use of org.apache.hadoop.hive.ql.plan.HashTableDummyDesc in project hive by apache.
the class SparkReduceSinkMapJoinProc method process.
/* (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(NullOrdering.defaultNullOrder(context.conf));
}
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);
// 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;
}
use of org.apache.hadoop.hive.ql.plan.HashTableDummyDesc in project hive by apache.
the class GenSparkSkewJoinProcessor method insertSHTS.
/**
* Insert SparkHashTableSink and HashTableDummy between small dir TS and MJ.
*/
@SuppressWarnings("unchecked")
private static void insertSHTS(byte tag, TableScanOperator tableScan, MapWork bigMapWork) {
Preconditions.checkArgument(tableScan.getChildOperators().size() == 1 && tableScan.getChildOperators().get(0) instanceof MapJoinOperator);
HashTableDummyDesc desc = new HashTableDummyDesc();
HashTableDummyOperator dummyOp = (HashTableDummyOperator) OperatorFactory.get(tableScan.getCompilationOpContext(), desc);
dummyOp.getConf().setTbl(tableScan.getTableDescSkewJoin());
MapJoinOperator mapJoinOp = (MapJoinOperator) tableScan.getChildOperators().get(0);
mapJoinOp.replaceParent(tableScan, dummyOp);
List<Operator<? extends OperatorDesc>> mapJoinChildren = new ArrayList<Operator<? extends OperatorDesc>>();
mapJoinChildren.add(mapJoinOp);
dummyOp.setChildOperators(mapJoinChildren);
bigMapWork.addDummyOp(dummyOp);
MapJoinDesc mjDesc = mapJoinOp.getConf();
// mapjoin should not be affected by join reordering
mjDesc.resetOrder();
SparkHashTableSinkDesc hashTableSinkDesc = new SparkHashTableSinkDesc(mjDesc);
SparkHashTableSinkOperator hashTableSinkOp = (SparkHashTableSinkOperator) OperatorFactory.get(tableScan.getCompilationOpContext(), hashTableSinkDesc);
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);
}
tableScan.replaceChild(mapJoinOp, hashTableSinkOp);
List<Operator<? extends OperatorDesc>> tableScanParents = new ArrayList<Operator<? extends OperatorDesc>>();
tableScanParents.add(tableScan);
hashTableSinkOp.setParentOperators(tableScanParents);
hashTableSinkOp.getConf().setTag(tag);
}
use of org.apache.hadoop.hive.ql.plan.HashTableDummyDesc in project hive by apache.
the class ReduceSinkMapJoinProc method processReduceSinkToHashJoin.
public static Object processReduceSinkToHashJoin(ReduceSinkOperator parentRS, MapJoinOperator mapJoinOp, GenTezProcContext context) throws SemanticException {
// remove the tag for in-memory side of mapjoin
parentRS.getConf().setSkipTag(true);
parentRS.setSkipTag(true);
// Mark this small table as being processed
if (mapJoinOp.getConf().isDynamicPartitionHashJoin()) {
context.mapJoinToUnprocessedSmallTableReduceSinks.get(mapJoinOp).remove(parentRS);
}
List<BaseWork> mapJoinWork = null;
/*
* 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);
BaseWork parentWork = getMapJoinParentWork(context, parentRS);
// 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");
}
MapJoinDesc joinConf = mapJoinOp.getConf();
long keyCount = Long.MAX_VALUE, rowCount = Long.MAX_VALUE, bucketCount = 1;
long tableSize = Long.MAX_VALUE;
Statistics stats = parentRS.getStatistics();
if (stats != null) {
keyCount = rowCount = stats.getNumRows();
if (keyCount <= 0) {
keyCount = rowCount = Long.MAX_VALUE;
}
tableSize = stats.getDataSize();
List<String> keyCols = parentRS.getConf().getOutputKeyColumnNames();
if (keyCols != null && !keyCols.isEmpty()) {
// See if we can arrive at a smaller number using distinct stats from key columns.
long maxKeyCount = 1;
String prefix = Utilities.ReduceField.KEY.toString();
for (String keyCol : keyCols) {
ExprNodeDesc realCol = parentRS.getColumnExprMap().get(prefix + "." + keyCol);
ColStatistics cs = StatsUtils.getColStatisticsFromExpression(context.conf, stats, realCol);
if (cs == null || cs.getCountDistint() <= 0) {
maxKeyCount = Long.MAX_VALUE;
break;
}
maxKeyCount *= cs.getCountDistint();
if (maxKeyCount >= keyCount) {
break;
}
}
keyCount = Math.min(maxKeyCount, keyCount);
}
if (joinConf.isBucketMapJoin()) {
OpTraits opTraits = mapJoinOp.getOpTraits();
bucketCount = (opTraits == null) ? -1 : opTraits.getNumBuckets();
if (bucketCount > 0) {
// We cannot obtain a better estimate without CustomPartitionVertex providing it
// to us somehow; in which case using statistics would be completely unnecessary.
keyCount /= bucketCount;
tableSize /= bucketCount;
}
} else if (joinConf.isDynamicPartitionHashJoin()) {
// For dynamic partitioned hash join, assuming table is split evenly among the reduce tasks.
