@Explain(displayName = "condition expressions")
public Map<Byte, String> getExprsStringMap() {
if (getExprs() == null) {
return null;
}
LinkedHashMap<Byte, String> ret = new LinkedHashMap<Byte, String>();
for (Map.Entry<Byte, List<ExprNodeDesc>> ent : getExprs().entrySet()) {
StringBuilder sb = new StringBuilder();
boolean first = true;
if (ent.getValue() != null) {
for (ExprNodeDesc expr : ent.getValue()) {
if (!first) {
sb.append(" ");
}
first = false;
sb.append("{");
sb.append(expr.getExprString());
sb.append("}");
}
}
ret.put(ent.getKey(), sb.toString());
}
return ret;
}
public String getExprString() {
StringBuilder sb = new StringBuilder();
sb.append(genericUDAFName);
sb.append("(");
if (distinct) {
sb.append("DISTINCT ");
}
boolean first = true;
for (ExprNodeDesc exp : parameters) {
if (first) {
first = false;
} else {
sb.append(", ");
}
sb.append(exp.getExprString());
}
sb.append(")");
return sb.toString();
}
/**
* Get the string representation of filters.
*
* <p>Returns null if they are no filters.
*
* @return Map from alias to filters on the alias.
*/
@Explain(displayName = "filter predicates")
public Map<Byte, String> getFiltersStringMap() {
if (getFilters() == null || getFilters().size() == 0) {
return null;
}
LinkedHashMap<Byte, String> ret = new LinkedHashMap<Byte, String>();
boolean filtersPresent = false;
for (Map.Entry<Byte, List<ExprNodeDesc>> ent : getFilters().entrySet()) {
StringBuilder sb = new StringBuilder();
boolean first = true;
if (ent.getValue() != null) {
if (ent.getValue().size() != 0) {
filtersPresent = true;
}
for (ExprNodeDesc expr : ent.getValue()) {
if (!first) {
sb.append(" ");
}
first = false;
sb.append("{");
sb.append(expr.getExprString());
sb.append("}");
}
}
ret.put(ent.getKey(), sb.toString());
}
if (filtersPresent) {
return ret;
} else {
return null;
}
}
@Override
protected void initializeOp(Configuration hconf) throws HiveException {
super.initializeOp(hconf);
try {
numRows = 0;
cntr = 1;
logEveryNRows = HiveConf.getLongVar(hconf, HiveConf.ConfVars.HIVE_LOG_N_RECORDS);
statsMap.put(getCounterName(Counter.RECORDS_OUT_INTERMEDIATE, hconf), recordCounter);
List<ExprNodeDesc> keys = conf.getKeyCols();
if (isLogDebugEnabled) {
LOG.debug("keys size is " + keys.size());
for (ExprNodeDesc k : keys) {
LOG.debug("Key exprNodeDesc " + k.getExprString());
}
}
keyEval = new ExprNodeEvaluator[keys.size()];
int i = 0;
for (ExprNodeDesc e : keys) {
keyEval[i++] = ExprNodeEvaluatorFactory.get(e);
}
numDistributionKeys = conf.getNumDistributionKeys();
distinctColIndices = conf.getDistinctColumnIndices();
numDistinctExprs = distinctColIndices.size();
valueEval = new ExprNodeEvaluator[conf.getValueCols().size()];
i = 0;
for (ExprNodeDesc e : conf.getValueCols()) {
valueEval[i++] = ExprNodeEvaluatorFactory.get(e);
}
partitionEval = new ExprNodeEvaluator[conf.getPartitionCols().size()];
i = 0;
for (ExprNodeDesc e : conf.getPartitionCols()) {
int index = ExprNodeDescUtils.indexOf(e, keys);
partitionEval[i++] = index < 0 ? ExprNodeEvaluatorFactory.get(e) : keyEval[index];
}
if (conf.getBucketCols() != null && !conf.getBucketCols().isEmpty()) {
bucketEval = new ExprNodeEvaluator[conf.getBucketCols().size()];
i = 0;
for (ExprNodeDesc e : conf.getBucketCols()) {
int index = ExprNodeDescUtils.indexOf(e, keys);
bucketEval[i++] = index < 0 ? ExprNodeEvaluatorFactory.get(e) : keyEval[index];
}
buckColIdxInKey = conf.getPartitionCols().size();
}
tag = conf.getTag();
tagByte[0] = (byte) tag;
skipTag = conf.getSkipTag();
if (isLogInfoEnabled) {
LOG.info("Using tag = " + tag);
}
TableDesc keyTableDesc = conf.getKeySerializeInfo();
keySerializer = (Serializer) keyTableDesc.getDeserializerClass().newInstance();
keySerializer.initialize(null, keyTableDesc.getProperties());
keyIsText = keySerializer.getSerializedClass().equals(Text.class);
TableDesc valueTableDesc = conf.getValueSerializeInfo();
valueSerializer = (Serializer) valueTableDesc.getDeserializerClass().newInstance();
valueSerializer.initialize(null, valueTableDesc.getProperties());
int limit = conf.getTopN();
float memUsage = conf.getTopNMemoryUsage();
if (limit >= 0 && memUsage > 0) {
reducerHash = conf.isPTFReduceSink() ? new PTFTopNHash() : reducerHash;
reducerHash.initialize(limit, memUsage, conf.isMapGroupBy(), this);
}
useUniformHash = conf.getReducerTraits().contains(UNIFORM);
firstRow = true;
} catch (Exception e) {
String msg = "Error initializing ReduceSinkOperator: " + e.getMessage();
LOG.error(msg, e);
throw new RuntimeException(e);
}
}
/**
* Converts a filter (which has been pushed down from Hive's optimizer) into corresponding
* restrictions on the HBase scan. The filter should already be in a form which can be fully
* converted.
