Many resources are needed to download a project. Please understand that we have to compensate our server costs. Thank you in advance. Project price only 1 $
You can buy this project and download/modify it how often you want.
/*
* 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.util.ArrayList;
import java.util.LinkedList;
import java.util.List;
import java.util.Map;
import java.util.Map.Entry;
import java.util.Stack;
import org.apache.hadoop.hive.conf.HiveConf;
import org.apache.hadoop.hive.ql.exec.CommonMergeJoinOperator;
import org.apache.hadoop.hive.ql.exec.DummyStoreOperator;
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.lib.Node;
import org.apache.hadoop.hive.ql.lib.NodeProcessor;
import org.apache.hadoop.hive.ql.lib.NodeProcessorCtx;
import org.apache.hadoop.hive.ql.optimizer.GenMapRedUtils;
import org.apache.hadoop.hive.ql.optimizer.ReduceSinkMapJoinProc;
import org.apache.hadoop.hive.ql.plan.BaseWork;
import org.apache.hadoop.hive.ql.plan.MergeJoinWork;
import org.apache.hadoop.hive.ql.plan.OperatorDesc;
import org.apache.hadoop.hive.ql.plan.ReduceSinkDesc;
import org.apache.hadoop.hive.ql.plan.ReduceWork;
import org.apache.hadoop.hive.ql.plan.TezEdgeProperty;
import org.apache.hadoop.hive.ql.plan.TezEdgeProperty.EdgeType;
import org.apache.hadoop.hive.ql.plan.TezWork;
import org.apache.hadoop.hive.ql.plan.TezWork.VertexType;
import org.apache.hadoop.hive.ql.plan.UnionWork;
import com.facebook.presto.hive.$internal.org.slf4j.Logger;
import com.facebook.presto.hive.$internal.org.slf4j.LoggerFactory;
/**
* GenTezWork separates the operator tree into tez tasks.
* It is called once per leaf operator (operator that forces
* a new execution unit.) and break the operators into work
* and tasks along the way.
*/
public class GenTezWork implements NodeProcessor {
private static final Logger LOG = LoggerFactory.getLogger(GenTezWork.class.getName());
private final GenTezUtils utils;
public GenTezWork(GenTezUtils utils) {
this.utils = utils;
}
@Override
public Object process(Node nd, Stack stack,
NodeProcessorCtx procContext, Object... nodeOutputs)
throws SemanticException {
GenTezProcContext context = (GenTezProcContext) procContext;
assert context != null && context.currentTask != null
&& context.currentRootOperator != null;
// Operator is a file sink or reduce sink. Something that forces
// a new vertex.
Operator operator = (Operator) nd;
// root is the start of the operator pipeline we're currently
// packing into a vertex, typically a table scan, union or join
Operator root = context.currentRootOperator;
LOG.debug("Root operator: " + root);
LOG.debug("Leaf operator: " + operator);
if (context.clonedReduceSinks.contains(operator)) {
// if we're visiting a terminal we've created ourselves,
// just skip and keep going
return null;
}
TezWork tezWork = context.currentTask.getWork();
// Right now the work graph is pretty simple. If there is no
// Preceding work we have a root and will generate a map
// vertex. If there is a preceding work we will generate
// a reduce vertex
BaseWork work;
if (context.rootToWorkMap.containsKey(root)) {
// having seen the root operator before means there was a branch in the
// operator graph. There's typically two reasons for that: a) mux/demux
// b) multi insert. Mux/Demux will hit the same leaf again, multi insert
// will result into a vertex with multiple FS or RS operators.
if (context.childToWorkMap.containsKey(operator)) {
// if we've seen both root and child, we can bail.
// clear out the mapjoin set. we don't need it anymore.
context.currentMapJoinOperators.clear();
// clear out the union set. we don't need it anymore.
context.currentUnionOperators.clear();
return null;
} else {
// At this point we don't have to do anything special. Just
// run through the regular paces w/o creating a new task.
work = context.rootToWorkMap.get(root);
}
} else {
// create a new vertex
if (context.preceedingWork == null) {
work = utils.createMapWork(context, root, tezWork, null);
} else {
work = GenTezUtils.createReduceWork(context, root, tezWork);
}
context.rootToWorkMap.put(root, work);
}
// this is where we set the sort columns that we will be using for KeyValueInputMerge
if (operator instanceof DummyStoreOperator) {
work.addSortCols(root.getOpTraits().getSortCols().get(0));
}
if (!context.childToWorkMap.containsKey(operator)) {
List workItems = new LinkedList();
workItems.add(work);
context.childToWorkMap.put(operator, workItems);
} else {
context.childToWorkMap.get(operator).add(work);
}
// this transformation needs to be first because it changes the work item itself.
