/**
* Construct a conditional task given the current leaf task, the MoveWork and the MapredWork.
*
* @param conf HiveConf
* @param currTask current leaf task
* @param mvWork MoveWork for the move task
* @param mergeWork MapredWork for the merge task.
* @param inputPath the input directory of the merge/move task
* @return The conditional task
*/
private ConditionalTask createCondTask(
HiveConf conf,
Task<? extends Serializable> currTask,
MoveWork mvWork,
MapredWork mergeWork,
String inputPath) {
Task<? extends Serializable> mergeTask = TaskFactory.get(mergeWork, conf);
Task<? extends Serializable> moveTask = TaskFactory.get(mvWork, conf);
List<Serializable> listWorks = new ArrayList<Serializable>();
listWorks.add(mvWork);
listWorks.add(mergeWork);
ConditionalWork cndWork = new ConditionalWork(listWorks);
List<Task<? extends Serializable>> listTasks = new ArrayList<Task<? extends Serializable>>();
listTasks.add(moveTask);
listTasks.add(mergeTask);
ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork, conf);
cndTsk.setListTasks(listTasks);
// create resolver
cndTsk.setResolver(new ConditionalResolverMergeFiles());
ConditionalResolverMergeFilesCtx mrCtx =
new ConditionalResolverMergeFilesCtx(listTasks, inputPath);
cndTsk.setResolverCtx(mrCtx);
// make the conditional task as the child of the current leaf task
currTask.addDependentTask(cndTsk);
return cndTsk;
}
/**
* Construct a conditional task given the current leaf task, the MoveWork and the MapredWork.
*
* @param conf HiveConf
* @param currTask current leaf task
* @param mvWork MoveWork for the move task
* @param mergeWork MapredWork for the merge task.
* @param inputPath the input directory of the merge/move task
* @return The conditional task
*/
private ConditionalTask createCondTask(
HiveConf conf,
Task<? extends Serializable> currTask,
MoveWork mvWork,
MapredWork mergeWork,
String inputPath) {
// There are 3 options for this ConditionalTask:
// 1) Merge the partitions
// 2) Move the partitions (i.e. don't merge the partitions)
// 3) Merge some partitions and move other partitions (i.e. merge some partitions and don't
// merge others) in this case the merge is done first followed by the move to prevent
// conflicts.
Task<? extends Serializable> mergeOnlyMergeTask = TaskFactory.get(mergeWork, conf);
Task<? extends Serializable> moveOnlyMoveTask = TaskFactory.get(mvWork, conf);
Task<? extends Serializable> mergeAndMoveMergeTask = TaskFactory.get(mergeWork, conf);
Task<? extends Serializable> mergeAndMoveMoveTask = TaskFactory.get(mvWork, conf);
// NOTE! It is necessary merge task is the parent of the move task, and not
// the other way around, for the proper execution of the execute method of
// ConditionalTask
mergeAndMoveMergeTask.addDependentTask(mergeAndMoveMoveTask);
List<Serializable> listWorks = new ArrayList<Serializable>();
listWorks.add(mvWork);
listWorks.add(mergeWork);
ConditionalWork cndWork = new ConditionalWork(listWorks);
List<Task<? extends Serializable>> listTasks = new ArrayList<Task<? extends Serializable>>();
listTasks.add(moveOnlyMoveTask);
listTasks.add(mergeOnlyMergeTask);
listTasks.add(mergeAndMoveMergeTask);
ConditionalTask cndTsk = (ConditionalTask) TaskFactory.get(cndWork, conf);
cndTsk.setListTasks(listTasks);
// create resolver
cndTsk.setResolver(new ConditionalResolverMergeFiles());
ConditionalResolverMergeFilesCtx mrCtx =
new ConditionalResolverMergeFilesCtx(listTasks, inputPath);
cndTsk.setResolverCtx(mrCtx);
// make the conditional task as the child of the current leaf task
currTask.addDependentTask(cndTsk);
return cndTsk;
}
/**
* Make the move task in the GenMRProcContext following the FileSinkOperator a dependent of all
* possible subtrees branching from the ConditionalTask.
