/**
* 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 java.io.IOException;
import java.io.Serializable;
import java.util.ArrayList;
import java.util.Arrays;
import java.util.HashMap;
import java.util.HashSet;
import java.util.List;
import java.util.Map;
import java.util.Set;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.FileSystem;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hive.common.FileUtils;
import org.apache.hadoop.hive.common.StatsSetupConst;
import org.apache.hadoop.hive.conf.HiveConf;
import org.apache.hadoop.hive.ql.ErrorMsg;
import org.apache.hadoop.hive.ql.io.FSRecordWriter;
import org.apache.hadoop.hive.ql.io.FSRecordWriter.StatsProvidingRecordWriter;
import org.apache.hadoop.hive.ql.io.HiveFileFormatUtils;
import org.apache.hadoop.hive.ql.io.HiveKey;
import org.apache.hadoop.hive.ql.io.HiveOutputFormat;
import org.apache.hadoop.hive.ql.io.HivePartitioner;
import org.apache.hadoop.hive.ql.io.HivePassThroughOutputFormat;
import org.apache.hadoop.hive.ql.metadata.HiveException;
import org.apache.hadoop.hive.ql.metadata.HiveFatalException;
import org.apache.hadoop.hive.ql.plan.DynamicPartitionCtx;
import org.apache.hadoop.hive.ql.plan.ExprNodeDesc;
import org.apache.hadoop.hive.ql.plan.FileSinkDesc;
import org.apache.hadoop.hive.ql.plan.FileSinkDesc.DPSortState;
import org.apache.hadoop.hive.ql.plan.ListBucketingCtx;
import org.apache.hadoop.hive.ql.plan.PlanUtils;
import org.apache.hadoop.hive.ql.plan.SkewedColumnPositionPair;
import org.apache.hadoop.hive.ql.plan.api.OperatorType;
import org.apache.hadoop.hive.ql.stats.StatsCollectionTaskIndependent;
import org.apache.hadoop.hive.ql.stats.StatsPublisher;
import org.apache.hadoop.hive.serde2.SerDeException;
import org.apache.hadoop.hive.serde2.SerDeStats;
import org.apache.hadoop.hive.serde2.Serializer;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorUtils;
import org.apache.hadoop.hive.serde2.objectinspector.ObjectInspectorUtils.ObjectInspectorCopyOption;
import org.apache.hadoop.hive.serde2.objectinspector.StructField;
import org.apache.hadoop.hive.serde2.objectinspector.StructObjectInspector;
import org.apache.hadoop.hive.serde2.objectinspector.SubStructObjectInspector;
import org.apache.hadoop.io.LongWritable;
import org.apache.hadoop.io.Writable;
import org.apache.hadoop.mapred.JobConf;
import org.apache.hadoop.util.ReflectionUtils;
import com.google.common.collect.Lists;
/**
* File Sink operator implementation.
**/
public class FileSinkOperator extends TerminalOperator<FileSinkDesc> implements
Serializable {
protected transient HashMap<String, FSPaths> valToPaths;
protected transient int numDynParts;
protected transient List<String> dpColNames;
protected transient DynamicPartitionCtx dpCtx;
protected transient boolean isCompressed;
protected transient Path parent;
protected transient HiveOutputFormat<?, ?> hiveOutputFormat;
protected transient Path specPath;
protected transient String childSpecPathDynLinkedPartitions;
protected transient int dpStartCol; // start column # for DP columns
protected transient List<String> dpVals; // array of values corresponding to DP columns
protected transient List<Object> dpWritables;
protected transient FSRecordWriter[] rowOutWriters; // row specific RecordWriters
protected transient int maxPartitions;
protected transient ListBucketingCtx lbCtx;
protected transient boolean isSkewedStoredAsSubDirectories;
protected transient boolean statsCollectRawDataSize;
private transient boolean[] statsFromRecordWriter;
private transient boolean isCollectRWStats;
private transient FSPaths prevFsp;
private transient FSPaths fpaths;
private transient ObjectInspector keyOI;
private transient List<Object> keyWritables;
private transient List<String> keys;
private transient int numKeyColToRead;
/**
* RecordWriter.
*
*/
public static interface RecordWriter {
void write(Writable w) throws IOException;
void close(boolean abort) throws IOException;
}
public class FSPaths implements Cloneable {
Path tmpPath;
Path taskOutputTempPath;
Path[] outPaths;
Path[] finalPaths;
FSRecordWriter[] outWriters;
Stat stat;
public FSPaths() {
}
public FSPaths(Path specPath) {
tmpPath = Utilities.toTempPath(specPath);
taskOutputTempPath = Utilities.toTaskTempPath(specPath);
outPaths = new Path[numFiles];
finalPaths = new Path[numFiles];
outWriters = new FSRecordWriter[numFiles];
stat = new Stat();
}
/**
* Update OutPath according to tmpPath.
*/
public Path getTaskOutPath(String taskId) {
return new Path(this.taskOutputTempPath, Utilities.toTempPath(taskId));
}
/**
* Update the final paths according to tmpPath.