*
* @param jobConf configuration for the scan
* @param iKey 0-based offset of key column within Hive table
* @return converted table split if any
*/
private Scan createFilterScan(JobConf jobConf, int iKey, boolean isKeyBinary) throws IOException {
// TODO: assert iKey is HBaseSerDe#HBASE_KEY_COL
Scan scan = new Scan();
String filterObjectSerialized = jobConf.get(TableScanDesc.FILTER_OBJECT_CONF_STR);
if (filterObjectSerialized != null) {
HBaseScanRange range =
Utilities.deserializeObject(filterObjectSerialized, HBaseScanRange.class);
try {
range.setup(scan, jobConf);
} catch (Exception e) {
throw new IOException(e);
}
return scan;
}
String filterExprSerialized = jobConf.get(TableScanDesc.FILTER_EXPR_CONF_STR);
if (filterExprSerialized == null) {
return scan;
}
ExprNodeGenericFuncDesc filterExpr = Utilities.deserializeExpression(filterExprSerialized);
String colName = jobConf.get(serdeConstants.LIST_COLUMNS).split(",")[iKey];
String colType = jobConf.get(serdeConstants.LIST_COLUMN_TYPES).split(",")[iKey];
IndexPredicateAnalyzer analyzer = newIndexPredicateAnalyzer(colName, colType, isKeyBinary);
List<IndexSearchCondition> searchConditions = new ArrayList<IndexSearchCondition>();
ExprNodeDesc residualPredicate = analyzer.analyzePredicate(filterExpr, searchConditions);
// There should be no residual since we already negotiated that earlier in
// HBaseStorageHandler.decomposePredicate. However, with hive.optimize.index.filter
// OpProcFactory#pushFilterToStorageHandler pushes the original filter back down again.
// Since pushed-down filters are not ommitted at the higher levels (and thus the
// contract of negotiation is ignored anyway), just ignore the residuals.
// Re-assess this when negotiation is honored and the duplicate evaluation is removed.
// THIS IGNORES RESIDUAL PARSING FROM HBaseStorageHandler#decomposePredicate
if (residualPredicate != null) {
LOG.debug("Ignoring residual predicate " + residualPredicate.getExprString());
}
// Convert the search condition into a restriction on the HBase scan
byte[] startRow = HConstants.EMPTY_START_ROW, stopRow = HConstants.EMPTY_END_ROW;
for (IndexSearchCondition sc : searchConditions) {
ExprNodeConstantEvaluator eval = new ExprNodeConstantEvaluator(sc.getConstantDesc());
PrimitiveObjectInspector objInspector;
Object writable;
try {
objInspector = (PrimitiveObjectInspector) eval.initialize(null);
writable = eval.evaluate(null);
} catch (ClassCastException cce) {
throw new IOException(
"Currently only primitve types are supported. Found: "
+ sc.getConstantDesc().getTypeString());
} catch (HiveException e) {
throw new IOException(e);
}
byte[] constantVal = getConstantVal(writable, objInspector, isKeyBinary);
String comparisonOp = sc.getComparisonOp();
if ("org.apache.hadoop.hive.ql.udf.generic.GenericUDFOPEqual".equals(comparisonOp)) {
startRow = constantVal;
stopRow = getNextBA(constantVal);
} else if ("org.apache.hadoop.hive.ql.udf.generic.GenericUDFOPLessThan"
.equals(comparisonOp)) {
stopRow = constantVal;
} else if ("org.apache.hadoop.hive.ql.udf.generic.GenericUDFOPEqualOrGreaterThan"
.equals(comparisonOp)) {
startRow = constantVal;
} else if ("org.apache.hadoop.hive.ql.udf.generic.GenericUDFOPGreaterThan"
.equals(comparisonOp)) {
startRow = getNextBA(constantVal);
} else if ("org.apache.hadoop.hive.ql.udf.generic.GenericUDFOPEqualOrLessThan"
.equals(comparisonOp)) {
stopRow = getNextBA(constantVal);
} else {
throw new IOException(comparisonOp + " is not a supported comparison operator");
}
}
scan.setStartRow(startRow);
