/** Reduce Sink Operator sends output to the reduce stage. */
public class ReduceSinkOperator extends TerminalOperator<ReduceSinkDesc>
implements Serializable, TopNHash.BinaryCollector {
/** Counters. */
public static enum Counter {
RECORDS_OUT_INTERMEDIATE
}
private static final long serialVersionUID = 1L;
private static final MurmurHash hash = (MurmurHash) MurmurHash.getInstance();
private transient ObjectInspector[] partitionObjectInspectors;
private transient ObjectInspector[] bucketObjectInspectors;
private transient int buckColIdxInKey;
private boolean firstRow;
private transient int tag;
private boolean skipTag = false;
private transient InspectableObject tempInspectableObject = new InspectableObject();
private transient int[] valueIndex; // index for value(+ from keys, - from values)
protected transient OutputCollector out;
/**
* The evaluators for the key columns. Key columns decide the sort order on the reducer side. Key
* columns are passed to the reducer in the "key".
*/
protected transient ExprNodeEvaluator[] keyEval;
/** The evaluators for the value columns. Value columns are passed to reducer in the "value". */
protected transient ExprNodeEvaluator[] valueEval;
/**
* The evaluators for the partition columns (CLUSTER BY or DISTRIBUTE BY in Hive language).
* Partition columns decide the reducer that the current row goes to. Partition columns are not
* passed to reducer.
*/
protected transient ExprNodeEvaluator[] partitionEval;
/** Evaluators for bucketing columns. This is used to compute bucket number. */
protected transient ExprNodeEvaluator[] bucketEval = null;
// TODO: we use MetadataTypedColumnsetSerDe for now, till DynamicSerDe is ready
protected transient Serializer keySerializer;
protected transient boolean keyIsText;
protected transient Serializer valueSerializer;
protected transient byte[] tagByte = new byte[1];
protected transient int numDistributionKeys;
protected transient int numDistinctExprs;
protected transient String[] inputAliases; // input aliases of this RS for join (used for PPD)
protected transient boolean useUniformHash = false;
// picks topN K:V pairs from input.
protected transient TopNHash reducerHash = new TopNHash();
protected transient HiveKey keyWritable = new HiveKey();
protected transient ObjectInspector keyObjectInspector;
protected transient ObjectInspector valueObjectInspector;
protected transient Object[] cachedValues;
protected transient List<List<Integer>> distinctColIndices;
protected transient Random random;
/**
* This two dimensional array holds key data and a corresponding Union object which contains the
* tag identifying the aggregate expression for distinct columns.
*
* <p>If there is no distict expression, cachedKeys is simply like this. cachedKeys[0] =
* [col0][col1]
*
* <p>with two distict expression, union(tag:key) is attatched for each distinct expression
* cachedKeys[0] = [col0][col1][0:dist1] cachedKeys[1] = [col0][col1][1:dist2]
*
* <p>in this case, child GBY evaluates distict values with expression like KEY.col2:0.dist1 see
* {@link ExprNodeColumnEvaluator}
*/
// TODO: we only ever use one row of these at a time. Why do we need to cache multiple?
protected transient Object[][] cachedKeys;
private StructField recIdField; // field to look for record identifier in
private StructField bucketField; // field to look for bucket in record identifier
private StructObjectInspector acidRowInspector; // row inspector used by acid options
private StructObjectInspector recIdInspector; // OI for the record identifier
private IntObjectInspector bucketInspector; // OI for the bucket field in the record id
protected transient long numRows = 0;
protected transient long cntr = 1;
protected transient long logEveryNRows = 0;
private final transient LongWritable recordCounter = new LongWritable();
/** Kryo ctor. */
protected ReduceSinkOperator() {
super();
}
public ReduceSinkOperator(CompilationOpContext ctx) {
super(ctx);
}
@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);
}
}
public String getCounterName(Counter counter, Configuration hconf) {
String context = hconf.get(Operator.CONTEXT_NAME_KEY, "");
if (context != null && !context.isEmpty()) {
context = "_" + context.replace(" ", "_");
}
return counter + context;
}
/**
* Initializes array of ExprNodeEvaluator. Adds Union field for distinct column indices for group
* by. Puts the return values into a StructObjectInspector with output column names.
