/**
* 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.optimizer.spark;
import java.util.ArrayList;
import java.util.HashMap;
import java.util.LinkedHashMap;
import java.util.List;
import java.util.Map;
import java.util.Stack;
import org.apache.hadoop.hive.conf.HiveConf;
import org.apache.hadoop.hive.ql.exec.GroupByOperator;
import org.apache.hadoop.hive.ql.exec.HashTableDummyOperator;
import org.apache.hadoop.hive.ql.exec.MapJoinOperator;
import org.apache.hadoop.hive.ql.exec.Operator;
import org.apache.hadoop.hive.ql.exec.OperatorFactory;
import org.apache.hadoop.hive.ql.exec.ReduceSinkOperator;
import org.apache.hadoop.hive.ql.exec.RowSchema;
import org.apache.hadoop.hive.ql.exec.SparkHashTableSinkOperator;
import org.apache.hadoop.hive.ql.lib.DefaultGraphWalker;
import org.apache.hadoop.hive.ql.lib.DefaultRuleDispatcher;
import org.apache.hadoop.hive.ql.lib.Dispatcher;
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.Rule;
import org.apache.hadoop.hive.ql.lib.RuleRegExp;
import org.apache.hadoop.hive.ql.parse.SemanticException;
import org.apache.hadoop.hive.ql.parse.spark.GenSparkProcContext;
import org.apache.hadoop.hive.ql.plan.BaseWork;
import org.apache.hadoop.hive.ql.plan.ExprNodeDesc;
import org.apache.hadoop.hive.ql.plan.HashTableDummyDesc;
import org.apache.hadoop.hive.ql.plan.MapJoinDesc;
import org.apache.hadoop.hive.ql.plan.OperatorDesc;
import org.apache.hadoop.hive.ql.plan.PlanUtils;
import org.apache.hadoop.hive.ql.plan.SparkEdgeProperty;
import org.apache.hadoop.hive.ql.plan.SparkHashTableSinkDesc;
import org.apache.hadoop.hive.ql.plan.SparkWork;
import org.apache.hadoop.hive.ql.plan.TableDesc;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import com.google.common.base.Preconditions;
public class SparkReduceSinkMapJoinProc implements NodeProcessor {
public static final Logger LOG = LoggerFactory.getLogger(SparkReduceSinkMapJoinProc.class.getName());
public static class SparkMapJoinFollowedByGroupByProcessor implements NodeProcessor {
private boolean hasGroupBy = false;
@Override
public Object process(Node nd, Stack<Node> stack, NodeProcessorCtx procCtx,
Object... nodeOutputs) throws SemanticException {
GenSparkProcContext context = (GenSparkProcContext) procCtx;
hasGroupBy = true;
GroupByOperator op = (GroupByOperator) nd;
float groupByMemoryUsage = context.conf.getFloatVar(
HiveConf.ConfVars.HIVEMAPJOINFOLLOWEDBYMAPAGGRHASHMEMORY);
op.getConf().setGroupByMemoryUsage(groupByMemoryUsage);
return null;
}
public boolean getHasGroupBy() {
return hasGroupBy;
}
}
private boolean hasGroupBy(Operator<? extends OperatorDesc> mapjoinOp,
GenSparkProcContext context) throws SemanticException {
List<Operator<? extends OperatorDesc>> childOps = mapjoinOp.getChildOperators();
Map<Rule, NodeProcessor> rules = new LinkedHashMap<Rule, NodeProcessor>();
SparkMapJoinFollowedByGroupByProcessor processor = new SparkMapJoinFollowedByGroupByProcessor();
rules.put(new RuleRegExp("GBY", GroupByOperator.getOperatorName() + "%"), processor);
Dispatcher disp = new DefaultRuleDispatcher(null, rules, context);
GraphWalker ogw = new DefaultGraphWalker(disp);
ArrayList<Node> topNodes = new ArrayList<Node>();
topNodes.addAll(childOps);
ogw.startWalking(topNodes, null);
return processor.getHasGroupBy();
}
/* (non-Javadoc)
* This processor addresses the RS-MJ case that occurs in spark on the small/hash
* table side of things. The work that RS will be a part of must be connected
* to the MJ work via be a broadcast edge.
* We should not walk down the tree when we encounter this pattern because:
* the type of work (map work or reduce work) needs to be determined
* on the basis of the big table side because it may be a mapwork (no need for shuffle)
* or reduce work.
