScanOperator ts = sjInfo.getTsOp();
        // Sanity check
        assert component.contains(ts);

        if (victimRS == null ||
                ts.getStatistics().getDataSize() <
                        victimTS.getStatistics().getDataSize()) {
          victimRS = (ReduceSinkOperator) o;
          victimTS = ts;
          removed = true;
        }
      }
    }

    // Always set the semijoin optimization as victim.
    Operator victim = victimRS;

    if (victimRS == null && victimAM != null ) {
        victim = victimAM;
    } else if (victimAM == null) {
      // do nothing
    } else {
      // Cycle consists of atleast one dynamic partition pruning(DPP)
      // optimization and atleast one min/max optimization.
      // DPP is a better optimization unless it ends up scanning the
      // bigger table for keys instead of the smaller table.

      // Get the parent TS of victimRS.
      Operator op = victimRS;
      while(!(op instanceof TableScanOperator)) {
        op = op.getParentOperators().get(0);
      }
      if ((2 * op.getStatistics().getDataSize()) <
              victimAM.getStatistics().getDataSize()) {
        victim = victimAM;
      }
    }

    if (hasHint && !removed) {
      // There is hint but none of the operators removed. Throw error
      throw new SemanticException("The user hint is causing an operator cycle. Please fix it and retry");
    }

    if (victim == null ||
            (!context.pruningOpsRemovedByPriorOpt.isEmpty() &&
                    context.pruningOpsRemovedByPriorOpt.contains(victim))) {
      return;
    }

    GenTezUtils.removeBranch(victim);

    if (victim == victimRS) {
      if (LOG.isDebugEnabled()) {
        LOG.debug("Cycle found. Removing semijoin "
            + OperatorUtils.getOpNamePretty(victimRS) + " - " + OperatorUtils.getOpNamePretty(victimTS));
      }
      GenTezUtils.removeSemiJoinOperator(context.parseContext, victimRS, victimTS);
    } else {
      // at this point we've found the fork in the op pipeline that has the pruning as a child plan.
      LOG.info("Disabling dynamic pruning for: "
          + ((DynamicPruningEventDesc) victim.getConf()).getTableScan().toString()
          + ". Needed to break cyclic dependency");
    }
    return;
  }

  // Tarjan's algo
  private Set>> getComponents(OptimizeTezProcContext procCtx) {
    Deque> deque = new LinkedList>();
    deque.addAll(procCtx.parseContext.getTopOps().values());

    AtomicInteger index = new AtomicInteger();
    Map, Integer> indexes = new HashMap, Integer>();
    Map, Integer> lowLinks = new HashMap, Integer>();
    Stack> nodes = new Stack>();
    Set>> components = new LinkedHashSet>>();

    for (Operator o : deque) {
      if (!indexes.containsKey(o)) {
        connect(o, index, nodes, indexes, lowLinks, components, procCtx.parseContext);
      }
    }

    return components;
  }

  private void connect(Operator o, AtomicInteger index, Stack> nodes,
      Map, Integer> indexes, Map, Integer> lowLinks,
      Set>> components, ParseContext parseContext) {

    indexes.put(o, index.get());
    lowLinks.put(o, index.get());
    index.incrementAndGet();
    nodes.push(o);

    List> children;
    if (o instanceof AppMasterEventOperator) {
      children = new ArrayList>();
      children.addAll(o.getChildOperators());
      TableScanOperator ts = ((DynamicPruningEventDesc) o.getConf()).getTableScan();
      LOG.debug("Adding special edge: " + o.getName() + " --> " + ts.toString());
      children.add(ts);
    } else if (o instanceof ReduceSinkOperator){
      // semijoin case
      children = new ArrayList>();
      children.addAll(o.getChildOperators());
      SemiJoinBranchInfo sjInfo = parseContext.getRsToSemiJoinBranchInfo().get(o);
      if (sjInfo != null ) {
        TableScanOperator ts = sjInfo.getTsOp();
        LOG.debug("Adding special edge: " + o.getName() + " --> " + ts.toString());
        children.add(ts);
      }
    } else {
      children = o.getChildOperators();
    }

    for (Operator child : children) {
      if (!indexes.containsKey(child)) {
        connect(child, index, nodes, indexes, lowLinks, components, parseContext);
        lowLinks.put(o, Math.min(lowLinks.get(o), lowLinks.get(child)));
      } else if (nodes.contains(child)) {
        lowLinks.put(o, Math.min(lowLinks.get(o), indexes.get(child)));
      }
    }

    if (lowLinks.get(o).equals(indexes.get(o))) {
      Set> component = new LinkedHashSet>();
      components.add(component);
      Operator current;
      do {
        current = nodes.pop();
        component.add(current);
      } while (current != o);
    }
  }

  private void runStatsAnnotation(OptimizeTezProcContext procCtx) throws SemanticException {
    new AnnotateWithStatistics().transform(procCtx.parseContext);
    new AnnotateWithOpTraits().transform(procCtx.parseContext);
  }

  private void runStatsDependentOptimizations(OptimizeTezProcContext procCtx,
      Set inputs, Set outputs) throws SemanticException {

    // Sequence of TableScan operators to be walked
    Deque> deque = new LinkedList>();
    deque.addAll(procCtx.parseContext.getTopOps().values());

    // create a walker which walks the tree in a DFS manner while maintaining
    // the operator stack.
    Map opRules = new LinkedHashMap();
    opRules.put(new RuleRegExp("Set parallelism - ReduceSink",
        ReduceSinkOperator.getOperatorName() + "%"),
        new SetReducerParallelism());

    opRules.put(new RuleRegExp("Convert Join to Map-join",
        JoinOperator.getOperatorName() + "%"), new ConvertJoinMapJoin());

