[
https://issues.apache.org/jira/browse/HIVE-24471?focusedWorklogId=522535&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-522535
]
ASF GitHub Bot logged work on HIVE-24471:
-----------------------------------------
Author: ASF GitHub Bot
Created on: 10/Dec/20 04:39
Start Date: 10/Dec/20 04:39
Worklog Time Spent: 10m
Work Description: t3rmin4t0r commented on a change in pull request #1736:
URL: https://github.com/apache/hive/pull/1736#discussion_r539838422
##########
File path: ql/src/java/org/apache/hadoop/hive/ql/exec/GroupByCombiner.java
##########
@@ -0,0 +1,246 @@
+/*
+ * 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.exec;
+
+import org.apache.hadoop.hive.conf.HiveConf;
+import org.apache.hadoop.hive.ql.exec.vector.VectorGroupByCombiner;
+import org.apache.hadoop.hive.ql.exec.vector.VectorGroupByOperator;
+import org.apache.hadoop.mapred.JobConf;
+import org.apache.hadoop.hive.ql.metadata.HiveException;
+import org.apache.hadoop.hive.ql.plan.BaseWork;
+import org.apache.hadoop.hive.ql.plan.GroupByDesc;
+import org.apache.hadoop.hive.ql.plan.ReduceWork;
+import org.apache.hadoop.hive.ql.plan.TableDesc;
+import org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator;
+import org.apache.hadoop.hive.serde2.AbstractSerDe;
+import org.apache.hadoop.hive.serde2.Deserializer;
+import org.apache.hadoop.hive.serde2.SerDeException;
+import org.apache.hadoop.hive.serde2.SerDeUtils;
+import org.apache.hadoop.hive.serde2.Serializer;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorFactory;
+import org.apache.hadoop.io.BytesWritable;
+import org.apache.hadoop.io.DataInputBuffer;
+import org.apache.hadoop.util.ReflectionUtils;
+import org.apache.tez.runtime.api.TaskContext;
+import org.apache.tez.runtime.library.common.sort.impl.IFile;
+import org.apache.tez.runtime.library.common.sort.impl.TezRawKeyValueIterator;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import org.apache.hadoop.fs.Path;
+
+import java.io.IOException;
+import java.util.ArrayList;
+
+import static org.apache.hadoop.hive.ql.exec.Utilities.HAS_REDUCE_WORK;
+import static org.apache.hadoop.hive.ql.exec.Utilities.REDUCE_PLAN_NAME;
+
+// Combiner for normal group by operator. In case of map side aggregate, the
partially
+// aggregated records are sorted based on group by key. If because of some
reasons, like hash
+// table memory exceeded the limit or the first few batches of records have
less ndvs, the
+// aggregation is not done, then here the aggregation can be done cheaply as
the records
+// are sorted based on group by key.
+public class GroupByCombiner extends VectorGroupByCombiner {
+
+ private static final Logger LOG = LoggerFactory.getLogger(
+ org.apache.hadoop.hive.ql.exec.GroupByCombiner.class.getName());
+
+ private transient GenericUDAFEvaluator[] aggregationEvaluators;
+ Deserializer valueDeserializer;
+ GenericUDAFEvaluator.AggregationBuffer[] aggregationBuffers;
+ GroupByOperator groupByOperator;
+ Serializer valueSerializer;
+ ObjectInspector aggrObjectInspector;
+ DataInputBuffer valueBuffer;
+ Object[] cachedValues;
+
+ public GroupByCombiner(TaskContext taskContext) throws HiveException,
IOException {
+ super(taskContext);
+ if (rw != null) {
+ try {
+ groupByOperator = (GroupByOperator) rw.getReducer();
+
+ ArrayList<ObjectInspector> ois = new ArrayList<ObjectInspector>();
+ ois.add(keyObjectInspector);
+ ois.add(valueObjectInspector);
+ ObjectInspector[] rowObjectInspector = new ObjectInspector[1];
+ rowObjectInspector[0] =
+
ObjectInspectorFactory.getStandardStructObjectInspector(Utilities.reduceFieldNameList,
+ ois);
+ groupByOperator.setInputObjInspectors(rowObjectInspector);
+ groupByOperator.initializeOp(conf);
+ aggregationBuffers = groupByOperator.getAggregationBuffers();
+ aggregationEvaluators = groupByOperator.getAggregationEvaluator();
+
+ TableDesc valueTableDesc = rw.getTagToValueDesc().get(0);
+ valueSerializer = (Serializer) valueTableDesc.getDeserializerClass()
+ .newInstance();
+ valueSerializer.initialize(null, valueTableDesc.getProperties());
+
+ valueDeserializer = (AbstractSerDe) ReflectionUtils.newInstance(
+ valueTableDesc.getDeserializerClass(), null);
+ SerDeUtils.initializeSerDe(valueDeserializer, null,
+ valueTableDesc.getProperties(), null);
+
+ aggrObjectInspector = groupByOperator.getAggrObjInspector();
+ valueBuffer = new DataInputBuffer();
+ cachedValues = new Object[aggregationEvaluators.length];
+ } catch (Exception e) {
+ LOG.error(" GroupByCombiner failed", e);
+ throw new RuntimeException(e.getMessage());
+ }
+ }
+ }
+
+ private void processAggregation(IFile.Writer writer, DataInputBuffer key)
+ throws Exception {
+ for (int i = 0; i < aggregationEvaluators.length; i++) {
+ cachedValues[i] =
aggregationEvaluators[i].evaluate(aggregationBuffers[i]);
+ }
+ BytesWritable result = (BytesWritable)
valueSerializer.serialize(cachedValues,
+ aggrObjectInspector);
+ valueBuffer.reset(result.getBytes(), result.getLength());
+ writer.append(key, valueBuffer);
+ combineOutputRecordsCounter.increment(1);
+ for (int i = 0; i < aggregationEvaluators.length; i++) {
+ aggregationEvaluators[i].reset(aggregationBuffers[i]);
+ }
+ }
+
+ private void updateAggregation(BytesWritable valWritable, DataInputBuffer
value)
+ throws HiveException, SerDeException {
+ valWritable.set(value.getData(), value.getPosition(),
+ value.getLength() - value.getPosition());
+ Object row = valueDeserializer.deserialize(valWritable);
+ groupByOperator.updateAggregation(row);
+ }
+
+ private void processRows(TezRawKeyValueIterator rawIter, IFile.Writer
writer) {
+ long numRows = 0;
+ try {
+ DataInputBuffer key = rawIter.getKey();
+ DataInputBuffer prevKey = new DataInputBuffer();
+ prevKey.reset(key.getData(), key.getPosition(), key.getLength() -
key.getPosition());
+ BytesWritable valWritable = new BytesWritable();
+ do {
+ key = rawIter.getKey();
+ // For first iteration, prevKey is always same as key.