*/
public Path getFinalPath(String taskId, Path tmpPath, String extension) {
if (extension != null) {
return new Path(tmpPath, taskId + extension);
} else {
return new Path(tmpPath, taskId);
}
}
public void setOutWriters(FSRecordWriter[] out) {
outWriters = out;
}
public FSRecordWriter[] getOutWriters() {
return outWriters;
}
public void closeWriters(boolean abort) throws HiveException {
for (int idx = 0; idx < outWriters.length; idx++) {
if (outWriters[idx] != null) {
try {
outWriters[idx].close(abort);
updateProgress();
} catch (IOException e) {
throw new HiveException(e);
}
}
}
}
private void commit(FileSystem fs) throws HiveException {
for (int idx = 0; idx < outPaths.length; ++idx) {
try {
if ((bDynParts || isSkewedStoredAsSubDirectories)
&& !fs.exists(finalPaths[idx].getParent())) {
fs.mkdirs(finalPaths[idx].getParent());
}
if (!fs.rename(outPaths[idx], finalPaths[idx])) {
throw new HiveException("Unable to rename output from: " +
outPaths[idx] + " to: " + finalPaths[idx]);
}
updateProgress();
} catch (IOException e) {
throw new HiveException("Unable to rename output from: " +
outPaths[idx] + " to: " + finalPaths[idx], e);
}
}
}
public void abortWriters(FileSystem fs, boolean abort, boolean delete) throws HiveException {
for (int idx = 0; idx < outWriters.length; idx++) {
if (outWriters[idx] != null) {
try {
outWriters[idx].close(abort);
if (delete) {
fs.delete(outPaths[idx], true);
}
updateProgress();
} catch (IOException e) {
throw new HiveException(e);
}
}
}
}
public Stat getStat() {
return stat;
}
} // class FSPaths
private static final long serialVersionUID = 1L;
protected transient FileSystem fs;
protected transient Serializer serializer;
protected transient LongWritable row_count;
private transient boolean isNativeTable = true;
/**
* The evaluators for the multiFile sprayer. If the table under consideration has 1000 buckets,
* it is not a good idea to start so many reducers - if the maximum number of reducers is 100,
* each reducer can write 10 files - this way we effectively get 1000 files.
*/
private transient ExprNodeEvaluator[] partitionEval;
protected transient int totalFiles;
private transient int numFiles;
protected transient boolean multiFileSpray;
protected transient final Map<Integer, Integer> bucketMap = new HashMap<Integer, Integer>();
private transient ObjectInspector[] partitionObjectInspectors;
protected transient HivePartitioner<HiveKey, Object> prtner;
protected transient final HiveKey key = new HiveKey();
private transient Configuration hconf;
protected transient FSPaths fsp;
protected transient boolean bDynParts;
private transient SubStructObjectInspector subSetOI;
private transient int timeOut; // JT timeout in msec.
private transient long lastProgressReport = System.currentTimeMillis();
protected transient boolean autoDelete = false;
protected transient JobConf jc;
Class<? extends Writable> outputClass;
String taskId;
protected boolean filesCreated = false;
private void initializeSpecPath() {
// For a query of the type:
// insert overwrite table T1
// select * from (subq1 union all subq2)u;
// subQ1 and subQ2 write to directories Parent/Child_1 and
// Parent/Child_2 respectively, and union is removed.
// The movetask that follows subQ1 and subQ2 tasks moves the directory
// 'Parent'
// However, if the above query contains dynamic partitions, subQ1 and
// subQ2 have to write to directories: Parent/DynamicPartition/Child_1
// and Parent/DynamicPartition/Child_1 respectively.
// The movetask that follows subQ1 and subQ2 tasks still moves the directory
// 'Parent'
if ((!conf.isLinkedFileSink()) || (dpCtx == null)) {
specPath = conf.getDirName();
childSpecPathDynLinkedPartitions = null;
return;
}
specPath = conf.getParentDir();
childSpecPathDynLinkedPartitions = conf.getDirName().getName();
}
@Override
protected void initializeOp(Configuration hconf) throws HiveException {
try {
this.hconf = hconf;
filesCreated = false;
isNativeTable = !conf.getTableInfo().isNonNative();
multiFileSpray = conf.isMultiFileSpray();
totalFiles = conf.getTotalFiles();
numFiles = conf.getNumFiles();
dpCtx = conf.getDynPartCtx();
lbCtx = conf.getLbCtx();
fsp = prevFsp = null;
valToPaths = new HashMap<String, FSPaths>();
taskId = Utilities.getTaskId(hconf);
initializeSpecPath();
fs = specPath.getFileSystem(hconf);
hiveOutputFormat = conf.getTableInfo().getOutputFileFormatClass().newInstance();
isCompressed = conf.getCompressed();
parent = Utilities.toTempPath(conf.getDirName());
statsCollectRawDataSize = conf.isStatsCollectRawDataSize();
statsFromRecordWriter = new boolean[numFiles];
serializer = (Serializer) conf.getTableInfo().getDeserializerClass().newInstance();
serializer.initialize(null, conf.getTableInfo().getProperties());
outputClass = serializer.getSerializedClass();
// Timeout is chosen to make sure that even if one iteration takes more than
// half of the script.timeout but less than script.timeout, we will still
// be able to report progress.