    // The dispatcher fires the processor corresponding to the closest matching
    // rule and passes the context along
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new ForwardWalker(disp);
    ogw.startWalking(topNodes, null);
  }

  private void runRemoveDynamicPruningOptimization(OptimizeTezProcContext procCtx,
      Set inputs, Set outputs) throws SemanticException {

    // Sequence of TableScan operators to be walked
    Deque> deque = new LinkedList>();
    deque.addAll(procCtx.parseContext.getTopOps().values());

    // create a walker which walks the tree in a DFS manner while maintaining
    // the operator stack.
    Map opRules = new LinkedHashMap();
    opRules.put(
        new RuleRegExp("Remove dynamic pruning by size",
        AppMasterEventOperator.getOperatorName() + "%"),
        new RemoveDynamicPruningBySize());

    // The dispatcher fires the processor corresponding to the closest matching
    // rule and passes the context along
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new ForwardWalker(disp);
    ogw.startWalking(topNodes, null);
  }

  private void runDynamicPartitionPruning(OptimizeTezProcContext procCtx, Set inputs,
      Set outputs) throws SemanticException {

    if (!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_PARTITION_PRUNING)) {
      return;
    }

    // Sequence of TableScan operators to be walked
    Deque> deque = new LinkedList>();
    deque.addAll(procCtx.parseContext.getTopOps().values());

    Map opRules = new LinkedHashMap();
    opRules.put(
        new RuleRegExp(new String("Dynamic Partition Pruning"), FilterOperator.getOperatorName()
            + "%"), new DynamicPartitionPruningOptimization());

    // The dispatcher fires the processor corresponding to the closest matching
    // rule and passes the context along
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new ForwardWalker(disp);
    ogw.startWalking(topNodes, null);
  }

  @Override
  protected void generateTaskTree(List> rootTasks, ParseContext pCtx,
      List> mvTask, Set inputs, Set outputs)
      throws SemanticException {

	PerfLogger perfLogger = SessionState.getPerfLogger();
    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    ParseContext tempParseContext = getParseContext(pCtx, rootTasks);
    GenTezUtils utils = new GenTezUtils();
    GenTezWork genTezWork = new GenTezWork(utils);

    GenTezProcContext procCtx = new GenTezProcContext(
        conf, tempParseContext, mvTask, rootTasks, inputs, outputs);

    // create a walker which walks the tree in a DFS manner while maintaining
    // the operator stack.
    // The dispatcher generates the plan from the operator tree
    Map opRules = new LinkedHashMap();
    opRules.put(new RuleRegExp("Split Work - ReduceSink",
        ReduceSinkOperator.getOperatorName() + "%"),
        genTezWork);

    opRules.put(new RuleRegExp("No more walking on ReduceSink-MapJoin",
        MapJoinOperator.getOperatorName() + "%"), new ReduceSinkMapJoinProc());

    opRules.put(new RuleRegExp("Recognize a Sorted Merge Join operator to setup the right edge and"
        + " stop traversing the DummyStore-MapJoin", CommonMergeJoinOperator.getOperatorName()
        + "%"), new MergeJoinProc());

    opRules.put(new RuleRegExp("Split Work + Move/Merge - FileSink",
        FileSinkOperator.getOperatorName() + "%"),
        new CompositeProcessor(new FileSinkProcessor(), genTezWork));

    opRules.put(new RuleRegExp("Split work - DummyStore", DummyStoreOperator.getOperatorName()
        + "%"), genTezWork);

    opRules.put(new RuleRegExp("Handle Potential Analyze Command",
        TableScanOperator.getOperatorName() + "%"),
        new ProcessAnalyzeTable(utils));

    opRules.put(new RuleRegExp("Remember union",
        UnionOperator.getOperatorName() + "%"),
        new UnionProcessor());

    opRules.put(new RuleRegExp("AppMasterEventOperator",
        AppMasterEventOperator.getOperatorName() + "%"),
        new AppMasterEventProcessor());

    // The dispatcher fires the processor corresponding to the closest matching
    // rule and passes the context along
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    List topNodes = new ArrayList();
    topNodes.addAll(pCtx.getTopOps().values());
    GraphWalker ogw = new GenTezWorkWalker(disp, procCtx);
    ogw.startWalking(topNodes, null);

    // we need to specify the reserved memory for each work that contains Map Join
    for (List baseWorkList : procCtx.mapJoinWorkMap.values()) {
      for (BaseWork w : baseWorkList) {
        // work should be the smallest unit for memory allocation
        w.setReservedMemoryMB(
            (int)(conf.getLongVar(ConfVars.HIVECONVERTJOINNOCONDITIONALTASKTHRESHOLD) / (1024 * 1024)));
      }
    }

    // we need to clone some operator plans and remove union operators still
    int indexForTezUnion = 0;
    for (BaseWork w: procCtx.workWithUnionOperators) {
      GenTezUtils.removeUnionOperators(procCtx, w, indexForTezUnion++);
    }

    // then we make sure the file sink operators are set up right
    for (FileSinkOperator fileSink: procCtx.fileSinkSet) {
      GenTezUtils.processFileSink(procCtx, fileSink);
    }

    // Connect any edges required for min/max pushdown
    if (pCtx.getRsToRuntimeValuesInfoMap().size() > 0) {
      for (ReduceSinkOperator rs : pCtx.getRsToRuntimeValuesInfoMap().keySet()) {
        // Process min/max
        GenTezUtils.processDynamicSemiJoinPushDownOperator(
                procCtx, pCtx.getRsToRuntimeValuesInfoMap().get(rs), rs);
      }
    }
    // and finally we hook up any events that need to be sent to the tez AM
    LOG.debug("There are " + procCtx.eventOperatorSet.size() + " app master events.");
    for (AppMasterEventOperator event : procCtx.eventOperatorSet) {
      LOG.debug("Handling AppMasterEventOperator: " + event);
      GenTezUtils.processAppMasterEvent(procCtx, event);
    }
    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "generateTaskTree");
  }