+ if (VectorGroupByCombiner.compare(key, prevKey) != 0) {
+ processAggregation(writer, prevKey);
+ prevKey.reset(key.getData(), key.getPosition(), key.getLength() -
key.getPosition());
+ }
+ updateAggregation(valWritable, rawIter.getValue());
+ numRows++;
+ } while (rawIter.next());
Review comment:
We only need to go into deserialization if there's >1 values for 1 key,
right?
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByCombiner.java
##########
@@ -0,0 +1,377 @@
+/*
+ * 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.exec.vector;
+
+import org.apache.hadoop.conf.Configuration;
+import org.apache.hadoop.fs.Path;
+import org.apache.hadoop.hive.conf.HiveConf;
+import org.apache.hadoop.hive.ql.exec.Utilities;
+import org.apache.hadoop.hive.ql.exec.mr.ExecReducer;
+import
org.apache.hadoop.hive.ql.exec.vector.expressions.aggregates.VectorAggregateExpression;
+import org.apache.hadoop.hive.ql.metadata.HiveException;
+import org.apache.hadoop.hive.ql.plan.ReduceWork;
+import org.apache.hadoop.hive.ql.plan.TableDesc;
+import org.apache.hadoop.hive.serde2.AbstractSerDe;
+import org.apache.hadoop.hive.serde2.ByteStream;
+import org.apache.hadoop.hive.serde2.Deserializer;
+import org.apache.hadoop.hive.serde2.SerDeException;
+import org.apache.hadoop.hive.serde2.SerDeUtils;
+import org.apache.hadoop.hive.serde2.lazybinary.fast.LazyBinaryDeserializeRead;
+import org.apache.hadoop.hive.serde2.lazybinary.fast.LazyBinarySerializeWrite;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
+import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;
+import org.apache.hadoop.io.DataInputBuffer;
+import org.apache.hadoop.mapreduce.TaskCounter;
+import org.apache.hadoop.util.ReflectionUtils;
+import org.apache.hadoop.util.StringUtils;
+import org.apache.tez.common.TezUtils;
+import org.apache.tez.common.counters.TezCounter;
+import org.apache.tez.mapreduce.combine.MRCombiner;
+import org.apache.tez.runtime.api.TaskContext;
+import org.apache.tez.runtime.library.common.sort.impl.IFile;
+import org.apache.tez.runtime.library.common.sort.impl.TezRawKeyValueIterator;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import java.io.IOException;
+
+import static org.apache.hadoop.hive.ql.exec.Utilities.HAS_REDUCE_WORK;
+import static org.apache.hadoop.hive.ql.exec.Utilities.MAPRED_REDUCER_CLASS;
+import static org.apache.hadoop.hive.ql.exec.Utilities.REDUCE_PLAN_NAME;
+import static
org.apache.hadoop.hive.serde2.lazy.fast.LazySimpleDeserializeRead.byteArrayCompareRanges;
+
+// Combiner for vectorized group by operator. In case of map side aggregate,
the partially
+// aggregated records are sorted based on group by key. If because of some
reasons, like hash
+// table memory exceeded the limit or the first few batches of records have
less ndvs, the
+// aggregation is not done, then here the aggregation can be done cheaply as
the records
+// are sorted based on group by key.
+public class VectorGroupByCombiner extends MRCombiner {
+ private static final Logger LOG = LoggerFactory.getLogger(
+ VectorGroupByCombiner.class.getName());
+ protected final Configuration conf;
+ protected final TezCounter combineInputRecordsCounter;
+ protected final TezCounter combineOutputRecordsCounter;
+ VectorAggregateExpression[] aggregators;
+ VectorAggregationBufferRow aggregationBufferRow;
+ protected transient LazyBinarySerializeWrite valueLazyBinarySerializeWrite;
+
+ // This helper object serializes LazyBinary format reducer values from
columns of a row
+ // in a vectorized row batch.
+ protected transient VectorSerializeRow<LazyBinarySerializeWrite>
valueVectorSerializeRow;
+
+ // The output buffer used to serialize a value into.