timeOut = hconf.getInt("mapred.healthChecker.script.timeout", 600000) / 2;
if (hconf instanceof JobConf) {
jc = (JobConf) hconf;
} else {
// test code path
jc = new JobConf(hconf);
}
if (multiFileSpray) {
partitionEval = new ExprNodeEvaluator[conf.getPartitionCols().size()];
int i = 0;
for (ExprNodeDesc e : conf.getPartitionCols()) {
partitionEval[i++] = ExprNodeEvaluatorFactory.get(e);
}
partitionObjectInspectors = initEvaluators(partitionEval, outputObjInspector);
prtner = (HivePartitioner<HiveKey, Object>) ReflectionUtils.newInstance(
jc.getPartitionerClass(), null);
}
row_count = new LongWritable();
if (dpCtx != null) {
dpSetup();
}
if (lbCtx != null) {
lbSetup();
}
int numPart = 0;
int numBuck = 0;
if (conf.getPartitionCols() != null && !conf.getPartitionCols().isEmpty()) {
numPart = conf.getPartitionCols().size();
}
// bucket number will exists only in PARTITION_BUCKET_SORTED mode
if (conf.getDpSortState().equals(DPSortState.PARTITION_BUCKET_SORTED)) {
numBuck = 1;
}
numKeyColToRead = numPart + numBuck;
keys = Lists.newArrayListWithCapacity(numKeyColToRead);
keyWritables = Lists.newArrayListWithCapacity(numKeyColToRead);
if (!bDynParts) {
fsp = new FSPaths(specPath);
// Create all the files - this is required because empty files need to be created for
// empty buckets
// createBucketFiles(fsp);
if (!this.isSkewedStoredAsSubDirectories) {
valToPaths.put("", fsp); // special entry for non-DP case
}
}
initializeChildren(hconf);
} catch (HiveException e) {
throw e;
} catch (Exception e) {
e.printStackTrace();
throw new HiveException(e);
}
}
/**
* Initialize list bucketing information
*/
private void lbSetup() {
this.isSkewedStoredAsSubDirectories = ((lbCtx == null) ? false : lbCtx.isSkewedStoredAsDir());
}
/**
* Set up for dynamic partitioning including a new ObjectInspector for the output row.
*/
private void dpSetup() {
this.bDynParts = false;
this.numDynParts = dpCtx.getNumDPCols();
this.dpColNames = dpCtx.getDPColNames();
this.maxPartitions = dpCtx.getMaxPartitionsPerNode();
assert numDynParts == dpColNames.size() : "number of dynamic paritions should be the same as the size of DP mapping";
if (dpColNames != null && dpColNames.size() > 0) {
this.bDynParts = true;
assert inputObjInspectors.length == 1 : "FileSinkOperator should have 1 parent, but it has "
+ inputObjInspectors.length;
StructObjectInspector soi = (StructObjectInspector) inputObjInspectors[0];
// remove the last dpMapping.size() columns from the OI
List<? extends StructField> fieldOI = soi.getAllStructFieldRefs();
ArrayList<ObjectInspector> newFieldsOI = new ArrayList<ObjectInspector>();
ArrayList<String> newFieldsName = new ArrayList<String>();
this.dpStartCol = 0;
for (StructField sf : fieldOI) {
String fn = sf.getFieldName();
if (!dpCtx.getInputToDPCols().containsKey(fn)) {
newFieldsOI.add(sf.getFieldObjectInspector());
newFieldsName.add(sf.getFieldName());
this.dpStartCol++;
}
}
assert newFieldsOI.size() > 0 : "new Fields ObjectInspector is empty";
this.subSetOI = new SubStructObjectInspector(soi, 0, this.dpStartCol);
this.dpVals = new ArrayList<String>(numDynParts);
this.dpWritables = new ArrayList<Object>(numDynParts);
}
}
protected void createBucketFiles(FSPaths fsp) throws HiveException {
try {
int filesIdx = 0;
Set<Integer> seenBuckets = new HashSet<Integer>();
for (int idx = 0; idx < totalFiles; idx++) {
if (this.getExecContext() != null && this.getExecContext().getFileId() != null) {
LOG.info("replace taskId from execContext ");
taskId = Utilities.replaceTaskIdFromFilename(taskId, this.getExecContext().getFileId());
LOG.info("new taskId: FS " + taskId);
assert !multiFileSpray;
assert totalFiles == 1;
}
if (multiFileSpray) {
key.setHashCode(idx);
// Does this hashcode belong to this reducer
int numReducers = totalFiles / numFiles;
if (numReducers > 1) {
int currReducer = Integer.valueOf(Utilities.getTaskIdFromFilename(Utilities
.getTaskId(hconf)));
int reducerIdx = prtner.getPartition(key, null, numReducers);
if (currReducer != reducerIdx) {
continue;
}
}
int bucketNum = prtner.getBucket(key, null, totalFiles);
if (seenBuckets.contains(bucketNum)) {
continue;
}
seenBuckets.add(bucketNum);
bucketMap.put(bucketNum, filesIdx);
taskId = Utilities.replaceTaskIdFromFilename(Utilities.getTaskId(hconf), bucketNum);
}
createBucketForFileIdx(fsp, filesIdx);
filesIdx++;
}
assert filesIdx == numFiles;
// in recent hadoop versions, use deleteOnExit to clean tmp files.