  @Override
  protected void setInputFormat(Task task) {
    if (task instanceof TezTask) {
      TezWork work = ((TezTask)task).getWork();
      List all = work.getAllWork();
      for (BaseWork w: all) {
        if (w instanceof MapWork) {
          MapWork mapWork = (MapWork) w;
          HashMap> opMap = mapWork.getAliasToWork();
          if (!opMap.isEmpty()) {
            for (Operator op : opMap.values()) {
              setInputFormat(mapWork, op);
            }
          }
        }
      }
    } else if (task instanceof ConditionalTask) {
      List> listTasks
        = ((ConditionalTask) task).getListTasks();
      for (Task tsk : listTasks) {
        setInputFormat(tsk);
      }
    }

    if (task.getChildTasks() != null) {
      for (Task childTask : task.getChildTasks()) {
        setInputFormat(childTask);
      }
    }
  }

  private void setInputFormat(MapWork work, Operator op) {
    if (op == null) {
      return;
    }
    if (op.isUseBucketizedHiveInputFormat()) {
      work.setUseBucketizedHiveInputFormat(true);
      return;
    }

    if (op.getChildOperators() != null) {
      for (Operator childOp : op.getChildOperators()) {
        setInputFormat(work, childOp);
      }
    }
  }

  @Override
  protected void decideExecMode(List> rootTasks, Context ctx,
      GlobalLimitCtx globalLimitCtx)
      throws SemanticException {
    // currently all Tez work is on the cluster
    return;
  }

  @Override
  protected void optimizeTaskPlan(List> rootTasks, ParseContext pCtx,
      Context ctx) throws SemanticException {
    PerfLogger perfLogger = SessionState.getPerfLogger();
    perfLogger.PerfLogBegin(this.getClass().getName(), PerfLogger.TEZ_COMPILER);
    PhysicalContext physicalCtx = new PhysicalContext(conf, pCtx, pCtx.getContext(), rootTasks,
       pCtx.getFetchTask());

    if (conf.getBoolVar(HiveConf.ConfVars.HIVENULLSCANOPTIMIZE)) {
      physicalCtx = new NullScanOptimizer().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping null scan query optimization");
    }

    if (conf.getBoolVar(HiveConf.ConfVars.HIVEMETADATAONLYQUERIES)) {
      physicalCtx = new MetadataOnlyOptimizer().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping metadata only query optimization");
    }

    if (conf.getBoolVar(HiveConf.ConfVars.HIVE_CHECK_CROSS_PRODUCT)) {
      physicalCtx = new CrossProductHandler().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping cross product analysis");
    }

    if ("llap".equalsIgnoreCase(conf.getVar(HiveConf.ConfVars.HIVE_EXECUTION_MODE))) {
      physicalCtx = new LlapPreVectorizationPass().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping llap pre-vectorization pass");
    }

    if (conf.getBoolVar(HiveConf.ConfVars.HIVE_VECTORIZATION_ENABLED)) {
      physicalCtx = new Vectorizer().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping vectorization");
    }

    if (!"none".equalsIgnoreCase(conf.getVar(HiveConf.ConfVars.HIVESTAGEIDREARRANGE))) {
      physicalCtx = new StageIDsRearranger().resolve(physicalCtx);
    } else {
      LOG.debug("Skipping stage id rearranger");
    }

    if ((conf.getBoolVar(HiveConf.ConfVars.HIVE_TEZ_ENABLE_MEMORY_MANAGER))
        && (conf.getBoolVar(HiveConf.ConfVars.HIVEUSEHYBRIDGRACEHASHJOIN))) {
      physicalCtx = new MemoryDecider().resolve(physicalCtx);
    }

    if ("llap".equalsIgnoreCase(conf.getVar(HiveConf.ConfVars.HIVE_EXECUTION_MODE))) {
      LlapClusterStateForCompile llapInfo = LlapClusterStateForCompile.getClusterInfo(conf);
      physicalCtx = new LlapDecider(llapInfo).resolve(physicalCtx);
    } else {
      LOG.debug("Skipping llap decider");
    }

    //  This optimizer will serialize all filters that made it to the
    //  table scan operator to avoid having to do it multiple times on
    //  the backend. If you have a physical optimization that changes
    //  table scans or filters, you have to invoke it before this one.
    physicalCtx = new SerializeFilter().resolve(physicalCtx);

    if (physicalCtx.getContext().getExplainAnalyze() != null) {
      new AnnotateRunTimeStatsOptimizer().resolve(physicalCtx);
    }

    perfLogger.PerfLogEnd(this.getClass().getName(), PerfLogger.TEZ_COMPILER, "optimizeTaskPlan");
    return;
  }

  private static class SMBJoinOpProcContext implements NodeProcessorCtx {
    HashMap JoinOpToTsOpMap = new HashMap();
  }

  private static class SMBJoinOpProc implements NodeProcessor {

    @Override
    public Object process(Node nd, Stack stack, NodeProcessorCtx procCtx,
                          Object... nodeOutputs) throws SemanticException {
      SMBJoinOpProcContext ctx = (SMBJoinOpProcContext) procCtx;
      ctx.JoinOpToTsOpMap.put((CommonMergeJoinOperator) nd,
              (TableScanOperator) stack.get(0));
      return null;
    }
  }

  private static void removeSemijoinOptimizationFromSMBJoins(
          OptimizeTezProcContext procCtx) throws SemanticException {
    if (!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION) ||
            procCtx.parseContext.getRsToSemiJoinBranchInfo().size() == 0) {
      return;
    }

    Map opRules = new LinkedHashMap();
    opRules.put(
            new RuleRegExp("R1", TableScanOperator.getOperatorName() + "%" +
                    ".*" + TezDummyStoreOperator.getOperatorName() + "%" +
                    CommonMergeJoinOperator.getOperatorName() + "%"),
            new SMBJoinOpProc());

