private static void pruneReduceSinkOperator(
boolean[] retainFlags, ReduceSinkOperator reduce, ColumnPrunerProcCtx cppCtx)
throws SemanticException {
ReduceSinkDesc reduceConf = reduce.getConf();
Map<String, ExprNodeDesc> oldMap = reduce.getColumnExprMap();
LOG.info("RS " + reduce.getIdentifier() + " oldColExprMap: " + oldMap);
RowResolver oldRR = cppCtx.getOpToParseCtxMap().get(reduce).getRowResolver();
ArrayList<ColumnInfo> old_signature = oldRR.getRowSchema().getSignature();
ArrayList<ColumnInfo> signature = new ArrayList<ColumnInfo>(old_signature);
List<String> valueColNames = reduceConf.getOutputValueColumnNames();
ArrayList<String> newValueColNames = new ArrayList<String>();
List<ExprNodeDesc> keyExprs = reduceConf.getKeyCols();
List<ExprNodeDesc> valueExprs = reduceConf.getValueCols();
ArrayList<ExprNodeDesc> newValueExprs = new ArrayList<ExprNodeDesc>();
for (int i = 0; i < retainFlags.length; i++) {
String outputCol = valueColNames.get(i);
ExprNodeDesc outputColExpr = valueExprs.get(i);
if (!retainFlags[i]) {
String[] nm = oldRR.reverseLookup(outputCol);
if (nm == null) {
outputCol = Utilities.ReduceField.VALUE.toString() + "." + outputCol;
nm = oldRR.reverseLookup(outputCol);
}
// In case there are multiple columns referenced to the same column name, we won't
// do row resolve once more because the ColumnInfo in row resolver is already removed
if (nm == null) {
continue;
}
// Only remove information of a column if it is not a key,
// i.e. this column is not appearing in keyExprs of the RS
if (ExprNodeDescUtils.indexOf(outputColExpr, keyExprs) == -1) {
ColumnInfo colInfo = oldRR.getFieldMap(nm[0]).remove(nm[1]);
oldRR.getInvRslvMap().remove(colInfo.getInternalName());
oldMap.remove(outputCol);
signature.remove(colInfo);
}
} else {
newValueColNames.add(outputCol);
newValueExprs.add(outputColExpr);
}
}
oldRR.getRowSchema().setSignature(signature);
reduce.getSchema().setSignature(signature);
reduceConf.setOutputValueColumnNames(newValueColNames);
reduceConf.setValueCols(newValueExprs);
TableDesc newValueTable =
PlanUtils.getReduceValueTableDesc(
PlanUtils.getFieldSchemasFromColumnList(
reduceConf.getValueCols(), newValueColNames, 0, ""));
reduceConf.setValueSerializeInfo(newValueTable);
LOG.info("RS " + reduce.getIdentifier() + " newColExprMap: " + oldMap);
}
/**
* Current RSDedup remove/replace child RS. For key columns, sorting order, and the number of
* reducers, copy more specific part of configurations of child RS to that of parent RS. For
* partitioning columns, if both child RS and parent RS have been assigned partitioning columns,
* we will choose the more general partitioning columns. If parent RS has not been assigned any
* partitioning column, we will use partitioning columns (if exist) of child RS.
*/
protected boolean merge(ReduceSinkOperator cRS, ReduceSinkOperator pRS, int minReducer)
throws SemanticException {
int[] result = checkStatus(cRS, pRS, minReducer);
if (result == null) {
return false;
}
if (result[0] > 0) {
// The sorting columns of the child RS are more specific than
// those of the parent RS. Assign sorting columns of the child RS
// to the parent RS.
List<ExprNodeDesc> childKCs = cRS.getConf().getKeyCols();
pRS.getConf().setKeyCols(ExprNodeDescUtils.backtrack(childKCs, cRS, pRS));
}
if (result[1] < 0) {
// The partitioning columns of the parent RS are more specific than
// those of the child RS.
List<ExprNodeDesc> childPCs = cRS.getConf().getPartitionCols();
if (childPCs != null && !childPCs.isEmpty()) {
// If partitioning columns of the child RS are assigned,
// assign these to the partitioning columns of the parent RS.
pRS.getConf().setPartitionCols(ExprNodeDescUtils.backtrack(childPCs, cRS, pRS));
}
} else if (result[1] > 0) {
// The partitioning columns of the child RS are more specific than
// those of the parent RS.
List<ExprNodeDesc> parentPCs = pRS.getConf().getPartitionCols();
if (parentPCs == null || parentPCs.isEmpty()) {
// If partitioning columns of the parent RS are not assigned,
// assign partitioning columns of the child RS to the parent RS.