if (isNativeTable) {
autoDelete = fs.deleteOnExit(fsp.outPaths[0]);
}
} catch (Exception e) {
e.printStackTrace();
throw new HiveException(e);
}
filesCreated = true;
}
protected void createBucketForFileIdx(FSPaths fsp, int filesIdx) throws HiveException {
try {
if (isNativeTable) {
fsp.finalPaths[filesIdx] = fsp.getFinalPath(taskId, fsp.tmpPath, null);
LOG.info("Final Path: FS " + fsp.finalPaths[filesIdx]);
fsp.outPaths[filesIdx] = fsp.getTaskOutPath(taskId);
LOG.info("Writing to temp file: FS " + fsp.outPaths[filesIdx]);
} else {
fsp.finalPaths[filesIdx] = fsp.outPaths[filesIdx] = specPath;
}
// The reason to keep these instead of using
// OutputFormat.getRecordWriter() is that
// getRecordWriter does not give us enough control over the file name that
// we create.
String extension = Utilities.getFileExtension(jc, isCompressed, hiveOutputFormat);
if (!bDynParts && !this.isSkewedStoredAsSubDirectories) {
fsp.finalPaths[filesIdx] = fsp.getFinalPath(taskId, parent, extension);
} else {
fsp.finalPaths[filesIdx] = fsp.getFinalPath(taskId, fsp.tmpPath, extension);
}
LOG.info("New Final Path: FS " + fsp.finalPaths[filesIdx]);
if (isNativeTable) {
// in recent hadoop versions, use deleteOnExit to clean tmp files.
autoDelete = fs.deleteOnExit(fsp.outPaths[filesIdx]);
}
Utilities.copyTableJobPropertiesToConf(conf.getTableInfo(), jc);
// only create bucket files only if no dynamic partitions,
// buckets of dynamic partitions will be created for each newly created partition
fsp.outWriters[filesIdx] = HiveFileFormatUtils.getHiveRecordWriter(jc, conf.getTableInfo(),
outputClass, conf, fsp.outPaths[filesIdx], reporter);
// If the record writer provides stats, get it from there instead of the serde
statsFromRecordWriter[filesIdx] = fsp.outWriters[filesIdx] instanceof StatsProvidingRecordWriter;
// increment the CREATED_FILES counter
if (reporter != null) {
reporter.incrCounter(HiveConf.getVar(hconf, HiveConf.ConfVars.HIVECOUNTERGROUP),
Operator.HIVECOUNTERCREATEDFILES, 1);
}
} catch (IOException e) {
throw new HiveException(e);
}
}
/**
* Report status to JT so that JT won't kill this task if closing takes too long
* due to too many files to close and the NN is overloaded.
*
* @return true if a new progress update is reported, false otherwise.
*/
protected boolean updateProgress() {
if (reporter != null &&
(System.currentTimeMillis() - lastProgressReport) > timeOut) {
reporter.progress();
lastProgressReport = System.currentTimeMillis();
return true;
} else {
return false;
}
}
protected Writable recordValue;
@Override
public void processOp(Object row, int tag) throws HiveException {
/* Create list bucketing sub-directory only if stored-as-directories is on. */
String lbDirName = null;
lbDirName = (lbCtx == null) ? null : generateListBucketingDirName(row);
if (!bDynParts && !filesCreated) {
if (lbDirName != null) {
FSPaths fsp2 = lookupListBucketingPaths(lbDirName);
} else {
createBucketFiles(fsp);
}
}
try {
updateProgress();
// if DP is enabled, get the final output writers and prepare the real output row
assert inputObjInspectors[0].getCategory() == ObjectInspector.Category.STRUCT : "input object inspector is not struct";
if (bDynParts) {
// copy the DP column values from the input row to dpVals
dpVals.clear();
dpWritables.clear();
ObjectInspectorUtils.partialCopyToStandardObject(dpWritables, row, dpStartCol, numDynParts,
(StructObjectInspector) inputObjInspectors[0], ObjectInspectorCopyOption.WRITABLE);
