private void handleSampling(DriverContext context, MapWork mWork, JobConf job, HiveConf conf)
throws Exception {
assert mWork.getAliasToWork().keySet().size() == 1;
String alias = mWork.getAliases().get(0);
Operator<?> topOp = mWork.getAliasToWork().get(alias);
PartitionDesc partDesc = mWork.getAliasToPartnInfo().get(alias);
ArrayList<String> paths = mWork.getPaths();
ArrayList<PartitionDesc> parts = mWork.getPartitionDescs();
List<Path> inputPaths = new ArrayList<Path>(paths.size());
for (String path : paths) {
inputPaths.add(new Path(path));
}
Path tmpPath = context.getCtx().getExternalTmpPath(inputPaths.get(0));
Path partitionFile = new Path(tmpPath, ".partitions");
ShimLoader.getHadoopShims().setTotalOrderPartitionFile(job, partitionFile);
PartitionKeySampler sampler = new PartitionKeySampler();
if (mWork.getSamplingType() == MapWork.SAMPLING_ON_PREV_MR) {
console.printInfo("Use sampling data created in previous MR");
// merges sampling data from previous MR and make partition keys for total sort
for (Path path : inputPaths) {
FileSystem fs = path.getFileSystem(job);
for (FileStatus status : fs.globStatus(new Path(path, ".sampling*"))) {
sampler.addSampleFile(status.getPath(), job);
}
}
} else if (mWork.getSamplingType() == MapWork.SAMPLING_ON_START) {
console.printInfo("Creating sampling data..");
assert topOp instanceof TableScanOperator;
TableScanOperator ts = (TableScanOperator) topOp;
FetchWork fetchWork;
if (!partDesc.isPartitioned()) {
assert paths.size() == 1;
fetchWork = new FetchWork(inputPaths.get(0), partDesc.getTableDesc());
} else {
fetchWork = new FetchWork(inputPaths, parts, partDesc.getTableDesc());
}
fetchWork.setSource(ts);
// random sampling
FetchOperator fetcher = PartitionKeySampler.createSampler(fetchWork, conf, job, ts);
try {
ts.initialize(conf, new ObjectInspector[] {fetcher.getOutputObjectInspector()});
OperatorUtils.setChildrenCollector(ts.getChildOperators(), sampler);
while (fetcher.pushRow()) {}
} finally {
fetcher.clearFetchContext();
}
} else {
throw new IllegalArgumentException("Invalid sampling type " + mWork.getSamplingType());
}
sampler.writePartitionKeys(partitionFile, conf, job);
}
/**
* Traverse all the partitions for a table, and get the OI for the table. Note that a conversion
* is required if any of the partition OI is different from the table OI. For eg. if the query
* references table T (partitions P1, P2), and P1's schema is same as T, whereas P2's scheme is
* different from T, conversion might be needed for both P1 and P2, since SettableOI might be
* needed for T
*/
private Map<TableDesc, StructObjectInspector> getConvertedOI(Configuration hconf)
throws HiveException {
Map<TableDesc, StructObjectInspector> tableDescOI =
new HashMap<TableDesc, StructObjectInspector>();
Set<TableDesc> identityConverterTableDesc = new HashSet<TableDesc>();
try {
Map<ObjectInspector, Boolean> oiSettableProperties = new HashMap<ObjectInspector, Boolean>();
for (String onefile : conf.getPathToAliases().keySet()) {
PartitionDesc pd = conf.getPathToPartitionInfo().get(onefile);
TableDesc tableDesc = pd.getTableDesc();
Deserializer partDeserializer = pd.getDeserializer(hconf);
StructObjectInspector partRawRowObjectInspector;
if (Utilities.isInputFileFormatSelfDescribing(pd)) {
Deserializer tblDeserializer = tableDesc.getDeserializer(hconf);
partRawRowObjectInspector = (StructObjectInspector) tblDeserializer.getObjectInspector();
} else {
partRawRowObjectInspector = (StructObjectInspector) partDeserializer.getObjectInspector();
}
StructObjectInspector tblRawRowObjectInspector = tableDescOI.get(tableDesc);
if ((tblRawRowObjectInspector == null)
|| (identityConverterTableDesc.contains(tableDesc))) {
Deserializer tblDeserializer = tableDesc.getDeserializer(hconf);
tblRawRowObjectInspector =
(StructObjectInspector)
ObjectInspectorConverters.getConvertedOI(
partRawRowObjectInspector,
tblDeserializer.getObjectInspector(),
oiSettableProperties);
if (identityConverterTableDesc.contains(tableDesc)) {
if (!partRawRowObjectInspector.equals(tblRawRowObjectInspector)) {
identityConverterTableDesc.remove(tableDesc);
}
} else if (partRawRowObjectInspector.equals(tblRawRowObjectInspector)) {
identityConverterTableDesc.add(tableDesc);
}
tableDescOI.put(tableDesc, tblRawRowObjectInspector);
}
}
} catch (Exception e) {
throw new HiveException(e);
}
return tableDescOI;
}
/*
* This is the same as the setChildren method below but for empty tables.