    SMBJoinOpProcContext ctx = new SMBJoinOpProcContext();
    // The dispatcher finds SMB and if there is semijoin optimization before it, removes it.
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, ctx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new PreOrderOnceWalker(disp);
    ogw.startWalking(topNodes, null);

    List tsOps = new ArrayList<>();
    // Iterate over the map and remove semijoin optimizations if needed.
    for (CommonMergeJoinOperator joinOp : ctx.JoinOpToTsOpMap.keySet()) {
      // Get one top level TS Op directly from the stack
      tsOps.add(ctx.JoinOpToTsOpMap.get(joinOp));

      // Get the other one by examining Join Op
      List> parents = joinOp.getParentOperators();
      for (Operator parent : parents) {
        if (parent instanceof TezDummyStoreOperator) {
          // already accounted for
          continue;
        }

        while (parent != null) {
          if (parent instanceof TableScanOperator) {
            tsOps.add((TableScanOperator) parent);
            break;
          }
          parent = parent.getParentOperators().get(0);
        }
      }
    }
    // Now the relevant TableScanOperators are known, find if there exists
    // a semijoin filter on any of them, if so, remove it.

    ParseContext pctx = procCtx.parseContext;
    Set rsSet = new HashSet<>(pctx.getRsToSemiJoinBranchInfo().keySet());
    for (TableScanOperator ts : tsOps) {
      for (ReduceSinkOperator rs : rsSet) {
        SemiJoinBranchInfo sjInfo = pctx.getRsToSemiJoinBranchInfo().get(rs);
        if (sjInfo != null && ts == sjInfo.getTsOp()) {
          // match!
          if (sjInfo.getIsHint()) {
            throw new SemanticException("Removing hinted semijoin as it is with SMB join " + rs + " : " + ts);
          }
          if (LOG.isDebugEnabled()) {
            LOG.debug("Semijoin optimization found going to SMB join. Removing semijoin "
                    + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(ts));
          }
          GenTezUtils.removeBranch(rs);
          GenTezUtils.removeSemiJoinOperator(pctx, rs, ts);
        }
      }
    }
  }

  private static class SemiJoinCycleRemovalDueTOMapsideJoinContext implements NodeProcessorCtx {
    HashMap,Operator> childParentMap = new HashMap,Operator>();
  }

  private static class SemiJoinCycleRemovalDueToMapsideJoins implements NodeProcessor {

    @Override
    public Object process(Node nd, Stack stack, NodeProcessorCtx procCtx,
                          Object... nodeOutputs) throws SemanticException {

      SemiJoinCycleRemovalDueTOMapsideJoinContext ctx =
              (SemiJoinCycleRemovalDueTOMapsideJoinContext) procCtx;
      ctx.childParentMap.put((Operator)stack.get(stack.size() - 2), (Operator) nd);
      return null;
    }
  }

  private static void removeSemiJoinCyclesDueToMapsideJoins(
          OptimizeTezProcContext procCtx) throws SemanticException {
    if (!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION) ||
            procCtx.parseContext.getRsToSemiJoinBranchInfo().size() == 0) {
      return;
    }

    Map opRules = new LinkedHashMap();
    opRules.put(
            new RuleRegExp("R1", MapJoinOperator.getOperatorName() + "%" +
                    MapJoinOperator.getOperatorName() + "%"),
            new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(
            new RuleRegExp("R2", MapJoinOperator.getOperatorName() + "%" +
                    CommonMergeJoinOperator.getOperatorName() + "%"),
            new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(
            new RuleRegExp("R3", CommonMergeJoinOperator.getOperatorName() + "%" +
                    MapJoinOperator.getOperatorName() + "%"),
            new SemiJoinCycleRemovalDueToMapsideJoins());
    opRules.put(
            new RuleRegExp("R4", CommonMergeJoinOperator.getOperatorName() + "%" +
                    CommonMergeJoinOperator.getOperatorName() + "%"),
            new SemiJoinCycleRemovalDueToMapsideJoins());

    SemiJoinCycleRemovalDueTOMapsideJoinContext ctx =
            new SemiJoinCycleRemovalDueTOMapsideJoinContext();
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, ctx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new PreOrderOnceWalker(disp);
    ogw.startWalking(topNodes, null);

    // process the list
    ParseContext pCtx = procCtx.parseContext;
    for (Operator parentJoin : ctx.childParentMap.keySet()) {
      Operator childJoin = ctx.childParentMap.get(parentJoin);

      if (parentJoin.getChildOperators().size() == 1) {
        continue;
      }

      for (Operator child : parentJoin.getChildOperators()) {
        if (!(child instanceof SelectOperator)) {
          continue;
        }

        while(child.getChildOperators().size() > 0) {
          child = child.getChildOperators().get(0);
        }

        if (!(child instanceof ReduceSinkOperator)) {
          continue;
        }

        ReduceSinkOperator rs = ((ReduceSinkOperator) child);
        SemiJoinBranchInfo sjInfo = pCtx.getRsToSemiJoinBranchInfo().get(rs);
        if (sjInfo == null) {
          continue;
        }