// which can affect the working of all downstream transformations.
if (context.currentMergeJoinOperator != null) {
// we are currently walking the big table side of the merge join. we need to create or hook up
// merge join work.
MergeJoinWork mergeJoinWork = null;
if (context.opMergeJoinWorkMap.containsKey(context.currentMergeJoinOperator)) {
// we have found a merge work corresponding to this closing operator. Hook up this work.
mergeJoinWork = context.opMergeJoinWorkMap.get(context.currentMergeJoinOperator);
} else {
// we need to create the merge join work
mergeJoinWork = new MergeJoinWork();
mergeJoinWork.setMergeJoinOperator(context.currentMergeJoinOperator);
tezWork.add(mergeJoinWork);
context.opMergeJoinWorkMap.put(context.currentMergeJoinOperator, mergeJoinWork);
}
// connect the work correctly.
work.addSortCols(root.getOpTraits().getSortCols().get(0));
mergeJoinWork.addMergedWork(work, null, context.leafOperatorToFollowingWork);
Operator parentOp =
getParentFromStack(context.currentMergeJoinOperator, stack);
// Set the big table position. Both the reduce work and merge join operator
// should be set with the same value.
// int pos = context.currentMergeJoinOperator.getTagForOperator(parentOp);
int pos = context.currentMergeJoinOperator.getConf().getBigTablePosition();
work.setTag(pos);
context.currentMergeJoinOperator.getConf().setBigTablePosition(pos);
tezWork.setVertexType(work, VertexType.MULTI_INPUT_UNINITIALIZED_EDGES);
for (BaseWork parentWork : tezWork.getParents(work)) {
TezEdgeProperty edgeProp = tezWork.getEdgeProperty(parentWork, work);
tezWork.disconnect(parentWork, work);
tezWork.connect(parentWork, mergeJoinWork, edgeProp);
}
for (BaseWork childWork : tezWork.getChildren(work)) {
TezEdgeProperty edgeProp = tezWork.getEdgeProperty(work, childWork);
tezWork.disconnect(work, childWork);
tezWork.connect(mergeJoinWork, childWork, edgeProp);
}
tezWork.remove(work);
context.rootToWorkMap.put(root, mergeJoinWork);
context.childToWorkMap.get(operator).remove(work);
context.childToWorkMap.get(operator).add(mergeJoinWork);
work = mergeJoinWork;
context.currentMergeJoinOperator = null;
}
// remember which mapjoin operator links with which work
if (!context.currentMapJoinOperators.isEmpty()) {
for (MapJoinOperator mj: context.currentMapJoinOperators) {
// For dynamic partitioned hash join, ReduceSinkMapJoinProc rule may not get run for all
// of the ReduceSink parents, because the parents of the MapJoin operator get
// removed later on in this method. Keep track of the parent to mapjoin mapping
// so we can later run the same logic that is run in ReduceSinkMapJoinProc.
if (mj.getConf().isDynamicPartitionHashJoin()) {
// Since this is a dynamic partitioned hash join, the work for this join should be a ReduceWork
ReduceWork reduceWork = (ReduceWork) work;
int bigTablePosition = mj.getConf().getPosBigTable();
reduceWork.setTag(bigTablePosition);
// Use context.mapJoinParentMap to get the original RS parents, because
// the MapJoin's parents may have been replaced by dummy operator.
List> mapJoinOriginalParents = context.mapJoinParentMap.get(mj);
if (mapJoinOriginalParents == null) {
throw new SemanticException("Unexpected error - context.mapJoinParentMap did not have an entry for " + mj);
}
for (int pos = 0; pos < mapJoinOriginalParents.size(); ++pos) {
// This processing only needs to happen for the small tables
if (pos == bigTablePosition) {
continue;
}
Operator parentOp = mapJoinOriginalParents.get(pos);
context.smallTableParentToMapJoinMap.put(parentOp, mj);
ReduceSinkOperator parentRS = (ReduceSinkOperator) parentOp;
// TableDesc needed for dynamic partitioned hash join
GenMapRedUtils.setKeyAndValueDesc(reduceWork, parentRS);
// For small table RS parents that have already been processed, we need to
// add the tag to the RS work to the reduce work that contains this map join.