+ protected transient ByteStream.Output valueOutput;
+ DataInputBuffer valueBytesWritable;
+
+ // Only required minimal configs are copied to the worker nodes. This hack
(file.) is
+ // done to include these configs to be copied to the worker node.
+ protected static String confPrefixForWorker = "file.";
+
+ VectorDeserializeRow<LazyBinaryDeserializeRead> batchValueDeserializer;
+ int firstValueColumnOffset;
+ VectorizedRowBatchCtx batchContext = null;
+ int numValueCol = 0;
+ protected ReduceWork rw;
+ VectorizedRowBatch outputBatch = null;
+ VectorizedRowBatch inputBatch = null;
+ protected Deserializer inputKeyDeserializer = null;
+ protected ObjectInspector keyObjectInspector = null;
+ protected ObjectInspector valueObjectInspector = null;
+ protected StructObjectInspector valueStructInspectors = null;
+ protected StructObjectInspector keyStructInspector = null;
+
+ public VectorGroupByCombiner(TaskContext taskContext) throws HiveException,
IOException {
+ super(taskContext);
+
+ combineInputRecordsCounter =
+
taskContext.getCounters().findCounter(TaskCounter.COMBINE_INPUT_RECORDS);
+ combineOutputRecordsCounter =
+
taskContext.getCounters().findCounter(TaskCounter.COMBINE_OUTPUT_RECORDS);
+
+ conf = TezUtils.createConfFromUserPayload(taskContext.getUserPayload());
+ rw = getReduceWork();
+ if (rw == null) {
+ return;
+ }
+
+ if (rw.getReducer() instanceof VectorGroupByOperator) {
+ VectorGroupByOperator vectorGroupByOperator = (VectorGroupByOperator)
rw.getReducer();
+ vectorGroupByOperator.initializeOp(this.conf);
+ this.aggregators = vectorGroupByOperator.getAggregators();
+ this.aggregationBufferRow = allocateAggregationBuffer();
+ batchContext = rw.getVectorizedRowBatchCtx();
+ }
+
+ try {
+ initObjectInspectors(rw.getTagToValueDesc().get(0), rw.getKeyDesc());
+ if (batchContext != null && numValueCol > 0) {
+ initVectorBatches();
+ }
+ } catch (SerDeException e) {
+ LOG.error("Fail to initialize VectorGroupByCombiner.", e);
+ throw new RuntimeException(e.getCause());
+ }
+ }
+
+ // Get the reduce work from the config. Here some hack is used to prefix the
config name with
+ // "file." to avoid the config being filtered out.
+ private ReduceWork getReduceWork() {
+ String plan = conf.get(confPrefixForWorker +
HiveConf.ConfVars.PLAN.varname);
+ this.conf.set(HiveConf.ConfVars.PLAN.varname, plan);
+ if (conf.getBoolean(confPrefixForWorker +
HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname,
+ true)) {
+ Path planPath = new Path(plan);
+ planPath = new Path(planPath, REDUCE_PLAN_NAME);
+ String planString = conf.get(confPrefixForWorker +
planPath.toUri().getPath());
+ this.conf.set(planPath.toUri().getPath(), planString);
+ this.conf.set(HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname, "true");
+ } else {
+ this.conf.set(HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname, "false");
+ }
+ this.conf.set(HAS_REDUCE_WORK, "true");
+ this.conf.set(MAPRED_REDUCER_CLASS, ExecReducer.class.getName());
+
+ return Utilities.getReduceWork(conf);
+ }
+
+ private void initObjectInspectors(TableDesc valueTableDesc,TableDesc
keyTableDesc)
+ throws SerDeException {
+ inputKeyDeserializer =
+ ReflectionUtils.newInstance(keyTableDesc.getDeserializerClass(),
null);
+ SerDeUtils.initializeSerDe(inputKeyDeserializer, null,
+ keyTableDesc.getProperties(), null);
+ keyObjectInspector = inputKeyDeserializer.getObjectInspector();
+
+ keyStructInspector = (StructObjectInspector) keyObjectInspector;
+ firstValueColumnOffset = keyStructInspector.getAllStructFieldRefs().size();
+
+ Deserializer inputValueDeserializer = (AbstractSerDe)
ReflectionUtils.newInstance(
+ valueTableDesc.getDeserializerClass(), null);
+ SerDeUtils.initializeSerDe(inputValueDeserializer, null,
+ valueTableDesc.getProperties(), null);
+ valueObjectInspector = inputValueDeserializer.getObjectInspector();
+ valueStructInspectors = (StructObjectInspector) valueObjectInspector;
+ numValueCol = valueStructInspectors.getAllStructFieldRefs().size();
+ }
+
+ void initVectorBatches() throws HiveException {
+ inputBatch = batchContext.createVectorizedRowBatch();
+
+ // Create data buffers for value bytes column vectors.