ArrayList<ExprNodeDesc> childPCs = cRS.getConf().getPartitionCols();
pRS.getConf().setPartitionCols(ExprNodeDescUtils.backtrack(childPCs, cRS, pRS));
}
}
if (result[2] > 0) {
// The sorting order of the child RS is more specific than
// that of the parent RS. Assign the sorting order of the child RS
// to the parent RS.
if (result[0] <= 0) {
// Sorting columns of the parent RS are more specific than those of the
// child RS but Sorting order of the child RS is more specific than
// that of the parent RS.
throw new SemanticException(
"Sorting columns and order don't match. "
+ "Try set "
+ HiveConf.ConfVars.HIVEOPTREDUCEDEDUPLICATION
+ "=false;");
}
pRS.getConf().setOrder(cRS.getConf().getOrder());
}
if (result[3] > 0) {
// The number of reducers of the child RS is more specific than
// that of the parent RS. Assign the number of reducers of the child RS
// to the parent RS.
pRS.getConf().setNumReducers(cRS.getConf().getNumReducers());
}
if (result[4] > 0) {
// This case happens only when pRS key is empty in which case we can use
// number of distribution keys and key serialization info from cRS
pRS.getConf().setNumDistributionKeys(cRS.getConf().getNumDistributionKeys());
List<FieldSchema> fields =
PlanUtils.getFieldSchemasFromColumnList(pRS.getConf().getKeyCols(), "reducesinkkey");
TableDesc keyTable = PlanUtils.getReduceKeyTableDesc(fields, pRS.getConf().getOrder());
ArrayList<String> outputKeyCols = Lists.newArrayList();
for (int i = 0; i < fields.size(); i++) {
outputKeyCols.add(fields.get(i).getName());
}
pRS.getConf().setOutputKeyColumnNames(outputKeyCols);
pRS.getConf().setKeySerializeInfo(keyTable);
}
return true;
}
public ReduceSinkOperator getReduceSinkOp(
List<Integer> partitionPositions,
List<Integer> sortPositions,
List<Integer> sortOrder,
List<Integer> sortNullOrder,
ArrayList<ExprNodeDesc> allCols,
ArrayList<ExprNodeDesc> bucketColumns,
int numBuckets,
Operator<? extends OperatorDesc> parent,
AcidUtils.Operation writeType)
throws SemanticException {
// Order of KEY columns
// 1) Partition columns
// 2) Bucket number column
// 3) Sort columns
Set<Integer> keyColsPosInVal = Sets.newLinkedHashSet();
ArrayList<ExprNodeDesc> keyCols = Lists.newArrayList();
List<Integer> newSortOrder = Lists.newArrayList();
List<Integer> newSortNullOrder = Lists.newArrayList();
int numPartAndBuck = partitionPositions.size();
keyColsPosInVal.addAll(partitionPositions);
if (!bucketColumns.isEmpty()
|| writeType == Operation.DELETE
|| writeType == Operation.UPDATE) {
keyColsPosInVal.add(-1);
numPartAndBuck += 1;
}
keyColsPosInVal.addAll(sortPositions);
// by default partition and bucket columns are sorted in ascending order
Integer order = 1;
if (sortOrder != null && !sortOrder.isEmpty()) {
if (sortOrder.get(0).intValue() == 0) {
order = 0;
}
}
for (int i = 0; i < numPartAndBuck; i++) {
newSortOrder.add(order);
}
newSortOrder.addAll(sortOrder);
String orderStr = "";
for (Integer i : newSortOrder) {
if (i.intValue() == 1) {
orderStr += "+";
} else {
orderStr += "-";
}
}
// if partition and bucket columns are sorted in ascending order, by default
// nulls come first; otherwise nulls come last
Integer nullOrder = order == 1 ? 0 : 1;
if (sortNullOrder != null && !sortNullOrder.isEmpty()) {
if (sortNullOrder.get(0).intValue() == 0) {
nullOrder = 0;
} else {
nullOrder = 1;
}
}
for (int i = 0; i < numPartAndBuck; i++) {
newSortNullOrder.add(nullOrder);
}
newSortNullOrder.addAll(sortNullOrder);
String nullOrderStr = "";
for (Integer i : newSortNullOrder) {
if (i.intValue() == 0) {
nullOrderStr += "a";
} else {
nullOrderStr += "z";
}
}
Map<String, ExprNodeDesc> colExprMap = Maps.newHashMap();
ArrayList<ExprNodeDesc> partCols = Lists.newArrayList();
// we will clone here as RS will update bucket column key with its