// This was not being done for normal mapjoins, where the small table typically
// has its ReduceSink parent removed.
if (!context.mapJoinToUnprocessedSmallTableReduceSinks.get(mj).contains(parentRS)) {
// This reduce sink has been processed already, so the work for the parentRS exists
BaseWork parentWork = ReduceSinkMapJoinProc.getMapJoinParentWork(context, parentRS);
int tag = parentRS.getConf().getTag();
tag = (tag == -1 ? 0 : tag);
reduceWork.getTagToInput().put(tag, parentWork.getName());
}
}
}
LOG.debug("Processing map join: " + mj);
// remember the mapping in case we scan another branch of the
// mapjoin later
if (!context.mapJoinWorkMap.containsKey(mj)) {
List workItems = new LinkedList();
workItems.add(work);
context.mapJoinWorkMap.put(mj, workItems);
} else {
context.mapJoinWorkMap.get(mj).add(work);
}
/*
* this happens in case of map join operations.
* The tree looks like this:
*
* RS <--- we are here perhaps
* |
* MapJoin
* / \
* RS TS
* /
* TS
*
* If we are at the RS pointed above, and we may have already visited the
* RS following the TS, we have already generated work for the TS-RS.
* We need to hook the current work to this generated work.
*/
if (context.linkOpWithWorkMap.containsKey(mj)) {
Map linkWorkMap = context.linkOpWithWorkMap.get(mj);
if (linkWorkMap != null) {
// Note: it's not quite clear why this is done inside this if. Seems like it should be on the top level.
if (context.linkChildOpWithDummyOp.containsKey(mj)) {
if (LOG.isDebugEnabled()) {
LOG.debug("Adding dummy ops to work: " + work.getName() + ": " + context.linkChildOpWithDummyOp.get(mj));
}
for (Operator dummy: context.linkChildOpWithDummyOp.get(mj)) {
work.addDummyOp((HashTableDummyOperator) dummy);
}
}
for (Entry parentWorkMap : linkWorkMap.entrySet()) {
BaseWork parentWork = parentWorkMap.getKey();
LOG.debug("connecting "+parentWork.getName()+" with "+work.getName());
TezEdgeProperty edgeProp = parentWorkMap.getValue();
tezWork.connect(parentWork, work, edgeProp);
if (edgeProp.getEdgeType() == EdgeType.CUSTOM_EDGE) {
tezWork.setVertexType(work, VertexType.INITIALIZED_EDGES);
}
// need to set up output name for reduce sink now that we know the name
// of the downstream work
for (ReduceSinkOperator r:
context.linkWorkWithReduceSinkMap.get(parentWork)) {
if (!context.mapJoinParentMap.get(mj).contains(r)) {
// We might be visiting twice because of reutilization of intermediary results.
// If that is the case, we do not need to do anything because either we have
// already connected this RS operator or we will connect it at subsequent pass.
continue;
}
if (r.getConf().getOutputName() != null) {
LOG.debug("Cloning reduce sink " + r + " for multi-child broadcast edge");
// we've already set this one up. Need to clone for the next work.
r = (ReduceSinkOperator) OperatorFactory.getAndMakeChild(
r.getCompilationOpContext(), (ReduceSinkDesc)r.getConf().clone(),
new RowSchema(r.getSchema()), r.getParentOperators());
context.clonedReduceSinks.add(r);
}
r.getConf().setOutputName(work.getName());
context.connectedReduceSinks.add(r);
}
}
}
}
}
// clear out the set. we don't need it anymore.
context.currentMapJoinOperators.clear();
}
// This is where we cut the tree as described above. We also remember that
// we might have to connect parent work with this work later.
for (Operator parent : new ArrayList>(root.getParentOperators())) {
if (LOG.isDebugEnabled()) {
LOG.debug("Removing " + parent + " as parent from " + root);
}
context.leafOperatorToFollowingWork.remove(parent);
context.leafOperatorToFollowingWork.put(parent, work);
root.removeParent(parent);
}
if (!context.currentUnionOperators.isEmpty()) {
// if there are union all operators, it means that the walking context contains union all operators.