*/
@SuppressWarnings("unchecked")
@Override
public Object process(Node nd, Stack<Node> stack,
NodeProcessorCtx procContext, Object... nodeOutputs)
throws SemanticException {
GenSparkProcContext context = (GenSparkProcContext) procContext;
if (!nd.getClass().equals(MapJoinOperator.class)) {
return null;
}
MapJoinOperator mapJoinOp = (MapJoinOperator)nd;
if (stack.size() < 2 || !(stack.get(stack.size() - 2) instanceof ReduceSinkOperator)) {
context.currentMapJoinOperators.add(mapJoinOp);
return null;
}
context.preceedingWork = null;
context.currentRootOperator = null;
ReduceSinkOperator parentRS = (ReduceSinkOperator)stack.get(stack.size() - 2);
// remove the tag for in-memory side of mapjoin
parentRS.getConf().setSkipTag(true);
parentRS.setSkipTag(true);
// remember the original parent list before we start modifying it.
if (!context.mapJoinParentMap.containsKey(mapJoinOp)) {
List<Operator<?>> parents = new ArrayList<Operator<?>>(mapJoinOp.getParentOperators());
context.mapJoinParentMap.put(mapJoinOp, parents);
}
List<BaseWork> mapJoinWork;
/*
* If there was a pre-existing work generated for the big-table mapjoin side,
* we need to hook the work generated for the RS (associated with the RS-MJ pattern)
* with the pre-existing work.
*
* Otherwise, we need to associate that the mapjoin op
* to be linked to the RS work (associated with the RS-MJ pattern).
*
*/
mapJoinWork = context.mapJoinWorkMap.get(mapJoinOp);
int workMapSize = context.childToWorkMap.get(parentRS).size();
Preconditions.checkArgument(workMapSize == 1,
"AssertionError: expected context.childToWorkMap.get(parentRS).size() to be 1, but was " + workMapSize);
BaseWork parentWork = context.childToWorkMap.get(parentRS).get(0);
// set the link between mapjoin and parent vertex
int pos = context.mapJoinParentMap.get(mapJoinOp).indexOf(parentRS);
if (pos == -1) {
throw new SemanticException("Cannot find position of parent in mapjoin");
}
LOG.debug("Mapjoin "+mapJoinOp+", pos: "+pos+" --> "+parentWork.getName());
mapJoinOp.getConf().getParentToInput().put(pos, parentWork.getName());
SparkEdgeProperty edgeProp = new SparkEdgeProperty(SparkEdgeProperty.SHUFFLE_NONE);
if (mapJoinWork != null) {
for (BaseWork myWork: mapJoinWork) {
// link the work with the work associated with the reduce sink that triggered this rule
SparkWork sparkWork = context.currentTask.getWork();
LOG.debug("connecting "+parentWork.getName()+" with "+myWork.getName());
sparkWork.connect(parentWork, myWork, edgeProp);
}
}
// remember in case we need to connect additional work later
Map<BaseWork, SparkEdgeProperty> linkWorkMap = null;
if (context.linkOpWithWorkMap.containsKey(mapJoinOp)) {
linkWorkMap = context.linkOpWithWorkMap.get(mapJoinOp);
} else {
linkWorkMap = new HashMap<BaseWork, SparkEdgeProperty>();
}
linkWorkMap.put(parentWork, edgeProp);
context.linkOpWithWorkMap.put(mapJoinOp, linkWorkMap);
List<ReduceSinkOperator> reduceSinks
= context.linkWorkWithReduceSinkMap.get(parentWork);
if (reduceSinks == null) {
reduceSinks = new ArrayList<ReduceSinkOperator>();
}
reduceSinks.add(parentRS);
context.linkWorkWithReduceSinkMap.put(parentWork, reduceSinks);
// create the dummy operators
List<Operator<?>> dummyOperators = new ArrayList<Operator<?>>();
// create an new operator: HashTableDummyOperator, which share the table desc
HashTableDummyDesc desc = new HashTableDummyDesc();
HashTableDummyOperator dummyOp = (HashTableDummyOperator) OperatorFactory.get(