bucketCount = parentRS.getConf().getNumReducers();
keyCount /= bucketCount;
tableSize /= bucketCount;
}
}
if (keyCount == 0) {
keyCount = 1;
}
if (tableSize == 0) {
tableSize = 1;
}
LOG.info("Mapjoin " + mapJoinOp + "(bucket map join = " + joinConf.isBucketMapJoin() + "), pos: " + pos + " --> " + parentWork.getName() + " (" + keyCount + " keys estimated from " + rowCount + " rows, " + bucketCount + " buckets)");
joinConf.getParentToInput().put(pos, parentWork.getName());
if (keyCount != Long.MAX_VALUE) {
joinConf.getParentKeyCounts().put(pos, keyCount);
}
joinConf.getParentDataSizes().put(pos, tableSize);
int numBuckets = -1;
EdgeType edgeType = EdgeType.BROADCAST_EDGE;
if (joinConf.isBucketMapJoin()) {
numBuckets = (Integer) joinConf.getBigTableBucketNumMapping().values().toArray()[0];
/*
* Here, we can be in one of 4 states.
*
* 1. If map join work is null implies that we have not yet traversed the big table side. We
* just need to see if we can find a reduce sink operator in the big table side. This would
* imply a reduce side operation.
*
* 2. If we don't find a reducesink in 1 it has to be the case that it is a map side operation.
*
* 3. If we have already created a work item for the big table side, we need to see if we can
* find a table scan operator in the big table side. This would imply a map side operation.
*
* 4. If we don't find a table scan operator, it has to be a reduce side operation.
*/
if (mapJoinWork == null) {
Operator<?> rootOp = OperatorUtils.findSingleOperatorUpstreamJoinAccounted(mapJoinOp.getParentOperators().get(joinConf.getPosBigTable()), ReduceSinkOperator.class);
if (rootOp == null) {
// likely we found a table scan operator
edgeType = EdgeType.CUSTOM_EDGE;
} else {
// we have found a reduce sink
edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
}
} else {
Operator<?> rootOp = OperatorUtils.findSingleOperatorUpstreamJoinAccounted(mapJoinOp.getParentOperators().get(joinConf.getPosBigTable()), TableScanOperator.class);
if (rootOp != null) {
// likely we found a table scan operator
edgeType = EdgeType.CUSTOM_EDGE;
} else {
// we have found a reduce sink
edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
}
}
} else if (mapJoinOp.getConf().isDynamicPartitionHashJoin()) {
if (parentRS.getConf().isForwarding()) {
edgeType = EdgeType.ONE_TO_ONE_EDGE;
} else {
edgeType = EdgeType.CUSTOM_SIMPLE_EDGE;
}
}
if (edgeType == EdgeType.CUSTOM_EDGE) {
// disable auto parallelism for bucket map joins
parentRS.getConf().setReducerTraits(EnumSet.of(FIXED));
}
TezEdgeProperty edgeProp = new TezEdgeProperty(null, edgeType, numBuckets);
if (mapJoinWork != null) {
for (BaseWork myWork : mapJoinWork) {
// link the work with the work associated with the reduce sink that triggered this rule
TezWork tezWork = context.currentTask.getWork();
LOG.debug("connecting " + parentWork.getName() + " with " + myWork.getName());
tezWork.connect(parentWork, myWork, edgeProp);
if (edgeType == EdgeType.CUSTOM_EDGE) {
tezWork.setVertexType(myWork, VertexType.INITIALIZED_EDGES);
}
ReduceSinkOperator r = null;
if (context.connectedReduceSinks.contains(parentRS)) {
LOG.debug("Cloning reduce sink " + parentRS + " for multi-child broadcast edge");
// we've already set this one up. Need to clone for the next work.
r = (ReduceSinkOperator) OperatorFactory.getAndMakeChild(parentRS.getCompilationOpContext(), (ReduceSinkDesc) parentRS.getConf().clone(), new RowSchema(parentRS.getSchema()), parentRS.getParentOperators());
context.clonedReduceSinks.add(r);
} else {
r = parentRS;
}
// remember the output name of the reduce sink
r.getConf().setOutputName(myWork.getName());
context.connectedReduceSinks.add(r);
}
}
// remember in case we need to connect additional work later
Map<BaseWork, TezEdgeProperty> linkWorkMap = null;
if (context.linkOpWithWorkMap.containsKey(mapJoinOp)) {
linkWorkMap = context.linkOpWithWorkMap.get(mapJoinOp);
} else {
linkWorkMap = new HashMap<BaseWork, TezEdgeProperty>();
}
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(parentRS.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 (ExprNodeDesc k : keyCols) {
keyOrder.append("+");
keyNullOrder.append(NullOrdering.defaultNullOrder(context.conf).getSign());
}
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);
// at task startup
if (mapJoinWork != null) {
for (BaseWork myWork : mapJoinWork) {
LOG.debug("adding dummy op to work " + myWork.getName() + " from MJ work: " + dummyOp);
myWork.addDummyOp(dummyOp);
}
}
if (context.linkChildOpWithDummyOp.containsKey(mapJoinOp)) {
for (Operator<?> op : context.linkChildOpWithDummyOp.get(mapJoinOp)) {
dummyOperators.add(op);
}
}
context.linkChildOpWithDummyOp.put(mapJoinOp, dummyOperators);
return true;
}
Aggregations