        TableScanOperator ts = sjInfo.getTsOp();
        // This is a semijoin branch. Find if this is creating a potential
        // cycle with childJoin.
        for (Operator parent : childJoin.getParentOperators()) {
          if (parent == parentJoin) {
            continue;
          }

          assert parent instanceof ReduceSinkOperator;
          while (parent.getParentOperators().size() > 0) {
            parent = parent.getParentOperators().get(0);
          }

          if (parent == ts) {
            // We have a cycle!
            if (sjInfo.getIsHint()) {
              throw new SemanticException("Removing hinted semijoin as it is creating cycles with mapside joins " + rs + " : " + ts);
            }
            if (LOG.isDebugEnabled()) {
              LOG.debug("Semijoin cycle due to mapjoin. Removing semijoin "
                  + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(ts));
            }
            GenTezUtils.removeBranch(rs);
            GenTezUtils.removeSemiJoinOperator(pCtx, rs, ts);
          }
        }
      }
    }
  }

  private static class SemiJoinRemovalIfNoStatsProc implements NodeProcessor {

    @Override
    public Object process(Node nd, Stack stack, NodeProcessorCtx procCtx,
                          Object... nodeOutputs) throws SemanticException {
      assert nd instanceof ReduceSinkOperator;
      ReduceSinkOperator rs = (ReduceSinkOperator) nd;
      ParseContext pCtx = ((OptimizeTezProcContext) procCtx).parseContext;
      SemiJoinBranchInfo sjInfo = pCtx.getRsToSemiJoinBranchInfo().get(rs);
      if (sjInfo == null) {
        // nothing to do here.
        return null;
      }

      // This is a semijoin branch. The stack should look like,
      // -SEL-GB1-RS1-GB2-RS2
      GroupByOperator gbOp = (GroupByOperator) (stack.get(stack.size() - 2));
      GroupByDesc gbDesc = gbOp.getConf();
      ArrayList aggregationDescs = gbDesc.getAggregators();
      for (AggregationDesc agg : aggregationDescs) {
        if (!"bloom_filter".equals(agg.getGenericUDAFName())) {
          continue;
        }

        GenericUDAFBloomFilterEvaluator udafBloomFilterEvaluator =
                (GenericUDAFBloomFilterEvaluator) agg.getGenericUDAFEvaluator();
        if (udafBloomFilterEvaluator.hasHintEntries())
         {
          return null; // Created using hint, skip it
        }

        long expectedEntries = udafBloomFilterEvaluator.getExpectedEntries();
        if (expectedEntries == -1 || expectedEntries >
                pCtx.getConf().getLongVar(ConfVars.TEZ_MAX_BLOOM_FILTER_ENTRIES)) {
          if (sjInfo.getIsHint()) {
            throw new SemanticException("Removing hinted semijoin due to lack to stats" +
            " or exceeding max bloom filter entries");
          }
          // Remove the semijoin optimization branch along with ALL the mappings
          // The parent GB2 has all the branches. Collect them and remove them.
          for (Node node : gbOp.getChildren()) {
            ReduceSinkOperator rsFinal = (ReduceSinkOperator) node;
            TableScanOperator ts = pCtx.getRsToSemiJoinBranchInfo().
                    get(rsFinal).getTsOp();
            if (LOG.isDebugEnabled()) {
              LOG.debug("expectedEntries=" + expectedEntries + ". "
                      + "Either stats unavailable or expectedEntries exceeded max allowable bloomfilter size. "
                      + "Removing semijoin "
                      + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(ts));
            }
            GenTezUtils.removeBranch(rsFinal);
            GenTezUtils.removeSemiJoinOperator(pCtx, rsFinal, ts);
          }
          return null;
        }
      }

      // At this point, hinted semijoin case has been handled already
      // Check if big table is big enough that runtime filtering is
      // worth it.
      TableScanOperator ts = sjInfo.getTsOp();
      if (ts.getStatistics() != null) {
        long numRows = ts.getStatistics().getNumRows();
        if (numRows < pCtx.getConf().getLongVar(ConfVars.TEZ_BIGTABLE_MIN_SIZE_SEMIJOIN_REDUCTION)) {
          if (sjInfo.getShouldRemove()) {
            if (LOG.isDebugEnabled()) {
              LOG.debug("Insufficient rows (" + numRows + ") to justify semijoin optimization. Removing semijoin "
                      + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(ts));
            }
            GenTezUtils.removeBranch(rs);
            GenTezUtils.removeSemiJoinOperator(pCtx, rs, ts);
          }
        }
      }
      return null;
    }
  }

  private void removeSemiJoinIfNoStats(OptimizeTezProcContext procCtx)
          throws SemanticException {
    if(!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION)) {
      // Not needed without semi-join reduction
      return;
    }

    Map opRules = new LinkedHashMap();
    opRules.put(
            new RuleRegExp("R1", GroupByOperator.getOperatorName() + "%" +
                    ReduceSinkOperator.getOperatorName() + "%" +
                    GroupByOperator.getOperatorName() + "%" +
                    ReduceSinkOperator.getOperatorName() + "%"),
            new SemiJoinRemovalIfNoStatsProc());
    Dispatcher disp = new DefaultRuleDispatcher(null, opRules, procCtx);
    List topNodes = new ArrayList();
    topNodes.addAll(procCtx.parseContext.getTopOps().values());
    GraphWalker ogw = new PreOrderOnceWalker(disp);
    ogw.startWalking(topNodes, null);
  }

  private boolean findParallelSemiJoinBranch(Operator mapjoin, TableScanOperator bigTableTS,
                                             ParseContext parseContext,
                                             Map semijoins) {

    boolean parallelEdges = false;
    for (Operator op : mapjoin.getParentOperators()) {
      if (!(op instanceof ReduceSinkOperator)) {
        continue;
      }

      op = op.getParentOperators().get(0);

      // Follow the Reducesink operator upstream which is on small table side.
      while (!(op instanceof ReduceSinkOperator) &&
              !(op instanceof TableScanOperator) &&
              !(op.getChildren() != null && op.getChildren().size() > 1)) {
        if (op instanceof MapJoinOperator) {
          // Pick the correct parent, only one of the parents is not
          // ReduceSink, that is what we are looking for.
          for (Operator parentOp : op.getParentOperators()) {
            if (parentOp instanceof ReduceSinkOperator) {
              continue;
            }
            op = parentOp; // parent in current pipeline
            continue;
          }
        }
        op = op.getParentOperators().get(0);
      }