+ for (int i = firstValueColumnOffset; i < inputBatch.numCols; i++) {
+ ColumnVector colVector = inputBatch.cols[i];
+ if (colVector instanceof BytesColumnVector) {
+ BytesColumnVector bytesColumnVector = (BytesColumnVector) colVector;
+ bytesColumnVector.initBuffer();
+ }
+ }
+
+ batchValueDeserializer =
+ new VectorDeserializeRow<>(
+ new LazyBinaryDeserializeRead(
+
VectorizedBatchUtil.typeInfosFromStructObjectInspector(
+ valueStructInspectors),
+ true));
+ batchValueDeserializer.init(firstValueColumnOffset);
+
+ int[] valueColumnMap = new int[numValueCol];
+ for (int i = 0; i < numValueCol; i++) {
+ valueColumnMap[i] = i + firstValueColumnOffset;
+ }
+
+ valueLazyBinarySerializeWrite = new LazyBinarySerializeWrite(numValueCol);
+ valueVectorSerializeRow = new
VectorSerializeRow<>(valueLazyBinarySerializeWrite);
+
valueVectorSerializeRow.init(VectorizedBatchUtil.typeInfosFromStructObjectInspector(
+ valueStructInspectors), valueColumnMap);
+ valueOutput = new ByteStream.Output();
+ valueVectorSerializeRow.setOutput(valueOutput);
+ outputBatch = batchContext.createVectorizedRowBatch();
+ valueBytesWritable = new DataInputBuffer();
+ }
+
+ private VectorAggregationBufferRow allocateAggregationBuffer() throws
HiveException {
+ VectorAggregateExpression.AggregationBuffer[] aggregationBuffers =
+ new
VectorAggregateExpression.AggregationBuffer[aggregators.length];
+ for (int i=0; i < aggregators.length; ++i) {
+ aggregationBuffers[i] = aggregators[i].getNewAggregationBuffer();
+ aggregators[i].reset(aggregationBuffers[i]);
+ }
+ return new VectorAggregationBufferRow(aggregationBuffers);
+ }
+
+ private void finishAggregation(DataInputBuffer key, IFile.Writer writer,
boolean needFlush)
+ throws HiveException, IOException {
+ for (int i = 0; i < aggregators.length; ++i) {
+ try {
+
aggregators[i].aggregateInput(aggregationBufferRow.getAggregationBuffer(i),
inputBatch);
+ } catch (HiveException e) {
+ throw new RuntimeException(e.getCause());
+ }
+ }
+
+ // In case the input batch is full but the keys are still same we need not
flush.
+ // Only evaluate the aggregates and store it in the aggregationBufferRow.
The aggregate
+ // functions are incremental and will take care of correctness when next
batch comes for
+ // aggregation.
+ if (!needFlush) {
+ return;
+ }
+
+ int colNum = firstValueColumnOffset;
+ for (int i = 0; i < aggregators.length; ++i) {
+ aggregators[i].assignRowColumn(outputBatch, 0, colNum++,
+ aggregationBufferRow.getAggregationBuffer(i));
+ }
+
+ valueLazyBinarySerializeWrite.reset();
+ valueVectorSerializeRow.serializeWrite(outputBatch, 0);
+ valueBytesWritable.reset(valueOutput.getData(), 0,
valueOutput.getLength());
+ writer.append(key, valueBytesWritable);
+ combineOutputRecordsCounter.increment(1);
+ aggregationBufferRow.reset();
+ outputBatch.reset();
+ }
+
+ private void addValueToBatch(DataInputBuffer val, DataInputBuffer key,
+ IFile.Writer writer, boolean needFLush) throws
IOException, HiveException {
+ batchValueDeserializer.setBytes(val.getData(), val.getPosition(),
+ val.getLength() - val.getPosition());
+ batchValueDeserializer.deserialize(inputBatch, inputBatch.size);
+ inputBatch.size++;
+ if (needFLush || (inputBatch.size >= VectorizedRowBatch.DEFAULT_SIZE)) {
+ processVectorGroup(key, writer, needFLush);
+ }
+ }
+
+ private void processVectorGroup(DataInputBuffer key, IFile.Writer writer,
boolean needFlush)
+ throws HiveException {
+ try {
+ finishAggregation(key, writer, needFlush);
+ inputBatch.reset();
+ } catch (Exception e) {
+ String rowString;
+ try {
+ rowString = inputBatch.toString();
+ } catch (Exception e2) {
+ rowString = "[Error getting row data with exception "
+ + StringUtils.stringifyException(e2) + " ]";
+ }
+ LOG.error("Hive Runtime Error while processing vector batch" +
rowString, e);
+ throw new HiveException("Hive Runtime Error while processing vector
batch", e);
+ }
+ }
+
+ protected void appendDirectlyToWriter(TezRawKeyValueIterator rawIter,
IFile.Writer writer) {
+ long numRows = 0;
+ try {
+ do {
+ numRows++;
+ writer.append(rawIter.getKey(), rawIter.getValue());
+ } while (rawIter.next());
+ combineInputRecordsCounter.increment(numRows);
+ combineOutputRecordsCounter.increment(numRows);
+ } catch(IOException e) {
+ LOG.error("Append to writer failed", e);
+ throw new RuntimeException(e.getMessage());
+ }
+ }
+
+ private void appendToWriter(DataInputBuffer val, DataInputBuffer key,
IFile.Writer writer) {
+ try {
+ writer.append(key, val);
+ combineOutputRecordsCounter.increment(1);
+ } catch(IOException e) {
+ LOG.error("Append value list to writer failed", e);
+ throw new RuntimeException(e.getMessage());
+ }
+ }
+
+ public static int compare(DataInputBuffer buf1, DataInputBuffer buf2) {
Review comment:
Is it compare or is it equals? (Equals has a fast-path exit for
different length)
##########
File path: ql/src/java/org/apache/hadoop/hive/ql/exec/GroupByCombiner.java
##########
@@ -0,0 +1,246 @@
+/*
+ * 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.exec;
+
+import org.apache.hadoop.hive.conf.HiveConf;
+import org.apache.hadoop.hive.ql.exec.vector.VectorGroupByCombiner;