scan.setStopRow(stopRow);
if (LOG.isDebugEnabled()) {
LOG.debug(Bytes.toStringBinary(startRow) + " ~ " + Bytes.toStringBinary(stopRow));
}
return scan;
}
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;
}
@Override
public Object process(
Node nd, Stack<Node> stack, NodeProcessorCtx procCtx, Object... nodeOutputs)
throws SemanticException {
// introduce RS and EX before FS. If the operator tree already contains
// RS then ReduceSinkDeDuplication optimization should merge them
FileSinkOperator fsOp = (FileSinkOperator) nd;
LOG.info("Sorted dynamic partitioning optimization kicked in..");
// if not dynamic partitioning then bail out
if (fsOp.getConf().getDynPartCtx() == null) {
LOG.debug(
"Bailing out of sort dynamic partition optimization as dynamic partitioning context is null");
return null;
}
// if list bucketing then bail out
ListBucketingCtx lbCtx = fsOp.getConf().getLbCtx();
if (lbCtx != null
&& !lbCtx.getSkewedColNames().isEmpty()
&& !lbCtx.getSkewedColValues().isEmpty()) {
LOG.debug(
"Bailing out of sort dynamic partition optimization as list bucketing is enabled");
return null;
}
Table destTable = fsOp.getConf().getTable();
if (destTable == null) {
LOG.debug(
"Bailing out of sort dynamic partition optimization as destination table is null");
return null;
}
// unlink connection between FS and its parent
Operator<? extends OperatorDesc> fsParent = fsOp.getParentOperators().get(0);
// if all dp columns got constant folded then disable this optimization
if (allStaticPartitions(fsParent, fsOp.getConf().getDynPartCtx())) {
LOG.debug(
"Bailing out of sorted dynamic partition optimizer as all dynamic partition"
+ " columns got constant folded (static partitioning)");
return null;
}
// if RS is inserted by enforce bucketing or sorting, we need to remove it
// since ReduceSinkDeDuplication will not merge them to single RS.
// RS inserted by enforce bucketing/sorting will have bucketing column in
// reduce sink key whereas RS inserted by this optimization will have
// partition columns followed by bucket number followed by sort columns in
// the reduce sink key. Since both key columns are not prefix subset
// ReduceSinkDeDuplication will not merge them together resulting in 2 MR jobs.
// To avoid that we will remove the RS (and EX) inserted by enforce bucketing/sorting.
if (!removeRSInsertedByEnforceBucketing(fsOp)) {
LOG.debug(
"Bailing out of sort dynamic partition optimization as some partition columns "
+ "got constant folded.");
return null;
}
// unlink connection between FS and its parent
fsParent = fsOp.getParentOperators().get(0);
fsParent.getChildOperators().clear();
DynamicPartitionCtx dpCtx = fsOp.getConf().getDynPartCtx();
int numBuckets = destTable.getNumBuckets();
// if enforce bucketing/sorting is disabled numBuckets will not be set.
// set the number of buckets here to ensure creation of empty buckets
dpCtx.setNumBuckets(numBuckets);
// Get the positions for partition, bucket and sort columns
List<Integer> bucketPositions =
getBucketPositions(destTable.getBucketCols(), destTable.getCols());
List<Integer> sortPositions = null;
List<Integer> sortOrder = null;
ArrayList<ExprNodeDesc> bucketColumns;
if (fsOp.getConf().getWriteType() == AcidUtils.Operation.UPDATE
|| fsOp.getConf().getWriteType() == AcidUtils.Operation.DELETE) {
// When doing updates and deletes we always want to sort on the rowid because the ACID
// reader will expect this sort order when doing reads. So
// ignore whatever comes from the table and enforce this sort order instead.