      // Bail out if RS or TS is encountered.
      if (op instanceof ReduceSinkOperator || op instanceof TableScanOperator) {
        continue;
      }

      // A branch is hit.
      for (Node nd : op.getChildren()) {
        if (nd instanceof SelectOperator) {
          Operator child = (Operator) nd;

          while (child.getChildOperators().size() > 0) {
            child = child.getChildOperators().get(0);
          }

          // If not ReduceSink Op, skip
          if (!(child instanceof ReduceSinkOperator)) {
            // This still could be DPP.
            if (child instanceof AppMasterEventOperator &&
                    ((AppMasterEventOperator) child).getConf() instanceof DynamicPruningEventDesc) {
              // DPP indeed, Set parallel edges true
              parallelEdges = true;
            }
            continue;
          }

          ReduceSinkOperator rs = (ReduceSinkOperator) child;
          SemiJoinBranchInfo sjInfo = parseContext.getRsToSemiJoinBranchInfo().get(rs);
          if (sjInfo == null) {
            continue;
          }

          TableScanOperator ts = sjInfo.getTsOp();
          if (ts != bigTableTS) {
            // skip, not the one we are looking for.
            continue;
          }

          parallelEdges = true;

          if (sjInfo.getIsHint() || !sjInfo.getShouldRemove()) {
            // Created by hint, skip it
            continue;
          }
          // Add the semijoin branch to the map
          semijoins.put(rs, ts);
        }
      }
    }
    return parallelEdges;
  }

  /*
   *  The algorithm looks at all the mapjoins in the operator pipeline until
   *  it hits RS Op and for each mapjoin examines if it has paralllel semijoin
   *  edge or dynamic partition pruning.
   */
  private void removeSemijoinsParallelToMapJoin(OptimizeTezProcContext procCtx)
          throws SemanticException {
    if(!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION) ||
            !procCtx.conf.getBoolVar(ConfVars.HIVECONVERTJOIN) ||
            procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION_FOR_MAPJOIN)) {
      // Not needed without semi-join reduction or mapjoins or when semijoins
      // are enabled for parallel mapjoins.
      return;
    }

    // Get all the TS ops.
    List> topOps = new ArrayList<>();
    topOps.addAll(procCtx.parseContext.getTopOps().values());

    Map semijoins = new HashMap<>();
    for (Operator parent : topOps) {
      // A TS can have multiple branches due to DPP Or Semijoin Opt.
      // USe DFS to traverse all the branches until RS is hit.
      Deque> deque = new LinkedList<>();
      deque.add(parent);
      while (!deque.isEmpty()) {
        Operator op = deque.pollLast();
        if (op instanceof ReduceSinkOperator) {
          // Done with this branch
          continue;
        }

        if (op instanceof MapJoinOperator) {
          // A candidate.
          if (!findParallelSemiJoinBranch(op, (TableScanOperator) parent,
                  procCtx.parseContext, semijoins)) {
            // No parallel edge was found for the given mapjoin op,
            // no need to go down further, skip this TS operator pipeline.
            break;
          }
        }
        deque.addAll(op.getChildOperators());
      }
    }

    if (semijoins.size() > 0) {
      for (ReduceSinkOperator rs : semijoins.keySet()) {
        if (LOG.isDebugEnabled()) {
          LOG.debug("Semijoin optimization with parallel edge to map join. Removing semijoin "
              + OperatorUtils.getOpNamePretty(rs) + " - " + OperatorUtils.getOpNamePretty(semijoins.get(rs)));
        }
        GenTezUtils.removeBranch(rs);
        GenTezUtils.removeSemiJoinOperator(procCtx.parseContext, rs,
                semijoins.get(rs));
      }
    }
  }

  private static boolean canUseNDV(ColStatistics colStats) {
    return (colStats != null) && (colStats.getCountDistint() >= 0);
  }

  private static double getBloomFilterCost(
      SelectOperator sel,
      FilterOperator fil) {
    double cost = -1;
    Statistics selStats = sel.getStatistics();
    if (selStats != null) {
      cost = selStats.getNumRows();

      // Some other things that could be added here to model cost:
      // Cost of computing/sending partial BloomFilter results? BloomFilterSize * # mappers
      // For reduce-side join, add the cost of the semijoin table scan/dependent tablescans?
    }
    return cost;
  }

  private static long getCombinedKeyDomainCardinality(
      ColStatistics selColStat,
      ColStatistics selColSourceStat,
      ColStatistics tsColStat) {
    long keyDomainCardinality = -1;
    if (!canUseNDV(selColStat) || !canUseNDV(tsColStat)) {
      return -1;
    }

    long selColSourceNdv = canUseNDV(selColSourceStat) ? selColSourceStat.getCountDistint() : -1;
    boolean semiJoinKeyIsPK = StatsUtils.inferForeignKey(selColStat, tsColStat);
    if (semiJoinKeyIsPK) {
      // PK/FQ relationship: NDV of selColSourceStat is a superset of what is in tsColStat
      if (selColSourceNdv >= 0) {
        // Most accurate domain cardinality would be source column NDV if available.
        keyDomainCardinality = selColSourceNdv;
      }
    } else {
      if (selColSourceNdv >= 0) {
        // If semijoin keys and ts keys completely unrelated, the cardinality of both sets
        // could be obtained by adding both cardinalities. Would there be an average case?
        keyDomainCardinality = selColSourceNdv + tsColStat.getCountDistint();