+import org.apache.hadoop.hive.ql.exec.vector.VectorGroupByOperator;
+import org.apache.hadoop.mapred.JobConf;
+import org.apache.hadoop.hive.ql.metadata.HiveException;
+import org.apache.hadoop.hive.ql.plan.BaseWork;
+import org.apache.hadoop.hive.ql.plan.GroupByDesc;
+import org.apache.hadoop.hive.ql.plan.ReduceWork;
+import org.apache.hadoop.hive.ql.plan.TableDesc;
+import org.apache.hadoop.hive.ql.udf.generic.GenericUDAFEvaluator;
+import org.apache.hadoop.hive.serde2.AbstractSerDe;
+import org.apache.hadoop.hive.serde2.Deserializer;
+import org.apache.hadoop.hive.serde2.SerDeException;
+import org.apache.hadoop.hive.serde2.SerDeUtils;
+import org.apache.hadoop.hive.serde2.Serializer;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorFactory;
+import org.apache.hadoop.io.BytesWritable;
+import org.apache.hadoop.io.DataInputBuffer;
+import org.apache.hadoop.util.ReflectionUtils;
+import org.apache.tez.runtime.api.TaskContext;
+import org.apache.tez.runtime.library.common.sort.impl.IFile;
+import org.apache.tez.runtime.library.common.sort.impl.TezRawKeyValueIterator;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import org.apache.hadoop.fs.Path;
+
+import java.io.IOException;
+import java.util.ArrayList;
+
+import static org.apache.hadoop.hive.ql.exec.Utilities.HAS_REDUCE_WORK;
+import static org.apache.hadoop.hive.ql.exec.Utilities.REDUCE_PLAN_NAME;
+
+// Combiner for normal group by operator. In case of map side aggregate, the
partially
+// aggregated records are sorted based on group by key. If because of some
reasons, like hash
+// table memory exceeded the limit or the first few batches of records have
less ndvs, the
+// aggregation is not done, then here the aggregation can be done cheaply as
the records
+// are sorted based on group by key.
+public class GroupByCombiner extends VectorGroupByCombiner {
+
+ private static final Logger LOG = LoggerFactory.getLogger(
+ org.apache.hadoop.hive.ql.exec.GroupByCombiner.class.getName());
+
+ private transient GenericUDAFEvaluator[] aggregationEvaluators;
+ Deserializer valueDeserializer;
+ GenericUDAFEvaluator.AggregationBuffer[] aggregationBuffers;
+ GroupByOperator groupByOperator;
+ Serializer valueSerializer;
+ ObjectInspector aggrObjectInspector;
+ DataInputBuffer valueBuffer;
+ Object[] cachedValues;
+
+ public GroupByCombiner(TaskContext taskContext) throws HiveException,
IOException {
+ super(taskContext);
+ if (rw != null) {
+ try {
+ groupByOperator = (GroupByOperator) rw.getReducer();
+
+ ArrayList<ObjectInspector> ois = new ArrayList<ObjectInspector>();
+ ois.add(keyObjectInspector);
+ ois.add(valueObjectInspector);
+ ObjectInspector[] rowObjectInspector = new ObjectInspector[1];
+ rowObjectInspector[0] =
+
ObjectInspectorFactory.getStandardStructObjectInspector(Utilities.reduceFieldNameList,
+ ois);
+ groupByOperator.setInputObjInspectors(rowObjectInspector);
+ groupByOperator.initializeOp(conf);
+ aggregationBuffers = groupByOperator.getAggregationBuffers();
+ aggregationEvaluators = groupByOperator.getAggregationEvaluator();
+
+ TableDesc valueTableDesc = rw.getTagToValueDesc().get(0);
+ valueSerializer = (Serializer) valueTableDesc.getDeserializerClass()
+ .newInstance();
+ valueSerializer.initialize(null, valueTableDesc.getProperties());
+
+ valueDeserializer = (AbstractSerDe) ReflectionUtils.newInstance(
+ valueTableDesc.getDeserializerClass(), null);
+ SerDeUtils.initializeSerDe(valueDeserializer, null,
+ valueTableDesc.getProperties(), null);
+
+ aggrObjectInspector = groupByOperator.getAggrObjInspector();
+ valueBuffer = new DataInputBuffer();
+ cachedValues = new Object[aggregationEvaluators.length];
+ } catch (Exception e) {
+ LOG.error(" GroupByCombiner failed", e);
+ throw new RuntimeException(e.getMessage());
+ }
+ }
+ }
+
+ private void processAggregation(IFile.Writer writer, DataInputBuffer key)
+ throws Exception {
+ for (int i = 0; i < aggregationEvaluators.length; i++) {
+ cachedValues[i] =
aggregationEvaluators[i].evaluate(aggregationBuffers[i]);
+ }
+ BytesWritable result = (BytesWritable)
valueSerializer.serialize(cachedValues,
+ aggrObjectInspector);
+ valueBuffer.reset(result.getBytes(), result.getLength());
+ writer.append(key, valueBuffer);
+ combineOutputRecordsCounter.increment(1);
+ for (int i = 0; i < aggregationEvaluators.length; i++) {
+ aggregationEvaluators[i].reset(aggregationBuffers[i]);
+ }
+ }
+
+ private void updateAggregation(BytesWritable valWritable, DataInputBuffer
value)
+ throws HiveException, SerDeException {
+ valWritable.set(value.getData(), value.getPosition(),
+ value.getLength() - value.getPosition());
+ Object row = valueDeserializer.deserialize(valWritable);
+ groupByOperator.updateAggregation(row);
+ }
+
+ private void processRows(TezRawKeyValueIterator rawIter, IFile.Writer
writer) {
+ long numRows = 0;
+ try {
+ DataInputBuffer key = rawIter.getKey();
+ DataInputBuffer prevKey = new DataInputBuffer();
+ prevKey.reset(key.getData(), key.getPosition(), key.getLength() -
key.getPosition());
+ BytesWritable valWritable = new BytesWritable();
+ do {
+ key = rawIter.getKey();
+ // For first iteration, prevKey is always same as key.