* It takes care of the following:
* 1. Create the right object inspector.
* 2. Set up the childrenOpToOI with the object inspector.
* So as to ensure that the initialization happens correctly.
*/
public void initEmptyInputChildren(List<Operator<?>> children, Configuration hconf)
throws SerDeException, Exception {
setChildOperators(children);
for (Operator<?> child : children) {
TableScanOperator tsOp = (TableScanOperator) child;
StructObjectInspector soi = null;
PartitionDesc partDesc = conf.getAliasToPartnInfo().get(tsOp.getConf().getAlias());
Deserializer serde = partDesc.getTableDesc().getDeserializer();
partDesc.setProperties(partDesc.getProperties());
MapOpCtx opCtx = new MapOpCtx(tsOp.getConf().getAlias(), child, partDesc);
StructObjectInspector tableRowOI = (StructObjectInspector) serde.getObjectInspector();
initObjectInspector(hconf, opCtx, tableRowOI);
soi = opCtx.rowObjectInspector;
child.getParentOperators().add(this);
childrenOpToOI.put(child, soi);
}
}
public void setChildren(Configuration hconf) throws Exception {
List<Operator<? extends OperatorDesc>> children =
new ArrayList<Operator<? extends OperatorDesc>>();
Map<TableDesc, StructObjectInspector> convertedOI = getConvertedOI(hconf);
for (Map.Entry<String, ArrayList<String>> entry : conf.getPathToAliases().entrySet()) {
String onefile = entry.getKey();
List<String> aliases = entry.getValue();
PartitionDesc partDesc = conf.getPathToPartitionInfo().get(onefile);
for (String alias : aliases) {
Operator<? extends OperatorDesc> op = conf.getAliasToWork().get(alias);
if (isLogDebugEnabled) {
LOG.debug("Adding alias " + alias + " to work list for file " + onefile);
}
Map<Operator<?>, MapOpCtx> contexts = opCtxMap.get(onefile);
if (contexts == null) {
opCtxMap.put(onefile, contexts = new LinkedHashMap<Operator<?>, MapOpCtx>());
}
if (contexts.containsKey(op)) {
continue;
}
MapOpCtx context = new MapOpCtx(alias, op, partDesc);
StructObjectInspector tableRowOI = convertedOI.get(partDesc.getTableDesc());
contexts.put(op, initObjectInspector(hconf, context, tableRowOI));
if (children.contains(op) == false) {
op.setParentOperators(new ArrayList<Operator<? extends OperatorDesc>>(1));
op.getParentOperators().add(this);
children.add(op);
}
}
}
initOperatorContext(children);
// we found all the operators that we are supposed to process.