// please see more details of context.currentUnionOperator in GenTezWorkWalker
UnionWork unionWork;
if (context.unionWorkMap.containsKey(operator)) {
// we've seen this terminal before and have created a union work object.
// just need to add this work to it. There will be no children of this one
// since we've passed this operator before.
assert operator.getChildOperators().isEmpty();
unionWork = (UnionWork) context.unionWorkMap.get(operator);
// finally connect the union work with work
connectUnionWorkWithWork(unionWork, work, tezWork, context);
} else {
// we've not seen this terminal before. we need to check
// rootUnionWorkMap which contains the information of mapping the root
// operator of a union work to a union work
unionWork = context.rootUnionWorkMap.get(root);
if (unionWork == null) {
// if unionWork is null, it means it is the first time. we need to
// create a union work object and add this work to it. Subsequent
// work should reference the union and not the actual work.
unionWork = GenTezUtils.createUnionWork(context, root, operator, tezWork);
// finally connect the union work with work
connectUnionWorkWithWork(unionWork, work, tezWork, context);
}
}
context.currentUnionOperators.clear();
work = unionWork;
}
// We're scanning a tree from roots to leaf (this is not technically
// correct, demux and mux operators might form a diamond shape, but
// we will only scan one path and ignore the others, because the
// diamond shape is always contained in a single vertex). The scan
// is depth first and because we remove parents when we pack a pipeline
// into a vertex we will never visit any node twice. But because of that
// we might have a situation where we need to connect 'work' that comes after
// the 'work' we're currently looking at.
//
// Also note: the concept of leaf and root is reversed in hive for historical
// reasons. Roots are data sources, leaves are data sinks. I know.
if (context.leafOperatorToFollowingWork.containsKey(operator)) {
BaseWork followingWork = context.leafOperatorToFollowingWork.get(operator);
long bytesPerReducer = context.conf.getLongVar(HiveConf.ConfVars.BYTESPERREDUCER);
LOG.debug("Second pass. Leaf operator: "+operator
+" has common downstream work: "+followingWork);
if (operator instanceof DummyStoreOperator) {
// this is the small table side.
assert (followingWork instanceof MergeJoinWork);
MergeJoinWork mergeJoinWork = (MergeJoinWork) followingWork;
CommonMergeJoinOperator mergeJoinOp = mergeJoinWork.getMergeJoinOperator();
work.setTag(mergeJoinOp.getTagForOperator(operator));
mergeJoinWork.addMergedWork(null, work, context.leafOperatorToFollowingWork);
tezWork.setVertexType(mergeJoinWork, VertexType.MULTI_INPUT_UNINITIALIZED_EDGES);
for (BaseWork parentWork : tezWork.getParents(work)) {
TezEdgeProperty edgeProp = tezWork.getEdgeProperty(parentWork, work);
tezWork.disconnect(parentWork, work);
tezWork.connect(parentWork, mergeJoinWork, edgeProp);
}
work = mergeJoinWork;
} else {
// need to add this branch to the key + value info
assert operator instanceof ReduceSinkOperator
&& ((followingWork instanceof ReduceWork) || (followingWork instanceof MergeJoinWork)
|| followingWork instanceof UnionWork);
ReduceSinkOperator rs = (ReduceSinkOperator) operator;
ReduceWork rWork = null;
if (followingWork instanceof MergeJoinWork) {
MergeJoinWork mergeJoinWork = (MergeJoinWork) followingWork;
rWork = (ReduceWork) mergeJoinWork.getMainWork();
} else if (followingWork instanceof UnionWork) {
// this can only be possible if there is merge work followed by the union
UnionWork unionWork = (UnionWork) followingWork;
int index = getFollowingWorkIndex(tezWork, unionWork, rs);
BaseWork baseWork = tezWork.getChildren(unionWork).get(index);
if (baseWork instanceof MergeJoinWork) {
MergeJoinWork mergeJoinWork = (MergeJoinWork) baseWork;
// disconnect the connection to union work and connect to merge work
followingWork = mergeJoinWork;
rWork = (ReduceWork) mergeJoinWork.getMainWork();
} else {
rWork = (ReduceWork) baseWork;
}
} else {
rWork = (ReduceWork) followingWork;
}
GenMapRedUtils.setKeyAndValueDesc(rWork, rs);
// remember which parent belongs to which tag
int tag = rs.getConf().getTag();
rWork.getTagToInput().put(tag == -1 ? 0 : tag, work.getName());
// remember the output name of the reduce sink
rs.getConf().setOutputName(rWork.getName());
// For dynamic partitioned hash join, run the ReduceSinkMapJoinProc logic for any
// ReduceSink parents that we missed.