+ if (VectorGroupByCombiner.compare(key, prevKey) != 0) {
Review comment:
Is this comparison necessary? Doesn't the combiner give you one iterator
per key?
##########
File path: ql/src/java/org/apache/hadoop/hive/ql/exec/GroupByOperator.java
##########
@@ -712,6 +751,12 @@ private void processKey(Object row,
@Override
public void process(Object row, int tag) throws HiveException {
+ if (hashAggr) {
+ if (getConfiguration().get("forced.streaming.mode",
"false").equals("true")) {
Review comment:
This should go into a GroupbyDesc
##########
File path:
ql/src/java/org/apache/hadoop/hive/ql/exec/vector/VectorGroupByCombiner.java
##########
@@ -0,0 +1,377 @@
+/*
+ * 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.exec.vector;
+
+import org.apache.hadoop.conf.Configuration;
+import org.apache.hadoop.fs.Path;
+import org.apache.hadoop.hive.conf.HiveConf;
+import org.apache.hadoop.hive.ql.exec.Utilities;
+import org.apache.hadoop.hive.ql.exec.mr.ExecReducer;
+import
org.apache.hadoop.hive.ql.exec.vector.expressions.aggregates.VectorAggregateExpression;
+import org.apache.hadoop.hive.ql.metadata.HiveException;
+import org.apache.hadoop.hive.ql.plan.ReduceWork;
+import org.apache.hadoop.hive.ql.plan.TableDesc;
+import org.apache.hadoop.hive.serde2.AbstractSerDe;
+import org.apache.hadoop.hive.serde2.ByteStream;
+import org.apache.hadoop.hive.serde2.Deserializer;
+import org.apache.hadoop.hive.serde2.SerDeException;
+import org.apache.hadoop.hive.serde2.SerDeUtils;
+import org.apache.hadoop.hive.serde2.lazybinary.fast.LazyBinaryDeserializeRead;
+import org.apache.hadoop.hive.serde2.lazybinary.fast.LazyBinarySerializeWrite;
+import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
+import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;
+import org.apache.hadoop.io.DataInputBuffer;
+import org.apache.hadoop.mapreduce.TaskCounter;
+import org.apache.hadoop.util.ReflectionUtils;
+import org.apache.hadoop.util.StringUtils;
+import org.apache.tez.common.TezUtils;
+import org.apache.tez.common.counters.TezCounter;
+import org.apache.tez.mapreduce.combine.MRCombiner;
+import org.apache.tez.runtime.api.TaskContext;
+import org.apache.tez.runtime.library.common.sort.impl.IFile;
+import org.apache.tez.runtime.library.common.sort.impl.TezRawKeyValueIterator;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+import java.io.IOException;
+
+import static org.apache.hadoop.hive.ql.exec.Utilities.HAS_REDUCE_WORK;
+import static org.apache.hadoop.hive.ql.exec.Utilities.MAPRED_REDUCER_CLASS;
+import static org.apache.hadoop.hive.ql.exec.Utilities.REDUCE_PLAN_NAME;
+import static
org.apache.hadoop.hive.serde2.lazy.fast.LazySimpleDeserializeRead.byteArrayCompareRanges;
+
+// Combiner for vectorized group by operator. In case of map side aggregate,
the partially
+// aggregated records are sorted based on group by key. If because of some
reasons, like hash
+// table memory exceeded the limit or the first few batches of records have
less ndvs, the
+// aggregation is not done, then here the aggregation can be done cheaply as
the records
+// are sorted based on group by key.
+public class VectorGroupByCombiner extends MRCombiner {
+ private static final Logger LOG = LoggerFactory.getLogger(
+ VectorGroupByCombiner.class.getName());
+ protected final Configuration conf;
+ protected final TezCounter combineInputRecordsCounter;
+ protected final TezCounter combineOutputRecordsCounter;
+ VectorAggregateExpression[] aggregators;
+ VectorAggregationBufferRow aggregationBufferRow;
+ protected transient LazyBinarySerializeWrite valueLazyBinarySerializeWrite;
+
+ // This helper object serializes LazyBinary format reducer values from
columns of a row
+ // in a vectorized row batch.
+ protected transient VectorSerializeRow<LazyBinarySerializeWrite>
valueVectorSerializeRow;
+
+ // The output buffer used to serialize a value into.
+ protected transient ByteStream.Output valueOutput;
+ DataInputBuffer valueBytesWritable;
+
+ // Only required minimal configs are copied to the worker nodes. This hack
(file.) is
+ // done to include these configs to be copied to the worker node.