*
* <p>If distinctColIndices is empty, the object inspector is same as {@link
* Operator#initEvaluatorsAndReturnStruct(ExprNodeEvaluator[], List, ObjectInspector)}
*/
protected static StructObjectInspector initEvaluatorsAndReturnStruct(
ExprNodeEvaluator[] evals,
List<List<Integer>> distinctColIndices,
List<String> outputColNames,
int length,
ObjectInspector rowInspector)
throws HiveException {
int inspectorLen = evals.length > length ? length + 1 : evals.length;
List<ObjectInspector> sois = new ArrayList<ObjectInspector>(inspectorLen);
// keys
ObjectInspector[] fieldObjectInspectors = initEvaluators(evals, 0, length, rowInspector);
sois.addAll(Arrays.asList(fieldObjectInspectors));
if (outputColNames.size() > length) {
// union keys
assert distinctColIndices != null;
List<ObjectInspector> uois = new ArrayList<ObjectInspector>();
for (List<Integer> distinctCols : distinctColIndices) {
List<String> names = new ArrayList<String>();
List<ObjectInspector> eois = new ArrayList<ObjectInspector>();
int numExprs = 0;
for (int i : distinctCols) {
names.add(HiveConf.getColumnInternalName(numExprs));
eois.add(evals[i].initialize(rowInspector));
numExprs++;
}
uois.add(ObjectInspectorFactory.getStandardStructObjectInspector(names, eois));
}
UnionObjectInspector uoi = ObjectInspectorFactory.getStandardUnionObjectInspector(uois);
sois.add(uoi);
}
return ObjectInspectorFactory.getStandardStructObjectInspector(outputColNames, sois);
}
@Override
@SuppressWarnings("unchecked")
public void process(Object row, int tag) throws HiveException {
try {
ObjectInspector rowInspector = inputObjInspectors[tag];
if (firstRow) {
firstRow = false;
// TODO: this is fishy - we init object inspectors based on first tag. We
// should either init for each tag, or if rowInspector doesn't really
// matter, then we can create this in ctor and get rid of firstRow.
if (conf.getWriteType() == AcidUtils.Operation.UPDATE
|| conf.getWriteType() == AcidUtils.Operation.DELETE) {
assert rowInspector instanceof StructObjectInspector
: "Exptected rowInspector to be instance of StructObjectInspector but it is a "
+ rowInspector.getClass().getName();
acidRowInspector = (StructObjectInspector) rowInspector;
// The record identifier is always in the first column
recIdField = acidRowInspector.getAllStructFieldRefs().get(0);
recIdInspector = (StructObjectInspector) recIdField.getFieldObjectInspector();
// The bucket field is in the second position
bucketField = recIdInspector.getAllStructFieldRefs().get(1);
bucketInspector = (IntObjectInspector) bucketField.getFieldObjectInspector();
}
if (isLogInfoEnabled) {
LOG.info(
"keys are "
+ conf.getOutputKeyColumnNames()
+ " num distributions: "
+ conf.getNumDistributionKeys());
}
keyObjectInspector =
initEvaluatorsAndReturnStruct(
keyEval,
distinctColIndices,
conf.getOutputKeyColumnNames(),
numDistributionKeys,
rowInspector);
valueObjectInspector =
initEvaluatorsAndReturnStruct(
valueEval, conf.getOutputValueColumnNames(), rowInspector);
partitionObjectInspectors = initEvaluators(partitionEval, rowInspector);
if (bucketEval != null) {
bucketObjectInspectors = initEvaluators(bucketEval, rowInspector);
}
int numKeys = numDistinctExprs > 0 ? numDistinctExprs : 1;
int keyLen = numDistinctExprs > 0 ? numDistributionKeys + 1 : numDistributionKeys;
cachedKeys = new Object[numKeys][keyLen];
cachedValues = new Object[valueEval.length];
}
// Determine distKeyLength (w/o distincts), and then add the first if present.
populateCachedDistributionKeys(row, 0);
// replace bucketing columns with hashcode % numBuckets
int bucketNumber = -1;
if (bucketEval != null) {
bucketNumber = computeBucketNumber(row, conf.getNumBuckets());
cachedKeys[0][buckColIdxInKey] = new Text(String.valueOf(bucketNumber));
} else if (conf.getWriteType() == AcidUtils.Operation.UPDATE
|| conf.getWriteType() == AcidUtils.Operation.DELETE) {
// In the non-partitioned case we still want to compute the bucket number for updates and
// deletes.
bucketNumber = computeBucketNumber(row, conf.getNumBuckets());
}
HiveKey firstKey = toHiveKey(cachedKeys[0], tag, null);
int distKeyLength = firstKey.getDistKeyLength();
if (numDistinctExprs > 0) {
populateCachedDistinctKeys(row, 0);
firstKey = toHiveKey(cachedKeys[0], tag, distKeyLength);
}
final int hashCode;
// distKeyLength doesn't include tag, but includes buckNum in cachedKeys[0]
if (useUniformHash && partitionEval.length > 0) {
hashCode = computeMurmurHash(firstKey);
} else {
hashCode = computeHashCode(row, bucketNumber);
}
firstKey.setHashCode(hashCode);
/*
* in case of TopN for windowing, we need to distinguish between rows with
* null partition keys and rows with value 0 for partition keys.