*
* @param ctx
* @param newOutput
* @param cndTsk
*/
private void linkMoveTask(
GenMRProcContext ctx, FileSinkOperator newOutput, ConditionalTask cndTsk) {
List<Task<MoveWork>> mvTasks = ctx.getMvTask();
Task<MoveWork> mvTask = findMoveTask(mvTasks, newOutput);
for (Task<? extends Serializable> tsk : cndTsk.getListTasks()) {
linkMoveTask(ctx, mvTask, tsk);
}
}
private void LinkMoveTask(
GenMRProcContext ctx, FileSinkOperator newOutput, ConditionalTask cndTsk) {
List<Task<? extends Serializable>> mvTasks = ctx.getMvTask();
Task<? extends Serializable> mvTask = findMoveTask(mvTasks, newOutput);
if (mvTask != null) {
for (Task<? extends Serializable> tsk : cndTsk.getListTasks()) {
tsk.addDependentTask(mvTask);
}
}
}
/**
* create a Map-only merge job with the following operators:
*
* @param fsInput
* @param ctx
* @param finalName MR job J0: ... | v FileSinkOperator_1 (fsInput) | v Merge job J1: | v
* TableScan (using CombineHiveInputFormat) (tsMerge) | v FileSinkOperator (fsMerge)
* <p>Here the pathToPartitionInfo & pathToAlias will remain the same, which means the paths
* do not contain the dynamic partitions (their parent). So after the dynamic partitions are
* created (after the first job finished before the moveTask or ConditionalTask start), we
* need to change the pathToPartitionInfo & pathToAlias to include the dynamic partition
* directories.
*/
private void createMap4Merge(FileSinkOperator fsInput, GenMRProcContext ctx, String finalName) {
//
// 1. create the operator tree
//
ParseContext parseCtx = ctx.getParseCtx();
FileSinkDesc fsInputDesc = fsInput.getConf();
// Create a TableScan operator
RowSchema inputRS = fsInput.getSchema();
Operator<? extends Serializable> tsMerge = OperatorFactory.get(TableScanDesc.class, inputRS);
// Create a FileSink operator
TableDesc ts = (TableDesc) fsInputDesc.getTableInfo().clone();
FileSinkDesc fsOutputDesc =
new FileSinkDesc(
finalName, ts, parseCtx.getConf().getBoolVar(HiveConf.ConfVars.COMPRESSRESULT));
FileSinkOperator fsOutput =
(FileSinkOperator) OperatorFactory.getAndMakeChild(fsOutputDesc, inputRS, tsMerge);
// If the input FileSinkOperator is a dynamic partition enabled, the tsMerge input schema
// needs to include the partition column, and the fsOutput should have
// a DynamicPartitionCtx to indicate that it needs to dynamically partitioned.
DynamicPartitionCtx dpCtx = fsInputDesc.getDynPartCtx();
if (dpCtx != null && dpCtx.getNumDPCols() > 0) {
// adding DP ColumnInfo to the RowSchema signature
ArrayList<ColumnInfo> signature = inputRS.getSignature();
String tblAlias = fsInputDesc.getTableInfo().getTableName();
LinkedHashMap<String, String> colMap = new LinkedHashMap<String, String>();
StringBuilder partCols = new StringBuilder();
for (String dpCol : dpCtx.getDPColNames()) {
ColumnInfo colInfo =
new ColumnInfo(
dpCol,
TypeInfoFactory.stringTypeInfo, // all partition column type should be string
tblAlias,
true); // partition column is virtual column
signature.add(colInfo);
colMap.put(dpCol, dpCol); // input and output have the same column name
partCols.append(dpCol).append('/');
}
partCols.setLength(partCols.length() - 1); // remove the last '/'
inputRS.setSignature(signature);
// create another DynamicPartitionCtx, which has a different input-to-DP column mapping
DynamicPartitionCtx dpCtx2 = new DynamicPartitionCtx(dpCtx);