// corresponding with bucket number and hence their OIs
for (Integer idx : keyColsPosInVal) {
if (idx < 0) {
ExprNodeConstantDesc bucketNumCol =
new ExprNodeConstantDesc(TypeInfoFactory.stringTypeInfo, BUCKET_NUMBER_COL_NAME);
keyCols.add(bucketNumCol);
colExprMap.put(
Utilities.ReduceField.KEY + ".'" + BUCKET_NUMBER_COL_NAME + "'", bucketNumCol);
} else {
keyCols.add(allCols.get(idx).clone());
}
}
ArrayList<ExprNodeDesc> valCols = Lists.newArrayList();
for (int i = 0; i < allCols.size(); i++) {
if (!keyColsPosInVal.contains(i)) {
valCols.add(allCols.get(i).clone());
}
}
for (Integer idx : partitionPositions) {
partCols.add(allCols.get(idx).clone());
}
// in the absence of SORTED BY clause, the sorted dynamic partition insert
// should honor the ordering of records provided by ORDER BY in SELECT statement
ReduceSinkOperator parentRSOp =
OperatorUtils.findSingleOperatorUpstream(parent, ReduceSinkOperator.class);
if (parentRSOp != null && parseCtx.getQueryProperties().hasOuterOrderBy()) {
String parentRSOpOrder = parentRSOp.getConf().getOrder();
String parentRSOpNullOrder = parentRSOp.getConf().getNullOrder();
if (parentRSOpOrder != null && !parentRSOpOrder.isEmpty() && sortPositions.isEmpty()) {
keyCols.addAll(parentRSOp.getConf().getKeyCols());
orderStr += parentRSOpOrder;
nullOrderStr += parentRSOpNullOrder;
}
}
// map _col0 to KEY._col0, etc
Map<String, String> nameMapping = new HashMap<>();
ArrayList<String> keyColNames = Lists.newArrayList();
for (ExprNodeDesc keyCol : keyCols) {
String keyColName = keyCol.getExprString();
keyColNames.add(keyColName);
colExprMap.put(Utilities.ReduceField.KEY + "." + keyColName, keyCol);
nameMapping.put(keyColName, Utilities.ReduceField.KEY + "." + keyColName);
}
ArrayList<String> valColNames = Lists.newArrayList();
for (ExprNodeDesc valCol : valCols) {
String colName = valCol.getExprString();
valColNames.add(colName);
colExprMap.put(Utilities.ReduceField.VALUE + "." + colName, valCol);
nameMapping.put(colName, Utilities.ReduceField.VALUE + "." + colName);
}
// Create Key/Value TableDesc. When the operator plan is split into MR tasks,
// the reduce operator will initialize Extract operator with information
// from Key and Value TableDesc
List<FieldSchema> fields =
PlanUtils.getFieldSchemasFromColumnList(keyCols, keyColNames, 0, "");
TableDesc keyTable = PlanUtils.getReduceKeyTableDesc(fields, orderStr, nullOrderStr);
List<FieldSchema> valFields =
PlanUtils.getFieldSchemasFromColumnList(valCols, valColNames, 0, "");
TableDesc valueTable = PlanUtils.getReduceValueTableDesc(valFields);
List<List<Integer>> distinctColumnIndices = Lists.newArrayList();
// Number of reducers is set to default (-1)
ReduceSinkDesc rsConf =
new ReduceSinkDesc(
keyCols,
keyCols.size(),
valCols,
keyColNames,
distinctColumnIndices,
valColNames,
-1,
partCols,
-1,
keyTable,
valueTable,
writeType);
rsConf.setBucketCols(bucketColumns);
rsConf.setNumBuckets(numBuckets);
ArrayList<ColumnInfo> signature = new ArrayList<>();
for (int index = 0; index < parent.getSchema().getSignature().size(); index++) {
ColumnInfo colInfo = new ColumnInfo(parent.getSchema().getSignature().get(index));
colInfo.setInternalName(nameMapping.get(colInfo.getInternalName()));
signature.add(colInfo);
}
ReduceSinkOperator op =
(ReduceSinkOperator)
OperatorFactory.getAndMakeChild(rsConf, new RowSchema(signature), parent);
op.setColumnExprMap(colExprMap);
return op;
}
private static void pruneJoinOperator(
NodeProcessorCtx ctx,
CommonJoinOperator op,
JoinDesc conf,
Map<String, ExprNodeDesc> columnExprMap,
Map<Byte, List<Integer>> retainMap,
boolean mapJoin)
throws SemanticException {
ColumnPrunerProcCtx cppCtx = (ColumnPrunerProcCtx) ctx;
List<Operator<? extends OperatorDesc>> childOperators = op.getChildOperators();
LOG.info("JOIN " + op.getIdentifier() + " oldExprs: " + conf.getExprs());