*/
boolean partKeyNull = conf.isPTFReduceSink() && partitionKeysAreNull(row);
// Try to store the first key. If it's not excluded, we will proceed.
int firstIndex = reducerHash.tryStoreKey(firstKey, partKeyNull);
if (firstIndex == TopNHash.EXCLUDE) return; // Nothing to do.
// Compute value and hashcode - we'd either store or forward them.
BytesWritable value = makeValueWritable(row);
if (firstIndex == TopNHash.FORWARD) {
collect(firstKey, value);
} else {
assert firstIndex >= 0;
reducerHash.storeValue(firstIndex, firstKey.hashCode(), value, false);
}
// All other distinct keys will just be forwarded. This could be optimized...
for (int i = 1; i < numDistinctExprs; i++) {
System.arraycopy(cachedKeys[0], 0, cachedKeys[i], 0, numDistributionKeys);
populateCachedDistinctKeys(row, i);
HiveKey hiveKey = toHiveKey(cachedKeys[i], tag, distKeyLength);
hiveKey.setHashCode(hashCode);
collect(hiveKey, value);
}
} catch (HiveException e) {
throw e;
} catch (Exception e) {
throw new HiveException(e);
}
}
private int computeBucketNumber(Object row, int numBuckets) throws HiveException {
if (conf.getWriteType() == AcidUtils.Operation.UPDATE
|| conf.getWriteType() == AcidUtils.Operation.DELETE) {
// We don't need to evaluate the hash code. Instead read the bucket number directly from
// the row. I don't need to evaluate any expressions as I know I am reading the ROW__ID
// column directly.
Object recIdValue = acidRowInspector.getStructFieldData(row, recIdField);
int buckNum = bucketInspector.get(recIdInspector.getStructFieldData(recIdValue, bucketField));
if (isLogTraceEnabled) {
LOG.trace("Acid choosing bucket number " + buckNum);
}
return buckNum;
} else {
Object[] bucketFieldValues = new Object[bucketEval.length];
for (int i = 0; i < bucketEval.length; i++) {
bucketFieldValues[i] = bucketEval[i].evaluate(row);
}
return ObjectInspectorUtils.getBucketNumber(
bucketFieldValues, bucketObjectInspectors, numBuckets);
}
}
private void populateCachedDistributionKeys(Object row, int index) throws HiveException {
for (int i = 0; i < numDistributionKeys; i++) {
cachedKeys[index][i] = keyEval[i].evaluate(row);
}
if (cachedKeys[0].length > numDistributionKeys) {
cachedKeys[index][numDistributionKeys] = null;
}
}
/**
* Populate distinct keys part of cachedKeys for a particular row.
*
* @param row the row
* @param index the cachedKeys index to write to
*/
private void populateCachedDistinctKeys(Object row, int index) throws HiveException {
StandardUnion union;
cachedKeys[index][numDistributionKeys] =
union = new StandardUnion((byte) index, new Object[distinctColIndices.get(index).size()]);
Object[] distinctParameters = (Object[]) union.getObject();
for (int distinctParamI = 0; distinctParamI < distinctParameters.length; distinctParamI++) {
distinctParameters[distinctParamI] =
keyEval[distinctColIndices.get(index).get(distinctParamI)].evaluate(row);
}
union.setTag((byte) index);
}
protected final int computeMurmurHash(HiveKey firstKey) {
return hash.hash(firstKey.getBytes(), firstKey.getDistKeyLength(), 0);
}
private int computeHashCode(Object row, int buckNum) throws HiveException {
// Evaluate the HashCode
int keyHashCode = 0;
if (partitionEval.length == 0) {
// If no partition cols and not doing an update or delete, just distribute the data uniformly
// to provide better load balance. If the requirement is to have a single reducer, we should
// set the number of reducers to 1. Use a constant seed to make the code deterministic.