setChildOperators(children);
}
private MapOpCtx initObjectInspector(
Configuration hconf, MapOpCtx opCtx, StructObjectInspector tableRowOI) throws Exception {
PartitionDesc pd = opCtx.partDesc;
TableDesc td = pd.getTableDesc();
// Use table properties in case of unpartitioned tables,
// and the union of table properties and partition properties, with partition
// taking precedence, in the case of partitioned tables
Properties overlayedProps =
SerDeUtils.createOverlayedProperties(td.getProperties(), pd.getProperties());
Map<String, String> partSpec = pd.getPartSpec();
opCtx.tableName = String.valueOf(overlayedProps.getProperty("name"));
opCtx.partName = String.valueOf(partSpec);
opCtx.deserializer = pd.getDeserializer(hconf);
StructObjectInspector partRawRowObjectInspector;
if (Utilities.isInputFileFormatSelfDescribing(pd)) {
partRawRowObjectInspector = tableRowOI;
} else {
partRawRowObjectInspector = (StructObjectInspector) opCtx.deserializer.getObjectInspector();
}
opCtx.partTblObjectInspectorConverter =
ObjectInspectorConverters.getConverter(partRawRowObjectInspector, tableRowOI);
// Next check if this table has partitions and if so
// get the list of partition names as well as allocate
// the serdes for the partition columns
String pcols = overlayedProps.getProperty(hive_metastoreConstants.META_TABLE_PARTITION_COLUMNS);
if (pcols != null && pcols.length() > 0) {
String[] partKeys = pcols.trim().split("/");
String pcolTypes =
overlayedProps.getProperty(hive_metastoreConstants.META_TABLE_PARTITION_COLUMN_TYPES);
String[] partKeyTypes = pcolTypes.trim().split(":");
if (partKeys.length > partKeyTypes.length) {
throw new HiveException(
"Internal error : partKeys length, "
+ partKeys.length
+ " greater than partKeyTypes length, "
+ partKeyTypes.length);
}
List<String> partNames = new ArrayList<String>(partKeys.length);
Object[] partValues = new Object[partKeys.length];
List<ObjectInspector> partObjectInspectors = new ArrayList<ObjectInspector>(partKeys.length);
for (int i = 0; i < partKeys.length; i++) {
String key = partKeys[i];
partNames.add(key);
ObjectInspector oi =
PrimitiveObjectInspectorFactory.getPrimitiveWritableObjectInspector(
TypeInfoFactory.getPrimitiveTypeInfo(partKeyTypes[i]));
// Partitions do not exist for this table
if (partSpec == null) {
// for partitionless table, initialize partValue to null
partValues[i] = null;
} else {
partValues[i] =
ObjectInspectorConverters.getConverter(
PrimitiveObjectInspectorFactory.javaStringObjectInspector, oi)
.convert(partSpec.get(key));
}
partObjectInspectors.add(oi);
}
opCtx.rowWithPart = new Object[] {null, partValues};
opCtx.partObjectInspector =
ObjectInspectorFactory.getStandardStructObjectInspector(partNames, partObjectInspectors);
}
// The op may not be a TableScan for mapjoins
// Consider the query: select /*+MAPJOIN(a)*/ count(*) FROM T1 a JOIN T2 b ON a.key = b.key;
// In that case, it will be a Select, but the rowOI need not be amended
if (opCtx.op instanceof TableScanOperator) {
TableScanOperator tsOp = (TableScanOperator) opCtx.op;
TableScanDesc tsDesc = tsOp.getConf();
if (tsDesc != null && tsDesc.hasVirtualCols()) {
opCtx.vcs = tsDesc.getVirtualCols();
opCtx.vcValues = new Object[opCtx.vcs.size()];
opCtx.vcsObjectInspector = VirtualColumn.getVCSObjectInspector(opCtx.vcs);
if (opCtx.isPartitioned()) {
opCtx.rowWithPartAndVC = Arrays.copyOfRange(opCtx.rowWithPart, 0, 3);
} else {
opCtx.rowWithPartAndVC = new Object[2];
}
}
}
if (!opCtx.hasVC() && !opCtx.isPartitioned()) {
opCtx.rowObjectInspector = tableRowOI;
return opCtx;
}
List<StructObjectInspector> inspectors = new ArrayList<StructObjectInspector>();
inspectors.add(tableRowOI);
if (opCtx.isPartitioned()) {
inspectors.add(opCtx.partObjectInspector);
}
if (opCtx.hasVC()) {
inspectors.add(opCtx.vcsObjectInspector);
}
opCtx.rowObjectInspector = ObjectInspectorFactory.getUnionStructObjectInspector(inspectors);
return opCtx;
}