+ protected static String confPrefixForWorker = "file.";
+
+ VectorDeserializeRow<LazyBinaryDeserializeRead> batchValueDeserializer;
+ int firstValueColumnOffset;
+ VectorizedRowBatchCtx batchContext = null;
+ int numValueCol = 0;
+ protected ReduceWork rw;
+ VectorizedRowBatch outputBatch = null;
+ VectorizedRowBatch inputBatch = null;
+ protected Deserializer inputKeyDeserializer = null;
+ protected ObjectInspector keyObjectInspector = null;
+ protected ObjectInspector valueObjectInspector = null;
+ protected StructObjectInspector valueStructInspectors = null;
+ protected StructObjectInspector keyStructInspector = null;
+
+ public VectorGroupByCombiner(TaskContext taskContext) throws HiveException,
IOException {
+ super(taskContext);
+
+ combineInputRecordsCounter =
+
taskContext.getCounters().findCounter(TaskCounter.COMBINE_INPUT_RECORDS);
+ combineOutputRecordsCounter =
+
taskContext.getCounters().findCounter(TaskCounter.COMBINE_OUTPUT_RECORDS);
+
+ conf = TezUtils.createConfFromUserPayload(taskContext.getUserPayload());
+ rw = getReduceWork();
+ if (rw == null) {
+ return;
+ }
+
+ if (rw.getReducer() instanceof VectorGroupByOperator) {
+ VectorGroupByOperator vectorGroupByOperator = (VectorGroupByOperator)
rw.getReducer();
+ vectorGroupByOperator.initializeOp(this.conf);
+ this.aggregators = vectorGroupByOperator.getAggregators();
+ this.aggregationBufferRow = allocateAggregationBuffer();
+ batchContext = rw.getVectorizedRowBatchCtx();
+ }
+
+ try {
+ initObjectInspectors(rw.getTagToValueDesc().get(0), rw.getKeyDesc());
+ if (batchContext != null && numValueCol > 0) {
+ initVectorBatches();
+ }
+ } catch (SerDeException e) {
+ LOG.error("Fail to initialize VectorGroupByCombiner.", e);
+ throw new RuntimeException(e.getCause());
+ }
+ }
+
+ // Get the reduce work from the config. Here some hack is used to prefix the
config name with
+ // "file." to avoid the config being filtered out.
+ private ReduceWork getReduceWork() {
+ String plan = conf.get(confPrefixForWorker +
HiveConf.ConfVars.PLAN.varname);
+ this.conf.set(HiveConf.ConfVars.PLAN.varname, plan);
+ if (conf.getBoolean(confPrefixForWorker +
HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname,
+ true)) {
+ Path planPath = new Path(plan);
+ planPath = new Path(planPath, REDUCE_PLAN_NAME);
+ String planString = conf.get(confPrefixForWorker +
planPath.toUri().getPath());
+ this.conf.set(planPath.toUri().getPath(), planString);
+ this.conf.set(HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname, "true");
+ } else {
+ this.conf.set(HiveConf.ConfVars.HIVE_RPC_QUERY_PLAN.varname, "false");
+ }
+ this.conf.set(HAS_REDUCE_WORK, "true");
+ this.conf.set(MAPRED_REDUCER_CLASS, ExecReducer.class.getName());
+
+ return Utilities.getReduceWork(conf);
+ }
+
+ private void initObjectInspectors(TableDesc valueTableDesc,TableDesc
keyTableDesc)
+ throws SerDeException {
+ inputKeyDeserializer =
+ ReflectionUtils.newInstance(keyTableDesc.getDeserializerClass(),
null);
+ SerDeUtils.initializeSerDe(inputKeyDeserializer, null,
+ keyTableDesc.getProperties(), null);
+ keyObjectInspector = inputKeyDeserializer.getObjectInspector();
+
+ keyStructInspector = (StructObjectInspector) keyObjectInspector;
+ firstValueColumnOffset = keyStructInspector.getAllStructFieldRefs().size();
+
+ Deserializer inputValueDeserializer = (AbstractSerDe)
ReflectionUtils.newInstance(
+ valueTableDesc.getDeserializerClass(), null);
+ SerDeUtils.initializeSerDe(inputValueDeserializer, null,
+ valueTableDesc.getProperties(), null);
+ valueObjectInspector = inputValueDeserializer.getObjectInspector();
+ valueStructInspectors = (StructObjectInspector) valueObjectInspector;
+ numValueCol = valueStructInspectors.getAllStructFieldRefs().size();
+ }
+
+ void initVectorBatches() throws HiveException {
+ inputBatch = batchContext.createVectorizedRowBatch();
+
+ // Create data buffers for value bytes column vectors.