// get a set of RecordWriter based on the DP column values
// pass the null value along to the escaping process to determine what the dir should be
for (Object o : dpWritables) {
if (o == null || o.toString().length() == 0) {
dpVals.add(dpCtx.getDefaultPartitionName());
} else {
dpVals.add(o.toString());
}
}
// use SubStructObjectInspector to serialize the non-partitioning columns in the input row
recordValue = serializer.serialize(row, subSetOI);
// when dynamic partition sorting is not used, the DPSortState will be NONE
// in which we will fall back to old method of file system path creation
// i.e, having as many record writers as distinct values in partition column
if (conf.getDpSortState().equals(DPSortState.NONE)) {
fpaths = getDynOutPaths(dpVals, lbDirName);
}
} else {
if (lbDirName != null) {
fpaths = lookupListBucketingPaths(lbDirName);
} else {
fpaths = fsp;
}
// use SerDe to serialize r, and write it out
recordValue = serializer.serialize(row, inputObjInspectors[0]);
}
rowOutWriters = fpaths.outWriters;
// check if all record writers implement statistics. if atleast one RW
// doesn't implement stats interface we will fallback to conventional way
// of gathering stats
isCollectRWStats = areAllTrue(statsFromRecordWriter);
if (conf.isGatherStats() && !isCollectRWStats) {
if (statsCollectRawDataSize) {
SerDeStats stats = serializer.getSerDeStats();
if (stats != null) {
fpaths.stat.addToStat(StatsSetupConst.RAW_DATA_SIZE, stats.getRawDataSize());
}
}
fpaths.stat.addToStat(StatsSetupConst.ROW_COUNT, 1);
}
FSRecordWriter rowOutWriter = null;
if (row_count != null) {
row_count.set(row_count.get() + 1);
}
if (!multiFileSpray) {
rowOutWriter = rowOutWriters[0];
} else {
int keyHashCode = 0;
for (int i = 0; i < partitionEval.length; i++) {
Object o = partitionEval[i].evaluate(row);
keyHashCode = keyHashCode * 31
+ ObjectInspectorUtils.hashCode(o, partitionObjectInspectors[i]);
}
key.setHashCode(keyHashCode);
int bucketNum = prtner.getBucket(key, null, totalFiles);
int idx = bucketMap.get(bucketNum);
rowOutWriter = rowOutWriters[idx];
}
rowOutWriter.write(recordValue);
} catch (IOException e) {
throw new HiveException(e);
} catch (SerDeException e) {
throw new HiveException(e);
}
}
private boolean areAllTrue(boolean[] statsFromRW) {
for(boolean b : statsFromRW) {
if (!b) {
return false;
}
}
return true;
}
/**
* Lookup list bucketing path.
* @param lbDirName
* @return
* @throws HiveException
*/
protected FSPaths lookupListBucketingPaths(String lbDirName) throws HiveException {
FSPaths fsp2 = valToPaths.get(lbDirName);
if (fsp2 == null) {
fsp2 = createNewPaths(lbDirName);
}
return fsp2;
}
/**
* create new path.
*
* @param dirName
* @return
* @throws HiveException
*/
private FSPaths createNewPaths(String dirName) throws HiveException {
FSPaths fsp2 = new FSPaths(specPath);
if (childSpecPathDynLinkedPartitions != null) {
fsp2.tmpPath = new Path(fsp2.tmpPath,
dirName + Path.SEPARATOR + childSpecPathDynLinkedPartitions);
fsp2.taskOutputTempPath =
new Path(fsp2.taskOutputTempPath,
dirName + Path.SEPARATOR + childSpecPathDynLinkedPartitions);
} else {
fsp2.tmpPath = new Path(fsp2.tmpPath, dirName);
fsp2.taskOutputTempPath =
new Path(fsp2.taskOutputTempPath, dirName);
}
if(!conf.getDpSortState().equals(DPSortState.PARTITION_BUCKET_SORTED)) {
createBucketFiles(fsp2);
valToPaths.put(dirName, fsp2);
}
return fsp2;
}
/**
* Generate list bucketing directory name from a row.
* @param row row to process.
* @return directory name.