// For acid operations make sure to send all records with the same key to the same
// FileSinkOperator, as the RecordUpdater interface can't manage multiple writers for a file.
if (conf.getWriteType() == AcidUtils.Operation.NOT_ACID) {
if (random == null) {
random = new Random(12345);
}
keyHashCode = random.nextInt();
} else {
keyHashCode = 1;
}
} else {
Object[] bucketFieldValues = new Object[partitionEval.length];
for (int i = 0; i < partitionEval.length; i++) {
bucketFieldValues[i] = partitionEval[i].evaluate(row);
}
keyHashCode =
ObjectInspectorUtils.getBucketHashCode(bucketFieldValues, partitionObjectInspectors);
}
int hashCode = buckNum < 0 ? keyHashCode : keyHashCode * 31 + buckNum;
if (isLogTraceEnabled) {
LOG.trace("Going to return hash code " + hashCode);
}
return hashCode;
}
private boolean partitionKeysAreNull(Object row) throws HiveException {
if (partitionEval.length != 0) {
for (int i = 0; i < partitionEval.length; i++) {
Object o = partitionEval[i].evaluate(row);
if (o != null) {
return false;
}
}
return true;
}
return false;
}
// Serialize the keys and append the tag
protected HiveKey toHiveKey(Object obj, int tag, Integer distLength) throws SerDeException {
BinaryComparable key = (BinaryComparable) keySerializer.serialize(obj, keyObjectInspector);
int keyLength = key.getLength();
if (tag == -1 || skipTag) {
keyWritable.set(key.getBytes(), 0, keyLength);
} else {
keyWritable.setSize(keyLength + 1);
System.arraycopy(key.getBytes(), 0, keyWritable.get(), 0, keyLength);
keyWritable.get()[keyLength] = tagByte[0];
}
keyWritable.setDistKeyLength((distLength == null) ? keyLength : distLength);
return keyWritable;
}
@Override
public void collect(byte[] key, byte[] value, int hash) throws IOException {
HiveKey keyWritable = new HiveKey(key, hash);
BytesWritable valueWritable = new BytesWritable(value);
collect(keyWritable, valueWritable);
}
protected void collect(BytesWritable keyWritable, Writable valueWritable) throws IOException {
// Since this is a terminal operator, update counters explicitly -
// forward is not called
if (null != out) {
numRows++;
if (isLogInfoEnabled) {
if (numRows == cntr) {
cntr = logEveryNRows == 0 ? cntr * 10 : numRows + logEveryNRows;
if (cntr < 0 || numRows < 0) {
cntr = 0;
numRows = 1;
}
LOG.info(toString() + ": records written - " + numRows);
}
}
out.collect(keyWritable, valueWritable);
}
}
private BytesWritable makeValueWritable(Object row) throws Exception {
int length = valueEval.length;
// Evaluate the value
for (int i = 0; i < length; i++) {
cachedValues[i] = valueEval[i].evaluate(row);
}
// Serialize the value
return (BytesWritable) valueSerializer.serialize(cachedValues, valueObjectInspector);
}
@Override
protected void closeOp(boolean abort) throws HiveException {
if (!abort) {
reducerHash.flush();
}
super.closeOp(abort);
out = null;
if (isLogInfoEnabled) {
LOG.info(toString() + ": records written - " + numRows);
}
recordCounter.set(numRows);
}
/** @return the name of the operator */
@Override
public String getName() {
return getOperatorName();
}
public static String getOperatorName() {
return "RS";
}
@Override
public OperatorType getType() {
return OperatorType.REDUCESINK;
}
@Override
public boolean opAllowedBeforeMapJoin() {
return false;
}
public void setSkipTag(boolean value) {
this.skipTag = value;
}
public void setValueIndex(int[] valueIndex) {
this.valueIndex = valueIndex;
}
public int[] getValueIndex() {
return valueIndex;
}
public void setInputAliases(String[] inputAliases) {
this.inputAliases = inputAliases;
}
public String[] getInputAliases() {
return inputAliases;
}
@Override
public boolean getIsReduceSink() {
return true;
}
@Override
public String getReduceOutputName() {
return conf.getOutputName();
}
@Override
public void setOutputCollector(OutputCollector _out) {
this.out = _out;
}
}
protected final int computeMurmurHash(HiveKey firstKey) {
return hash.hash(firstKey.getBytes(), firstKey.getDistKeyLength(), 0);
}