dpCtx2.setInputToDPCols(colMap);
fsOutputDesc.setDynPartCtx(dpCtx2);
// update the FileSinkOperator to include partition columns
fsInputDesc
.getTableInfo()
.getProperties()
.setProperty(
org.apache.hadoop.hive.metastore.api.Constants.META_TABLE_PARTITION_COLUMNS,
partCols.toString()); // list of dynamic partition column names
} else {
// non-partitioned table
fsInputDesc
.getTableInfo()
.getProperties()
.remove(org.apache.hadoop.hive.metastore.api.Constants.META_TABLE_PARTITION_COLUMNS);
}
//
// 2. Constructing a conditional task consisting of a move task and a map reduce task
//
MapRedTask currTask = (MapRedTask) ctx.getCurrTask();
MoveWork dummyMv =
new MoveWork(
null,
null,
null,
new LoadFileDesc(fsInputDesc.getDirName(), finalName, true, null, null),
false);
MapredWork cplan = createMergeTask(ctx.getConf(), tsMerge, fsInputDesc);
// use CombineHiveInputFormat for map-only merging
cplan.setInputformat("org.apache.hadoop.hive.ql.io.CombineHiveInputFormat");
// NOTE: we should gather stats in MR1 rather than MR2 at merge job since we don't
// know if merge MR2 will be triggered at execution time
ConditionalTask cndTsk =
createCondTask(ctx.getConf(), ctx.getCurrTask(), dummyMv, cplan, fsInputDesc.getDirName());
// keep the dynamic partition context in conditional task resolver context
ConditionalResolverMergeFilesCtx mrCtx =
(ConditionalResolverMergeFilesCtx) cndTsk.getResolverCtx();
mrCtx.setDPCtx(fsInputDesc.getDynPartCtx());
//
// 3. add the moveTask as the children of the conditional task
//
LinkMoveTask(ctx, fsOutput, cndTsk);
}
/**
* @param fsInput The FileSink operator.
* @param ctx The MR processing context.
* @param finalName the final destination path the merge job should output.
* @throws SemanticException
* <p>create a Map-only merge job using CombineHiveInputFormat for all partitions with
* following operators: MR job J0: ... | v FileSinkOperator_1 (fsInput) | v Merge job J1: | v
* TableScan (using CombineHiveInputFormat) (tsMerge) | v FileSinkOperator (fsMerge)
* <p>Here the pathToPartitionInfo & pathToAlias will remain the same, which means the paths
* do not contain the dynamic partitions (their parent). So after the dynamic partitions are
* created (after the first job finished before the moveTask or ConditionalTask start), we
* need to change the pathToPartitionInfo & pathToAlias to include the dynamic partition
* directories.
*/
private void createMRWorkForMergingFiles(
FileSinkOperator fsInput, GenMRProcContext ctx, String finalName) throws SemanticException {
//
// 1. create the operator tree
//
HiveConf conf = ctx.getParseCtx().getConf();
FileSinkDesc fsInputDesc = fsInput.getConf();
// Create a TableScan operator
RowSchema inputRS = fsInput.getSchema();
Operator<? extends OperatorDesc> tsMerge = OperatorFactory.get(TableScanDesc.class, inputRS);
// Create a FileSink operator
TableDesc ts = (TableDesc) fsInputDesc.getTableInfo().clone();
FileSinkDesc fsOutputDesc =
new FileSinkDesc(finalName, ts, conf.getBoolVar(ConfVars.COMPRESSRESULT));
FileSinkOperator fsOutput =
(FileSinkOperator) OperatorFactory.getAndMakeChild(fsOutputDesc, inputRS, tsMerge);
// If the input FileSinkOperator is a dynamic partition enabled, the tsMerge input schema
// needs to include the partition column, and the fsOutput should have
// a DynamicPartitionCtx to indicate that it needs to dynamically partitioned.