*/
protected String generateListBucketingDirName(Object row) {
if (!this.isSkewedStoredAsSubDirectories) {
return null;
}
String lbDirName = null;
List<Object> standObjs = new ArrayList<Object>();
List<String> skewedCols = lbCtx.getSkewedColNames();
List<List<String>> allSkewedVals = lbCtx.getSkewedColValues();
List<String> skewedValsCandidate = null;
Map<List<String>, String> locationMap = lbCtx.getLbLocationMap();
/* Convert input row to standard objects. */
ObjectInspectorUtils.copyToStandardObject(standObjs, row,
(StructObjectInspector) inputObjInspectors[0], ObjectInspectorCopyOption.WRITABLE);
assert (standObjs.size() >= skewedCols.size()) :
"The row has less number of columns than no. of skewed column.";
skewedValsCandidate = new ArrayList<String>(skewedCols.size());
for (SkewedColumnPositionPair posPair : lbCtx.getRowSkewedIndex()) {
skewedValsCandidate.add(posPair.getSkewColPosition(),
standObjs.get(posPair.getTblColPosition()).toString());
}
/* The row matches skewed column names. */
if (allSkewedVals.contains(skewedValsCandidate)) {
/* matches skewed values. */
lbDirName = FileUtils.makeListBucketingDirName(skewedCols, skewedValsCandidate);
locationMap.put(skewedValsCandidate, lbDirName);
} else {
/* create default directory. */
lbDirName = FileUtils.makeDefaultListBucketingDirName(skewedCols,
lbCtx.getDefaultDirName());
List<String> defaultKey = Arrays.asList(lbCtx.getDefaultKey());
if (!locationMap.containsKey(defaultKey)) {
locationMap.put(defaultKey, lbDirName);
}
}
return lbDirName;
}
protected FSPaths getDynOutPaths(List<String> row, String lbDirName) throws HiveException {
FSPaths fp;
// get the path corresponding to the dynamic partition columns,
String dpDir = getDynPartDirectory(row, dpColNames, numDynParts);
String pathKey = null;
if (dpDir != null) {
dpDir = appendToSource(lbDirName, dpDir);
pathKey = dpDir;
if(conf.getDpSortState().equals(DPSortState.PARTITION_BUCKET_SORTED)) {
String buckNum = row.get(row.size() - 1);
taskId = Utilities.replaceTaskIdFromFilename(Utilities.getTaskId(hconf), buckNum);
pathKey = appendToSource(taskId, dpDir);
}
FSPaths fsp2 = valToPaths.get(pathKey);
if (fsp2 == null) {
// check # of dp
if (valToPaths.size() > maxPartitions) {
// we cannot proceed and need to tell the hive client that retries won't succeed either
throw new HiveFatalException(
ErrorMsg.DYNAMIC_PARTITIONS_TOO_MANY_PER_NODE_ERROR.getErrorCodedMsg()
+ "Maximum was set to: " + maxPartitions);
}
if (!conf.getDpSortState().equals(DPSortState.NONE) && prevFsp != null) {
// close the previous fsp as it is no longer needed
prevFsp.closeWriters(false);
// since we are closing the previous fsp's record writers, we need to see if we can get
// stats from the record writer and store in the previous fsp that is cached
if (conf.isGatherStats() && isCollectRWStats) {
FSRecordWriter outWriter = prevFsp.outWriters[0];
if (outWriter != null) {
SerDeStats stats = ((StatsProvidingRecordWriter) outWriter).getStats();
if (stats != null) {
prevFsp.stat.addToStat(StatsSetupConst.RAW_DATA_SIZE, stats.getRawDataSize());
prevFsp.stat.addToStat(StatsSetupConst.ROW_COUNT, stats.getRowCount());
}
}
}
// let writers release the memory for garbage collection
prevFsp.outWriters[0] = null;
prevFsp = null;
}
fsp2 = createNewPaths(dpDir);
if (prevFsp == null) {
prevFsp = fsp2;
}
if(conf.getDpSortState().equals(DPSortState.PARTITION_BUCKET_SORTED)) {
createBucketForFileIdx(fsp2, 0);
valToPaths.put(pathKey, fsp2);
}
}
fp = fsp2;
} else {
fp = fsp;
}
return fp;
}
/**
* Append dir to source dir
* @param appendDir
* @param srcDir
* @return
*/
private String appendToSource(String appendDir, String srcDir) {
StringBuilder builder = new StringBuilder(srcDir);
srcDir = (appendDir == null) ? srcDir : builder.append(Path.SEPARATOR).append(appendDir)
.toString();
return srcDir;
}
// given the current input row, the mapping for input col info to dp columns, and # of dp cols,
// return the relative path corresponding to the row.
// e.g., ds=2008-04-08/hr=11
private String getDynPartDirectory(List<String> row, List<String> dpColNames, int numDynParts) {
assert row.size() == numDynParts && numDynParts == dpColNames.size() : "data length is different from num of DP columns";
return FileUtils.makePartName(dpColNames, row);
}
@Override
public void startGroup() throws HiveException {
if (!conf.getDpSortState().equals(DPSortState.NONE)) {
keyOI = getGroupKeyObjectInspector();
keys.clear();
keyWritables.clear();
ObjectInspectorUtils.partialCopyToStandardObject(keyWritables, getGroupKeyObject(), 0,
numKeyColToRead, (StructObjectInspector) keyOI, ObjectInspectorCopyOption.WRITABLE);
for (Object o : keyWritables) {
if (o == null || o.toString().length() == 0) {
keys.add(dpCtx.getDefaultPartitionName());
} else {
keys.add(o.toString());
}
}
fpaths = getDynOutPaths(keys, null);
}
}
@Override
public void closeOp(boolean abort) throws HiveException {
if (!bDynParts && !filesCreated) {
createBucketFiles(fsp);
}
lastProgressReport = System.currentTimeMillis();
if (!abort) {
for (FSPaths fsp : valToPaths.values()) {
fsp.closeWriters(abort);
// before closing the operator check if statistics gathering is requested
// and is provided by record writer. this is different from the statistics
// gathering done in processOp(). In processOp(), for each row added
// serde statistics about the row is gathered and accumulated in hashmap.