MapJoinOperator mj = context.smallTableParentToMapJoinMap.get(rs);
if (mj != null) {
// Only need to run the logic for tables we missed
if (context.mapJoinToUnprocessedSmallTableReduceSinks.get(mj).contains(rs)) {
// ReduceSinkMapJoinProc logic does not work unless the ReduceSink is connected as
// a parent of the MapJoin, but at this point we have already removed all of the
// parents from the MapJoin.
// Try temporarily adding the RS as a parent
ArrayList> tempMJParents = new ArrayList>();
tempMJParents.add(rs);
mj.setParentOperators(tempMJParents);
// ReduceSink also needs MapJoin as child
List> rsChildren = rs.getChildOperators();
rsChildren.add(mj);
// Since the MapJoin has had all of its other parents removed at this point,
// it would be bad here if processReduceSinkToHashJoin() tries to do anything
// with the RS parent based on its position in the list of parents.
ReduceSinkMapJoinProc.processReduceSinkToHashJoin(rs, mj, context);
// Remove any parents from MapJoin again
mj.removeParents();
// TODO: do we also need to remove the MapJoin from the list of RS's children?
}
}
if (!context.connectedReduceSinks.contains(rs)) {
// add dependency between the two work items
TezEdgeProperty edgeProp;
EdgeType edgeType = GenTezUtils.determineEdgeType(work, followingWork, rs);
if (rWork.isAutoReduceParallelism()) {
edgeProp =
new TezEdgeProperty(context.conf, edgeType, true, rWork.isSlowStart(),
rWork.getMinReduceTasks(), rWork.getMaxReduceTasks(), bytesPerReducer);
} else {
edgeProp = new TezEdgeProperty(edgeType);
edgeProp.setSlowStart(rWork.isSlowStart());
}
tezWork.connect(work, followingWork, edgeProp);
context.connectedReduceSinks.add(rs);
}
}
} else {
LOG.debug("First pass. Leaf operator: "+operator);
}
// No children means we're at the bottom. If there are more operators to scan
// the next item will be a new root.
if (!operator.getChildOperators().isEmpty()) {
assert operator.getChildOperators().size() == 1;
context.parentOfRoot = operator;
context.currentRootOperator = operator.getChildOperators().get(0);
context.preceedingWork = work;
}
return null;
}
private int getFollowingWorkIndex(TezWork tezWork, UnionWork unionWork, ReduceSinkOperator rs)
throws SemanticException {
int index = 0;
for (BaseWork baseWork : tezWork.getChildren(unionWork)) {
TezEdgeProperty edgeProperty = tezWork.getEdgeProperty(unionWork, baseWork);
if (edgeProperty.getEdgeType() != TezEdgeProperty.EdgeType.CONTAINS) {
return index;
}
index++;
}
throw new SemanticException("Following work not found for the reduce sink: " + rs.getName());
}
@SuppressWarnings("unchecked")
private Operator getParentFromStack(Node currentMergeJoinOperator,
Stack stack) {
int pos = stack.indexOf(currentMergeJoinOperator);
return (Operator) stack.get(pos - 1);
}
private void connectUnionWorkWithWork(UnionWork unionWork, BaseWork work, TezWork tezWork,
GenTezProcContext context) {
LOG.debug("Connecting union work (" + unionWork + ") with work (" + work + ")");
TezEdgeProperty edgeProp = new TezEdgeProperty(EdgeType.CONTAINS);
tezWork.connect(unionWork, work, edgeProp);
unionWork.addUnionOperators(context.currentUnionOperators);
context.workWithUnionOperators.add(work);
}
}
