private void createMapReduce4Merge(FileSinkOperator fsOp, GenMRProcContext ctx, String finalName)
throws SemanticException {
Task<? extends Serializable> currTask = ctx.getCurrTask();
RowSchema inputRS = fsOp.getSchema();
// create a reduce Sink operator - key is the first column
ArrayList<ExprNodeDesc> keyCols = new ArrayList<ExprNodeDesc>();
keyCols.add(TypeCheckProcFactory.DefaultExprProcessor.getFuncExprNodeDesc("rand"));
// value is all the columns in the FileSink operator input
ArrayList<ExprNodeDesc> valueCols = new ArrayList<ExprNodeDesc>();
for (ColumnInfo ci : inputRS.getSignature()) {
valueCols.add(
new ExprNodeColumnDesc(
ci.getType(), ci.getInternalName(), ci.getTabAlias(), ci.getIsVirtualCol()));
}
// create a dummy tableScan operator
Operator<? extends Serializable> tsMerge = OperatorFactory.get(TableScanDesc.class, inputRS);
ArrayList<String> outputColumns = new ArrayList<String>();
for (int i = 0; i < valueCols.size(); i++) {
outputColumns.add(SemanticAnalyzer.getColumnInternalName(i));
}
ReduceSinkDesc rsDesc =
PlanUtils.getReduceSinkDesc(
new ArrayList<ExprNodeDesc>(), valueCols, outputColumns, false, -1, -1, -1);
OperatorFactory.getAndMakeChild(rsDesc, inputRS, tsMerge);
ParseContext parseCtx = ctx.getParseCtx();
FileSinkDesc fsConf = fsOp.getConf();
// Add the extract operator to get the value fields
RowResolver out_rwsch = new RowResolver();
RowResolver interim_rwsch = ctx.getParseCtx().getOpParseCtx().get(fsOp).getRowResolver();
Integer pos = Integer.valueOf(0);
for (ColumnInfo colInfo : interim_rwsch.getColumnInfos()) {
String[] info = interim_rwsch.reverseLookup(colInfo.getInternalName());
out_rwsch.put(
info[0],
info[1],
new ColumnInfo(
pos.toString(),
colInfo.getType(),
info[0],
colInfo.getIsVirtualCol(),
colInfo.isHiddenVirtualCol()));
pos = Integer.valueOf(pos.intValue() + 1);
}
Operator<ExtractDesc> extract =
OperatorFactory.getAndMakeChild(
new ExtractDesc(
new ExprNodeColumnDesc(
TypeInfoFactory.stringTypeInfo,
Utilities.ReduceField.VALUE.toString(),
"",
false)),
new RowSchema(out_rwsch.getColumnInfos()));
TableDesc ts = (TableDesc) fsConf.getTableInfo().clone();
fsConf
.getTableInfo()
.getProperties()
.remove(org.apache.hadoop.hive.metastore.api.Constants.META_TABLE_PARTITION_COLUMNS);
FileSinkDesc newFSD =
new FileSinkDesc(
finalName, ts, parseCtx.getConf().getBoolVar(HiveConf.ConfVars.COMPRESSRESULT));
FileSinkOperator newOutput =
(FileSinkOperator) OperatorFactory.getAndMakeChild(newFSD, inputRS, extract);
HiveConf conf = parseCtx.getConf();
MapredWork cplan = createMergeTask(conf, tsMerge, fsConf);
cplan.setReducer(extract);
// NOTE: we should gather stats in MR1 (rather than the merge MR job)
// since it is unknown if the merge MR will be triggered at execution time.
MoveWork dummyMv =
new MoveWork(
null,
null,
null,
new LoadFileDesc(fsConf.getDirName(), finalName, true, null, null),
false);
ConditionalTask cndTsk = createCondTask(conf, currTask, dummyMv, cplan, fsConf.getDirName());
LinkMoveTask(ctx, newOutput, cndTsk);
}
/**
* 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);
}