// this adds more overhead to the actual processing of row. But if the
// record writer already gathers the statistics, it can simply return the
// accumulated statistics which will be aggregated in case of spray writers
if (conf.isGatherStats() && isCollectRWStats) {
for (int idx = 0; idx < fsp.outWriters.length; idx++) {
FSRecordWriter outWriter = fsp.outWriters[idx];
if (outWriter != null) {
SerDeStats stats = ((StatsProvidingRecordWriter) outWriter).getStats();
if (stats != null) {
fsp.stat.addToStat(StatsSetupConst.RAW_DATA_SIZE, stats.getRawDataSize());
fsp.stat.addToStat(StatsSetupConst.ROW_COUNT, stats.getRowCount());
}
}
}
}
if (isNativeTable) {
fsp.commit(fs);
}
}
// Only publish stats if this operator's flag was set to gather stats
if (conf.isGatherStats()) {
publishStats();
}
} else {
// Will come here if an Exception was thrown in map() or reduce().
// Hadoop always call close() even if an Exception was thrown in map() or
// reduce().
for (FSPaths fsp : valToPaths.values()) {
fsp.abortWriters(fs, abort, !autoDelete && isNativeTable);
}
}
fsp = prevFsp = null;
}
/**
* @return the name of the operator
*/
@Override
public String getName() {
return getOperatorName();
}
static public String getOperatorName() {
return "FS";
}
@Override
public void jobCloseOp(Configuration hconf, boolean success)
throws HiveException {
try {
if ((conf != null) && isNativeTable) {
Path specPath = conf.getDirName();
DynamicPartitionCtx dpCtx = conf.getDynPartCtx();
if (conf.isLinkedFileSink() && (dpCtx != null)) {
specPath = conf.getParentDir();
}
Utilities.mvFileToFinalPath(specPath, hconf, success, LOG, dpCtx, conf,
reporter);
}
} catch (IOException e) {
throw new HiveException(e);
}
super.jobCloseOp(hconf, success);
}
@Override
public OperatorType getType() {
return OperatorType.FILESINK;
}
@Override
public void augmentPlan() {
PlanUtils.configureOutputJobPropertiesForStorageHandler(
getConf().getTableInfo());
}
public void checkOutputSpecs(FileSystem ignored, JobConf job) throws IOException {
if (hiveOutputFormat == null) {
try {
if (getConf().getTableInfo().getJobProperties() != null) {
//Setting only for Storage Handler
if (getConf().getTableInfo().getJobProperties().get(HivePassThroughOutputFormat.HIVE_PASSTHROUGH_STORAGEHANDLER_OF_JOBCONFKEY) != null) {
job.set(HivePassThroughOutputFormat.HIVE_PASSTHROUGH_STORAGEHANDLER_OF_JOBCONFKEY,getConf().getTableInfo().getJobProperties().get(HivePassThroughOutputFormat.HIVE_PASSTHROUGH_STORAGEHANDLER_OF_JOBCONFKEY));
hiveOutputFormat = ReflectionUtils.newInstance(conf.getTableInfo().getOutputFileFormatClass(),job);
}
else {
hiveOutputFormat = conf.getTableInfo().getOutputFileFormatClass().newInstance();
}
}
else {
hiveOutputFormat = conf.getTableInfo().getOutputFileFormatClass().newInstance();
}
} catch (Exception ex) {
throw new IOException(ex);
}
}
Utilities.copyTableJobPropertiesToConf(conf.getTableInfo(), job);
if (conf.getTableInfo().isNonNative()) {
//check the ouput specs only if it is a storage handler (native tables's outputformats does
//not set the job's output properties correctly)
try {
hiveOutputFormat.checkOutputSpecs(ignored, job);
} catch (NoSuchMethodError e) {
//For BC, ignore this for now, but leave a log message
LOG.warn("HiveOutputFormat should implement checkOutputSpecs() method`");
}
}
}
private void publishStats() throws HiveException {
boolean isStatsReliable = conf.isStatsReliable();
// Initializing a stats publisher
StatsPublisher statsPublisher = Utilities.getStatsPublisher(jc);
if (statsPublisher == null) {
// just return, stats gathering should not block the main query
LOG.error("StatsPublishing error: StatsPublisher is not initialized.");
if (isStatsReliable) {
throw new HiveException(ErrorMsg.STATSPUBLISHER_NOT_OBTAINED.getErrorCodedMsg());
}
return;
}
if (!statsPublisher.connect(hconf)) {
// just return, stats gathering should not block the main query
LOG.error("StatsPublishing error: cannot connect to database");
if (isStatsReliable) {
throw new HiveException(ErrorMsg.STATSPUBLISHER_CONNECTION_ERROR.getErrorCodedMsg());
}
return;
}
String taskID = Utilities.getTaskIdFromFilename(Utilities.getTaskId(hconf));
String spSpec = conf.getStaticSpec();
int maxKeyLength = conf.getMaxStatsKeyPrefixLength();
boolean taskIndependent = statsPublisher instanceof StatsCollectionTaskIndependent;
for (Map.Entry<String, FSPaths> entry : valToPaths.entrySet()) {
String fspKey = entry.getKey(); // DP/LB
FSPaths fspValue = entry.getValue();
// for bucketed tables, hive.optimize.sort.dynamic.partition optimization
// adds the taskId to the fspKey.