        // Don't exceed the range if we have one.
        if (StatsUtils.hasDiscreteRange(selColStat)
            && StatsUtils.hasDiscreteRange(tsColStat)) {
          long range = 0;
          // Trying using the cardinality from the value range.
          ColStatistics.Range combinedRange = StatsUtils.combineRange(selColStat.getRange(), tsColStat.getRange());
          if (combinedRange != null) {
            range = StatsUtils.getRangeDelta(combinedRange);
          } else {
            range = StatsUtils.getRangeDelta(selColStat.getRange())
                + StatsUtils.getRangeDelta(tsColStat.getRange());
          }
          keyDomainCardinality = Math.min(keyDomainCardinality, range);
        }
      }
      // Otherwise, we tried ..
    }
    if (LOG.isDebugEnabled()) {
      LOG.debug("Computing key domain cardinality, keyDomainCardinality=" + keyDomainCardinality
          + ", semiJoinKeyIsPK=" + semiJoinKeyIsPK
          + ", selColStat=" + selColStat
          + ", selColSourceStat=" + selColSourceStat
          + ", tsColStat=" + tsColStat);
    }

    return keyDomainCardinality;
  }

  private static double getBloomFilterBenefit(
      SelectOperator sel, ExprNodeDesc selExpr,
      FilterOperator fil, ExprNodeDesc tsExpr) {
    double benefit = -1;
    Statistics selStats = sel.getStatistics();
    Statistics filStats = fil.getStatistics();
    if (selStats == null || filStats == null) {
      LOG.debug("No stats available to compute BloomFilter benefit");
      return benefit;
    }

    // For cardinality values use numRows as default, try to use ColStats if available
    long selKeyCardinality = selStats.getNumRows();
    long tsKeyCardinality = filStats.getNumRows();
    long tsRows = filStats.getNumRows();
    long tsRowSize = filStats.getAvgRowSize();
    long keyDomainCardinality = selKeyCardinality + tsKeyCardinality;

    ExprNodeColumnDesc selCol = ExprNodeDescUtils.getColumnExpr(selExpr);
    ExprNodeColumnDesc tsCol = ExprNodeDescUtils.getColumnExpr(tsExpr);
    if (selCol != null && tsCol != null) {
      // Check if there are column stats available for these columns
      ColStatistics selColStat = selStats.getColumnStatisticsFromColName(selCol.getColumn());
      ColStatistics filColStat = filStats.getColumnStatisticsFromColName(tsCol.getColumn());
      if (canUseNDV(selColStat)) {
        selKeyCardinality = selColStat.getCountDistint();
      }
      if (canUseNDV(filColStat)) {
        tsKeyCardinality = filColStat.getCountDistint();
      }
      // Get colstats for the original table column for selCol if possible, this would have
      // more accurate information about the original NDV of the column before any filtering.
      ColStatistics selColSourceStat = null;
      if (selColStat != null) {
        ExprNodeDescUtils.ColumnOrigin selColSource = ExprNodeDescUtils.findColumnOrigin(selCol, sel);
        if (selColSource != null && selColSource.op.getStatistics() != null) {
          selColSourceStat = selColSource.op.getStatistics().getColumnStatisticsFromColName(
              selColSource.col.getColumn());
        }
      }
      long domainCardinalityFromColStats = getCombinedKeyDomainCardinality(
          selColStat, selColSourceStat, filColStat);
      if (domainCardinalityFromColStats >= 0) {
        keyDomainCardinality = domainCardinalityFromColStats;
      }
    }

    // Selectivity: key cardinality of semijoin / domain cardinality
    // Benefit (rows filtered from ts): (1 - selectivity) * # ts rows
    double selectivity = selKeyCardinality / (double) keyDomainCardinality;
    selectivity = Math.min(selectivity, 1);
    benefit = tsRows * (1 - selectivity);

    if (LOG.isDebugEnabled()) {
      LOG.debug("BloomFilter benefit for " + selCol + " to " + tsCol
          + ", selKeyCardinality=" + selKeyCardinality
          + ", tsKeyCardinality=" + tsKeyCardinality
          + ", tsRows=" + tsRows
          + ", keyDomainCardinality=" + keyDomainCardinality);
      LOG.debug("SemiJoin key selectivity=" + selectivity
          + ", benefit=" + benefit);
    }

    return benefit;
  }

  private static double computeBloomFilterNetBenefit(
      SelectOperator sel, ExprNodeDesc selExpr,
      FilterOperator fil, ExprNodeDesc tsExpr) {
    double netBenefit = -1;
    double benefit = getBloomFilterBenefit(sel, selExpr, fil, tsExpr);
    Statistics filStats = fil.getStatistics();
    if (benefit > 0 && filStats != null) {
      double cost = getBloomFilterCost(sel, fil);
      if (cost > 0) {
        long filDataSize = filStats.getNumRows();
        netBenefit = (benefit - cost) / filDataSize;
        LOG.debug("BloomFilter benefit=" + benefit
            + ", cost=" + cost
            + ", tsDataSize=" + filDataSize
            + ", netBenefit=" + (benefit - cost));
      }
    }
    LOG.debug("netBenefit=" + netBenefit);
    return netBenefit;
  }

  private void removeSemijoinOptimizationByBenefit(OptimizeTezProcContext procCtx)
      throws SemanticException {
    if(!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION)) {
      // Not needed without semi-join reduction
      return;
    }

    List semijoinRsToRemove = new ArrayList();
    Map map = procCtx.parseContext.getRsToSemiJoinBranchInfo();
    double semijoinReductionThreshold = procCtx.conf.getFloatVar(
        HiveConf.ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION_THRESHOLD);
    for (ReduceSinkOperator rs : map.keySet()) {
      SemiJoinBranchInfo sjInfo = map.get(rs);
      if (sjInfo.getIsHint() || !sjInfo.getShouldRemove()) {
        // Semijoin created using hint or marked useful, skip it
        continue;
      }
      // rs is semijoin optimization branch, which should look like -SEL-GB1-RS1-GB2-RS2
      // Get to the SelectOperator ancestor
      SelectOperator sel = null;
      for (Operator currOp = rs; currOp.getParentOperators().size() > 0; currOp = currOp.getParentOperators().get(0)) {
        if (currOp instanceof SelectOperator) {
          sel = (SelectOperator) currOp;
          break;
        }
      }
      if (sel == null) {
        throw new SemanticException("Unexpected error - could not find SEL ancestor from semijoin branch of " + rs);
      }