List<String> childColLists = cppCtx.genColLists(op);
if (childColLists == null) {
return;
}
Map<Byte, List<String>> prunedColLists = new HashMap<Byte, List<String>>();
for (byte tag : conf.getTagOrder()) {
prunedColLists.put(tag, new ArrayList<String>());
}
// add the columns in join filters
Set<Map.Entry<Byte, List<ExprNodeDesc>>> filters = conf.getFilters().entrySet();
Iterator<Map.Entry<Byte, List<ExprNodeDesc>>> iter = filters.iterator();
while (iter.hasNext()) {
Map.Entry<Byte, List<ExprNodeDesc>> entry = iter.next();
Byte tag = entry.getKey();
for (ExprNodeDesc desc : entry.getValue()) {
List<String> cols = prunedColLists.get(tag);
cols = Utilities.mergeUniqElems(cols, desc.getCols());
prunedColLists.put(tag, cols);
}
}
RowResolver joinRR = cppCtx.getOpToParseCtxMap().get(op).getRowResolver();
RowResolver newJoinRR = new RowResolver();
ArrayList<String> outputCols = new ArrayList<String>();
ArrayList<ColumnInfo> rs = new ArrayList<ColumnInfo>();
Map<String, ExprNodeDesc> newColExprMap = new HashMap<String, ExprNodeDesc>();
for (int i = 0; i < conf.getOutputColumnNames().size(); i++) {
String internalName = conf.getOutputColumnNames().get(i);
ExprNodeDesc desc = columnExprMap.get(internalName);
Byte tag = conf.getReversedExprs().get(internalName);
if (!childColLists.contains(internalName)) {
int index = conf.getExprs().get(tag).indexOf(desc);
if (index < 0) {
continue;
}
conf.getExprs().get(tag).remove(desc);
if (retainMap != null) {
retainMap.get(tag).remove(index);
}
} else {
List<String> prunedRSList = prunedColLists.get(tag);
if (prunedRSList == null) {
prunedRSList = new ArrayList<String>();
prunedColLists.put(tag, prunedRSList);
}
prunedRSList = Utilities.mergeUniqElems(prunedRSList, desc.getCols());
outputCols.add(internalName);
newColExprMap.put(internalName, desc);
}
}
if (mapJoin) {
// regenerate the valueTableDesc
List<TableDesc> valueTableDescs = new ArrayList<TableDesc>();
for (int pos = 0; pos < op.getParentOperators().size(); pos++) {
List<ExprNodeDesc> valueCols = conf.getExprs().get(Byte.valueOf((byte) pos));
StringBuilder keyOrder = new StringBuilder();
for (int i = 0; i < valueCols.size(); i++) {
keyOrder.append("+");
}
TableDesc valueTableDesc =
PlanUtils.getMapJoinValueTableDesc(
PlanUtils.getFieldSchemasFromColumnList(valueCols, "mapjoinvalue"));
valueTableDescs.add(valueTableDesc);
}
((MapJoinDesc) conf).setValueTblDescs(valueTableDescs);
Set<Map.Entry<Byte, List<ExprNodeDesc>>> exprs = ((MapJoinDesc) conf).getKeys().entrySet();
Iterator<Map.Entry<Byte, List<ExprNodeDesc>>> iters = exprs.iterator();
while (iters.hasNext()) {
Map.Entry<Byte, List<ExprNodeDesc>> entry = iters.next();
List<ExprNodeDesc> lists = entry.getValue();
for (int j = 0; j < lists.size(); j++) {
ExprNodeDesc desc = lists.get(j);
Byte tag = entry.getKey();
List<String> cols = prunedColLists.get(tag);
cols = Utilities.mergeUniqElems(cols, desc.getCols());
prunedColLists.put(tag, cols);
}
}
}
for (Operator<? extends OperatorDesc> child : childOperators) {
if (child instanceof ReduceSinkOperator) {
boolean[] flags =
getPruneReduceSinkOpRetainFlags(childColLists, (ReduceSinkOperator) child);
pruneReduceSinkOperator(flags, (ReduceSinkOperator) child, cppCtx);
}
}
for (int i = 0; i < outputCols.size(); i++) {
String internalName = outputCols.get(i);
String[] nm = joinRR.reverseLookup(internalName);
ColumnInfo col = joinRR.get(nm[0], nm[1]);
newJoinRR.put(nm[0], nm[1], col);
rs.add(col);
}
LOG.info("JOIN " + op.getIdentifier() + " newExprs: " + conf.getExprs());
op.setColumnExprMap(newColExprMap);
conf.setOutputColumnNames(outputCols);
op.getSchema().setSignature(rs);
cppCtx.getOpToParseCtxMap().get(op).setRowResolver(newJoinRR);
cppCtx.getJoinPrunedColLists().put(op, prunedColLists);
}