DynamicPartitionCtx dpCtx = fsInputDesc.getDynPartCtx();
if (dpCtx != null && dpCtx.getNumDPCols() > 0) {
// adding DP ColumnInfo to the RowSchema signature
ArrayList<ColumnInfo> signature = inputRS.getSignature();
String tblAlias = fsInputDesc.getTableInfo().getTableName();
LinkedHashMap<String, String> colMap = new LinkedHashMap<String, String>();
StringBuilder partCols = new StringBuilder();
for (String dpCol : dpCtx.getDPColNames()) {
ColumnInfo colInfo =
new ColumnInfo(
dpCol,
TypeInfoFactory.stringTypeInfo, // all partition column type should be string
tblAlias,
true); // partition column is virtual column
signature.add(colInfo);
colMap.put(dpCol, dpCol); // input and output have the same column name
partCols.append(dpCol).append('/');
}
partCols.setLength(partCols.length() - 1); // remove the last '/'
inputRS.setSignature(signature);
// create another DynamicPartitionCtx, which has a different input-to-DP column mapping
DynamicPartitionCtx dpCtx2 = new DynamicPartitionCtx(dpCtx);
dpCtx2.setInputToDPCols(colMap);
fsOutputDesc.setDynPartCtx(dpCtx2);
// update the FileSinkOperator to include partition columns
fsInputDesc
.getTableInfo()
.getProperties()
.setProperty(
org.apache
.hadoop
.hive
.metastore
.api
.hive_metastoreConstants
.META_TABLE_PARTITION_COLUMNS,
partCols.toString()); // list of dynamic partition column names
} else {
// non-partitioned table
fsInputDesc
.getTableInfo()
.getProperties()
.remove(
org.apache
.hadoop
.hive
.metastore
.api
.hive_metastoreConstants
.META_TABLE_PARTITION_COLUMNS);
}
//
// 2. Constructing a conditional task consisting of a move task and a map reduce task
//
MoveWork dummyMv =
new MoveWork(
null,
null,
null,
new LoadFileDesc(fsInputDesc.getFinalDirName(), finalName, true, null, null),
false);
MapredWork cplan;
if (conf.getBoolVar(ConfVars.HIVEMERGERCFILEBLOCKLEVEL)
&& fsInputDesc.getTableInfo().getInputFileFormatClass().equals(RCFileInputFormat.class)) {
// Check if InputFormatClass is valid
String inputFormatClass = conf.getVar(ConfVars.HIVEMERGERCFILEINPUTFORMATBLOCKLEVEL);
try {
Class c = (Class<? extends InputFormat>) Class.forName(inputFormatClass);
LOG.info("RCFile format- Using block level merge");
cplan =
createBlockMergeTask(
fsInputDesc,
finalName,
dpCtx != null && dpCtx.getNumDPCols() > 0,
RCFileMergeMapper.class,
RCFileInputFormat.class,
RCFileBlockMergeInputFormat.class);
} catch (ClassNotFoundException e) {
String msg = "Illegal input format class: " + inputFormatClass;
throw new SemanticException(msg);
}
} else if (conf.getBoolVar(ConfVars.HIVEMERGEORCBLOCKLEVEL)
&& fsInputDesc.getTableInfo().getInputFileFormatClass().equals(OrcInputFormat.class)) {
// Check if InputFormatClass is valid
String inputFormatClass = conf.getVar(ConfVars.HIVEMERGEORCINPUTFORMATBLOCKLEVEL);
try {
Class c = (Class<? extends InputFormat>) Class.forName(inputFormatClass);
LOG.info("ORCFile format- Using block level merge");
cplan =
createBlockMergeTask(
fsInputDesc,
finalName,
dpCtx != null && dpCtx.getNumDPCols() > 0,
OrcMergeMapper.class,
OrcInputFormat.class,
OrcBlockMergeInputFormat.class);
} catch (ClassNotFoundException e) {
String msg = "Illegal input format class: " + inputFormatClass;
throw new SemanticException(msg);
}
} else {
cplan = createMRWorkForMergingFiles(conf, tsMerge, fsInputDesc);
// use CombineHiveInputFormat for map-only merging
}
cplan.setInputformat("org.apache.hadoop.hive.ql.io.CombineHiveInputFormat");
// NOTE: we should gather stats in MR1 rather than MR2 at merge job since we don't
// know if merge MR2 will be triggered at execution time
ConditionalTask cndTsk =
createCondTask(conf, ctx.getCurrTask(), dummyMv, cplan, fsInputDesc.getFinalDirName());
// keep the dynamic partition context in conditional task resolver context
ConditionalResolverMergeFilesCtx mrCtx =
(ConditionalResolverMergeFilesCtx) cndTsk.getResolverCtx();
mrCtx.setDPCtx(fsInputDesc.getDynPartCtx());
mrCtx.setLbCtx(fsInputDesc.getLbCtx());
//
// 3. add the moveTask as the children of the conditional task
//
linkMoveTask(ctx, fsOutput, cndTsk);
}