      // Check the ndv/rows from the SEL vs the destination tablescan the semijoin opt is going to.
      TableScanOperator ts = sjInfo.getTsOp();
      RuntimeValuesInfo rti = procCtx.parseContext.getRsToRuntimeValuesInfoMap().get(rs);
      ExprNodeDesc tsExpr = rti.getTsColExpr();
      // In the SEL operator of the semijoin branch, there should be only one column in the operator
      ExprNodeDesc selExpr = sel.getConf().getColList().get(0);

      if (LOG.isDebugEnabled()) {
        LOG.debug("Computing BloomFilter cost/benefit for " + OperatorUtils.getOpNamePretty(rs)
            + " - " + OperatorUtils.getOpNamePretty(ts) + " (" + tsExpr + ")");
      }

      double reductionFactor = computeBloomFilterNetBenefit(sel, selExpr,
              (FilterOperator)ts.getChildOperators().get(0), tsExpr);
      if (reductionFactor < semijoinReductionThreshold) {
        // This semijoin optimization should be removed. Do it after we're done iterating
        semijoinRsToRemove.add(rs);
      }
    }

    for (ReduceSinkOperator rs : semijoinRsToRemove) {
      TableScanOperator ts = map.get(rs).getTsOp();
      if (LOG.isDebugEnabled()) {
        LOG.debug("Reduction factor not satisfied for " + OperatorUtils.getOpNamePretty(rs)
            + "-" + OperatorUtils.getOpNamePretty(ts) + ". Removing semijoin optimization.");
      }
      GenTezUtils.removeBranch(rs);
      GenTezUtils.removeSemiJoinOperator(procCtx.parseContext, rs, ts);
    }
  }

  private void markSemiJoinForDPP(OptimizeTezProcContext procCtx)
          throws SemanticException {
    if(!procCtx.conf.getBoolVar(ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION)) {
      // Not needed without semi-join reduction
      return;
    }

    // Stores the Tablescan operators processed to avoid redoing them.
    Map tsOps = new HashMap<>();
    Map map = procCtx.parseContext.getRsToSemiJoinBranchInfo();

    for (ReduceSinkOperator rs : map.keySet()) {
      SemiJoinBranchInfo sjInfo = map.get(rs);
      TableScanOperator ts = sjInfo.getTsOp();
      TableScanOperator tsInMap = tsOps.putIfAbsent(ts, ts);
      if (tsInMap != null) {
        // Already processed, skip
        continue;
      }

      if (sjInfo.getIsHint() || !sjInfo.getShouldRemove()) {
        continue;
      }

      // A TS can have multiple branches due to DPP Or Semijoin Opt.
      // Use DFS to traverse all the branches until RS or DPP is hit.
      Deque> deque = new LinkedList<>();
      deque.add(ts);
      while (!deque.isEmpty()) {
        Operator op = deque.pollLast();
        if (op instanceof AppMasterEventOperator &&
                ((AppMasterEventOperator) op).getConf() instanceof DynamicPruningEventDesc) {
          // DPP. Now look up nDVs on both sides to see the selectivity.
          // -SEL-GB1-RS1-GB2-RS2
          SelectOperator selOp = null;
          try {
            selOp = (SelectOperator)
                    (rs.getParentOperators().get(0)
                            .getParentOperators().get(0)
                            .getParentOperators().get(0)
                            .getParentOperators().get(0));
          } catch (NullPointerException e) {
            LOG.warn("markSemiJoinForDPP : Null pointer exception caught while accessing semijoin operators");
            assert false;
            return;
          }
          try {
            // If stats are not available, just assume its a useful edge
            Statistics stats = selOp.getStatistics();
            ExprNodeColumnDesc colExpr = ExprNodeDescUtils.getColumnExpr(
                    selOp.getConf().getColList().get(0));
            long nDVs = stats.getColumnStatisticsFromColName(
                    colExpr.getColumn()).getCountDistint();
            if (nDVs > 0) {
              // Lookup nDVs on TS side.
              RuntimeValuesInfo rti = procCtx.parseContext
                      .getRsToRuntimeValuesInfoMap().get(rs);
              ExprNodeDesc tsExpr = rti.getTsColExpr();
              FilterOperator fil = (FilterOperator) (ts.getChildOperators().get(0));
              Statistics filStats = fil.getStatistics();
              ExprNodeColumnDesc tsColExpr = ExprNodeDescUtils.getColumnExpr(tsExpr);
              long nDVsOfTS = filStats.getColumnStatisticsFromColName(
                      tsColExpr.getColumn()).getCountDistint();
              double nDVsOfTSFactored = nDVsOfTS * procCtx.conf.getFloatVar(
                      ConfVars.TEZ_DYNAMIC_SEMIJOIN_REDUCTION_FOR_DPP_FACTOR);
              if ((long)nDVsOfTSFactored > nDVs) {
                if (LOG.isDebugEnabled()) {
                  LOG.debug("nDVs = " + nDVs + ", nDVsOfTS = " + nDVsOfTS + " and nDVsOfTSFactored = " + nDVsOfTSFactored
                  + "Adding semijoin branch from ReduceSink " + rs + " to TS " + sjInfo.getTsOp());
                }
                sjInfo.setShouldRemove(false);
              }
            }
          } catch (NullPointerException e) {
            sjInfo.setShouldRemove(false);
          }
          break;
        }
        if (op instanceof ReduceSinkOperator) {
          // Done with this branch
          continue;
        }
        deque.addAll(op.getChildOperators());
      }
    }
  }
}




© 2015 - 2024 Weber Informatics LLC | Privacy Policy