+ for (int i = firstValueColumnOffset; i < inputBatch.numCols; i++) {
+ ColumnVector colVector = inputBatch.cols[i];
+ if (colVector instanceof BytesColumnVector) {
+ BytesColumnVector bytesColumnVector = (BytesColumnVector) colVector;
+ bytesColumnVector.initBuffer();
+ }
+ }
+
+ batchValueDeserializer =
+ new VectorDeserializeRow<>(
+ new LazyBinaryDeserializeRead(
+
VectorizedBatchUtil.typeInfosFromStructObjectInspector(
+ valueStructInspectors),
+ true));
+ batchValueDeserializer.init(firstValueColumnOffset);
+
+ int[] valueColumnMap = new int[numValueCol];
+ for (int i = 0; i < numValueCol; i++) {
+ valueColumnMap[i] = i + firstValueColumnOffset;
+ }
+
+ valueLazyBinarySerializeWrite = new LazyBinarySerializeWrite(numValueCol);
+ valueVectorSerializeRow = new
VectorSerializeRow<>(valueLazyBinarySerializeWrite);
+
valueVectorSerializeRow.init(VectorizedBatchUtil.typeInfosFromStructObjectInspector(
+ valueStructInspectors), valueColumnMap);
+ valueOutput = new ByteStream.Output();
+ valueVectorSerializeRow.setOutput(valueOutput);
+ outputBatch = batchContext.createVectorizedRowBatch();
+ valueBytesWritable = new DataInputBuffer();
+ }
+
+ private VectorAggregationBufferRow allocateAggregationBuffer() throws
HiveException {
+ VectorAggregateExpression.AggregationBuffer[] aggregationBuffers =
+ new
VectorAggregateExpression.AggregationBuffer[aggregators.length];
+ for (int i=0; i < aggregators.length; ++i) {
+ aggregationBuffers[i] = aggregators[i].getNewAggregationBuffer();
+ aggregators[i].reset(aggregationBuffers[i]);
+ }
+ return new VectorAggregationBufferRow(aggregationBuffers);
+ }
+
+ private void finishAggregation(DataInputBuffer key, IFile.Writer writer,
boolean needFlush)
+ throws HiveException, IOException {
+ for (int i = 0; i < aggregators.length; ++i) {
+ try {
+
aggregators[i].aggregateInput(aggregationBufferRow.getAggregationBuffer(i),
inputBatch);
+ } catch (HiveException e) {
+ throw new RuntimeException(e.getCause());
+ }
+ }
+
+ // In case the input batch is full but the keys are still same we need not
flush.
+ // Only evaluate the aggregates and store it in the aggregationBufferRow.
The aggregate
+ // functions are incremental and will take care of correctness when next
batch comes for
+ // aggregation.
+ if (!needFlush) {
+ return;
+ }
+
+ int colNum = firstValueColumnOffset;
+ for (int i = 0; i < aggregators.length; ++i) {
+ aggregators[i].assignRowColumn(outputBatch, 0, colNum++,
+ aggregationBufferRow.getAggregationBuffer(i));
+ }
+
+ valueLazyBinarySerializeWrite.reset();
+ valueVectorSerializeRow.serializeWrite(outputBatch, 0);
+ valueBytesWritable.reset(valueOutput.getData(), 0,
valueOutput.getLength());
+ writer.append(key, valueBytesWritable);
+ combineOutputRecordsCounter.increment(1);
+ aggregationBufferRow.reset();
+ outputBatch.reset();
+ }
+
+ private void addValueToBatch(DataInputBuffer val, DataInputBuffer key,
+ IFile.Writer writer, boolean needFLush) throws
IOException, HiveException {
+ batchValueDeserializer.setBytes(val.getData(), val.getPosition(),
+ val.getLength() - val.getPosition());
+ batchValueDeserializer.deserialize(inputBatch, inputBatch.size);
+ inputBatch.size++;
+ if (needFLush || (inputBatch.size >= VectorizedRowBatch.DEFAULT_SIZE)) {
+ processVectorGroup(key, writer, needFLush);
+ }
+ }
+
+ private void processVectorGroup(DataInputBuffer key, IFile.Writer writer,
boolean needFlush)
+ throws HiveException {
+ try {
+ finishAggregation(key, writer, needFlush);
+ inputBatch.reset();
+ } catch (Exception e) {
+ String rowString;
+ try {
+ rowString = inputBatch.toString();
+ } catch (Exception e2) {
+ rowString = "[Error getting row data with exception "
+ + StringUtils.stringifyException(e2) + " ]";
+ }
+ LOG.error("Hive Runtime Error while processing vector batch" +
rowString, e);
+ throw new HiveException("Hive Runtime Error while processing vector
batch", e);
+ }
+ }
+
+ protected void appendDirectlyToWriter(TezRawKeyValueIterator rawIter,
IFile.Writer writer) {
+ long numRows = 0;
+ try {
+ do {
+ numRows++;
+ writer.append(rawIter.getKey(), rawIter.getValue());
Review comment:
Need a Tez internal SAME_KEY here, right?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
Issue Time Tracking
-------------------
Worklog Id: (was: 522535)
Time Spent: 20m (was: 10m)
> Add support for combiner in hash mode group aggregation
> --------------------------------------------------------
>
> Key: HIVE-24471
> URL: https://issues.apache.org/jira/browse/HIVE-24471
> Project: Hive
> Issue Type: Bug
> Components: Hive
> Reporter: mahesh kumar behera
> Assignee: mahesh kumar behera
> Priority: Major
> Labels: pull-request-available
> Time Spent: 20m
> Remaining Estimate: 0h
>
> In map side group aggregation, partial grouped aggregation is calculated to
> reduce the data written to disk by map task. In case of hash aggregation,
> where the input data is not sorted, hash table is used. If the hash table
> size increases beyond configurable limit, data is flushed to disk and new
> hash table is generated. If the reduction by hash table is less than min hash
> aggregation reduction calculated during compile time, the map side
> aggregation is converted to streaming mode. So if the first few batch of
> records does not result into significant reduction, then the mode is switched
> to streaming mode. This may have impact on performance, if the subsequent
> batch of records have less number of distinct values. To mitigate this
> situation, a combiner can be added to the map task after the keys are sorted.
> This will make sure that the aggregation is done if possible and reduce the
> data written to disk.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