mapJoinOp.getCompilationOpContext(), desc);
TableDesc tbl;
// need to create the correct table descriptor for key/value
RowSchema rowSchema = parentRS.getParentOperators().get(0).getSchema();
tbl = PlanUtils.getReduceValueTableDesc(PlanUtils.getFieldSchemasFromRowSchema(rowSchema, ""));
dummyOp.getConf().setTbl(tbl);
Map<Byte, List<ExprNodeDesc>> keyExprMap = mapJoinOp.getConf().getKeys();
List<ExprNodeDesc> keyCols = keyExprMap.get(Byte.valueOf((byte) 0));
StringBuilder keyOrder = new StringBuilder();
StringBuilder keyNullOrder = new StringBuilder();
for (int i = 0; i < keyCols.size(); i++) {
keyOrder.append("+");
keyNullOrder.append("a");
}
TableDesc keyTableDesc = PlanUtils.getReduceKeyTableDesc(PlanUtils
.getFieldSchemasFromColumnList(keyCols, "mapjoinkey"), keyOrder.toString(),
keyNullOrder.toString());
mapJoinOp.getConf().setKeyTableDesc(keyTableDesc);
// let the dummy op be the parent of mapjoin op
mapJoinOp.replaceParent(parentRS, dummyOp);
List<Operator<? extends OperatorDesc>> dummyChildren =
new ArrayList<Operator<? extends OperatorDesc>>();
dummyChildren.add(mapJoinOp);
dummyOp.setChildOperators(dummyChildren);
dummyOperators.add(dummyOp);
// cut the operator tree so as to not retain connections from the parent RS downstream
List<Operator<? extends OperatorDesc>> childOperators = parentRS.getChildOperators();
int childIndex = childOperators.indexOf(mapJoinOp);
childOperators.remove(childIndex);
// the "work" needs to know about the dummy operators. They have to be separately initialized
// at task startup
if (mapJoinWork != null) {
for (BaseWork myWork: mapJoinWork) {
myWork.addDummyOp(dummyOp);
}
}
if (context.linkChildOpWithDummyOp.containsKey(mapJoinOp)) {
for (Operator<?> op: context.linkChildOpWithDummyOp.get(mapJoinOp)) {
dummyOperators.add(op);
}
}
context.linkChildOpWithDummyOp.put(mapJoinOp, dummyOperators);
// replace ReduceSinkOp with HashTableSinkOp for the RSops which are parents of MJop
MapJoinDesc mjDesc = mapJoinOp.getConf();
HiveConf conf = context.conf;
// Unlike in MR, we may call this method multiple times, for each
// small table HTS. But, since it's idempotent, it should be OK.
mjDesc.resetOrder();
float hashtableMemoryUsage;
if (hasGroupBy(mapJoinOp, context)) {
hashtableMemoryUsage = conf.getFloatVar(
HiveConf.ConfVars.HIVEHASHTABLEFOLLOWBYGBYMAXMEMORYUSAGE);
} else {
hashtableMemoryUsage = conf.getFloatVar(
HiveConf.ConfVars.HIVEHASHTABLEMAXMEMORYUSAGE);
}
mjDesc.setHashTableMemoryUsage(hashtableMemoryUsage);
SparkHashTableSinkDesc hashTableSinkDesc = new SparkHashTableSinkDesc(mjDesc);
SparkHashTableSinkOperator hashTableSinkOp = (SparkHashTableSinkOperator) OperatorFactory.get(
mapJoinOp.getCompilationOpContext(), hashTableSinkDesc);
byte tag = (byte) pos;
int[] valueIndex = mjDesc.getValueIndex(tag);
if (valueIndex != null) {
List<ExprNodeDesc> newValues = new ArrayList<ExprNodeDesc>();
List<ExprNodeDesc> values = hashTableSinkDesc.getExprs().get(tag);
for (int index = 0; index < values.size(); index++) {
if (valueIndex[index] < 0) {
newValues.add(values.get(index));
}
}
hashTableSinkDesc.getExprs().put(tag, newValues);
}
//get all parents of reduce sink
List<Operator<? extends OperatorDesc>> rsParentOps = parentRS.getParentOperators();
for (Operator<? extends OperatorDesc> parent : rsParentOps) {
parent.replaceChild(parentRS, hashTableSinkOp);
}
hashTableSinkOp.setParentOperators(rsParentOps);
hashTableSinkOp.getConf().setTag(tag);
return true;
}
}