if (conf.getDpSortState().equals(DPSortState.PARTITION_BUCKET_SORTED)) {
taskID = Utilities.getTaskIdFromFilename(fspKey);
// if length of (prefix/ds=__HIVE_DEFAULT_PARTITION__/000000_0) is greater than max key prefix
// and if (prefix/ds=10/000000_0) is less than max key prefix, then former will get hashed
// to a smaller prefix (MD5hash/000000_0) and later will stored as such in staging stats table.
// When stats gets aggregated in StatsTask only the keys that starts with "prefix" will be fetched.
// Now that (prefix/ds=__HIVE_DEFAULT_PARTITION__) is hashed to a smaller prefix it will
// not be retrieved from staging table and hence not aggregated. To avoid this issue
// we will remove the taskId from the key which is redundant anyway.
fspKey = fspKey.split(taskID)[0];
}
// split[0] = DP, split[1] = LB
String[] split = splitKey(fspKey);
String dpSpec = split[0];
String lbSpec = split[1];
String prefix;
String postfix=null;
if (taskIndependent) {
// key = "database.table/SP/DP/"LB/
prefix = conf.getTableInfo().getTableName();
} else {
// key = "prefix/SP/DP/"LB/taskID/
prefix = conf.getStatsAggPrefix();
postfix = Utilities.join(lbSpec, taskID);
}
prefix = Utilities.join(prefix, spSpec, dpSpec);
prefix = Utilities.getHashedStatsPrefix(prefix, maxKeyLength);
String key = Utilities.join(prefix, postfix);
Map<String, String> statsToPublish = new HashMap<String, String>();
for (String statType : fspValue.stat.getStoredStats()) {
statsToPublish.put(statType, Long.toString(fspValue.stat.getStat(statType)));
}
if (!statsPublisher.publishStat(key, statsToPublish)) {
// The original exception is lost.
// Not changing the interface to maintain backward compatibility
if (isStatsReliable) {
throw new HiveException(ErrorMsg.STATSPUBLISHER_PUBLISHING_ERROR.getErrorCodedMsg());
}
}
}
if (!statsPublisher.closeConnection()) {
// The original exception is lost.
// Not changing the interface to maintain backward compatibility
if (isStatsReliable) {
throw new HiveException(ErrorMsg.STATSPUBLISHER_CLOSING_ERROR.getErrorCodedMsg());
}
}
}
/**
* This is server side code to create key in order to save statistics to stats database.
* Client side will read it via StatsTask.java aggregateStats().
* Client side reads it via db query prefix which is based on partition spec.
* Since store-as-subdir information is not part of partition spec, we have to
* remove store-as-subdir information from variable "keyPrefix" calculation.
* But we have to keep store-as-subdir information in variable "key" calculation
* since each skewed value has a row in stats db and "key" is db key,
* otherwise later value overwrites previous value.
* Performance impact due to string handling is minimum since this method is
* only called once in FileSinkOperator closeOp().
* For example,
* create table test skewed by (key, value) on (('484','val_484') stored as DIRECTORIES;
* skewedValueDirList contains 2 elements:
* 1. key=484/value=val_484
* 2. HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME
* Case #1: Static partition with store-as-sub-dir
* spSpec has SP path
* fspKey has either
* key=484/value=val_484 or
* HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME
* After filter, fspKey is empty, storedAsDirPostFix has either
* key=484/value=val_484 or
* HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME
* so, at the end, "keyPrefix" doesnt have subdir information but "key" has
* Case #2: Dynamic partition with store-as-sub-dir. Assume dp part is hr
* spSpec has SP path
* fspKey has either
* hr=11/key=484/value=val_484 or
* hr=11/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME
* After filter, fspKey is hr=11, storedAsDirPostFix has either
* key=484/value=val_484 or
* HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME/HIVE_LIST_BUCKETING_DEFAULT_DIR_NAME
* so, at the end, "keyPrefix" doesn't have subdir information from skewed but "key" has
*
* In a word, fspKey is consists of DP(dynamic partition spec) + LB(list bucketing spec)
* In stats publishing, full partition spec consists of prefix part of stat key
* but list bucketing spec is regarded as a postfix of stat key. So we split it here.
*/
private String[] splitKey(String fspKey) {
if (!fspKey.isEmpty() && isSkewedStoredAsSubDirectories) {
for (String dir : lbCtx.getSkewedValuesDirNames()) {
int index = fspKey.indexOf(dir);
if (index >= 0) {
return new String[] {fspKey.substring(0, index), fspKey.substring(index + 1)};
}
}
}
return new String[] {fspKey, null};
}
}