sortPositions = Arrays.asList(0);
sortOrder = Arrays.asList(1); // 1 means asc, could really use enum here in the thrift if
bucketColumns =
new ArrayList<>(); // Bucketing column is already present in ROW__ID, which is specially
// handled in ReduceSink
} else {
if (!destTable.getSortCols().isEmpty()) {
// Sort columns specified by table
sortPositions = getSortPositions(destTable.getSortCols(), destTable.getCols());
sortOrder = getSortOrders(destTable.getSortCols(), destTable.getCols());
} else {
// Infer sort columns from operator tree
sortPositions = Lists.newArrayList();
sortOrder = Lists.newArrayList();
inferSortPositions(fsParent, sortPositions, sortOrder);
}
List<ColumnInfo> colInfos = fsParent.getSchema().getSignature();
bucketColumns = getPositionsToExprNodes(bucketPositions, colInfos);
}
List<Integer> sortNullOrder = new ArrayList<Integer>();
for (int order : sortOrder) {
sortNullOrder.add(order == 1 ? 0 : 1); // for asc, nulls first; for desc, nulls last
}
LOG.debug("Got sort order");
for (int i : sortPositions) LOG.debug("sort position " + i);
for (int i : sortOrder) LOG.debug("sort order " + i);
for (int i : sortNullOrder) LOG.debug("sort null order " + i);
List<Integer> partitionPositions = getPartitionPositions(dpCtx, fsParent.getSchema());
// update file sink descriptor
fsOp.getConf().setMultiFileSpray(false);
fsOp.getConf().setNumFiles(1);
fsOp.getConf().setTotalFiles(1);
ArrayList<ColumnInfo> parentCols = Lists.newArrayList(fsParent.getSchema().getSignature());
ArrayList<ExprNodeDesc> allRSCols = Lists.newArrayList();
for (ColumnInfo ci : parentCols) {
allRSCols.add(new ExprNodeColumnDesc(ci));
}
// Create ReduceSink operator
ReduceSinkOperator rsOp =
getReduceSinkOp(
partitionPositions,
sortPositions,
sortOrder,
sortNullOrder,
allRSCols,
bucketColumns,
numBuckets,
fsParent,
fsOp.getConf().getWriteType());
List<ExprNodeDesc> descs = new ArrayList<ExprNodeDesc>(allRSCols.size());
List<String> colNames = new ArrayList<String>();
String colName;
for (int i = 0; i < allRSCols.size(); i++) {
ExprNodeDesc col = allRSCols.get(i);
colName = col.getExprString();
colNames.add(colName);
if (partitionPositions.contains(i) || sortPositions.contains(i)) {
descs.add(
new ExprNodeColumnDesc(
col.getTypeInfo(), ReduceField.KEY.toString() + "." + colName, null, false));
} else {
descs.add(
new ExprNodeColumnDesc(
col.getTypeInfo(), ReduceField.VALUE.toString() + "." + colName, null, false));
}
}
RowSchema selRS = new RowSchema(fsParent.getSchema());
if (!bucketColumns.isEmpty()
|| fsOp.getConf().getWriteType() == Operation.DELETE
|| fsOp.getConf().getWriteType() == Operation.UPDATE) {
descs.add(
new ExprNodeColumnDesc(
TypeInfoFactory.stringTypeInfo,
ReduceField.KEY.toString() + ".'" + BUCKET_NUMBER_COL_NAME + "'",
null,
false));
colNames.add("'" + BUCKET_NUMBER_COL_NAME + "'");
ColumnInfo ci =
new ColumnInfo(
BUCKET_NUMBER_COL_NAME,
TypeInfoFactory.stringTypeInfo,
selRS.getSignature().get(0).getTabAlias(),
true,
true);
selRS.getSignature().add(ci);
fsParent.getSchema().getSignature().add(ci);
}
// Create SelectDesc
SelectDesc selConf = new SelectDesc(descs, colNames);
// Create Select Operator
SelectOperator selOp = (SelectOperator) OperatorFactory.getAndMakeChild(selConf, selRS, rsOp);
// link SEL to FS
fsOp.getParentOperators().clear();
fsOp.getParentOperators().add(selOp);
selOp.getChildOperators().add(fsOp);
// Set if partition sorted or partition bucket sorted
fsOp.getConf().setDpSortState(FileSinkDesc.DPSortState.PARTITION_SORTED);
if (bucketColumns.size() > 0
|| fsOp.getConf().getWriteType() == Operation.DELETE
|| fsOp.getConf().getWriteType() == Operation.UPDATE) {
fsOp.getConf().setDpSortState(FileSinkDesc.DPSortState.PARTITION_BUCKET_SORTED);
}
// update partition column info in FS descriptor
fsOp.getConf().setPartitionCols(rsOp.getConf().getPartitionCols());
LOG.info(
"Inserted "
+ rsOp.getOperatorId()
+ " and "
+ selOp.getOperatorId()
+ " as parent of "
+ fsOp.getOperatorId()
+ " and child of "
+ fsParent.getOperatorId());
parseCtx.setReduceSinkAddedBySortedDynPartition(true);
return null;
}