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);
}
private void initializeOperators(Map<FetchOperator, JobConf> fetchOpJobConfMap)
throws HiveException {
for (Map.Entry<String, Operator<? extends OperatorDesc>> entry :
work.getAliasToWork().entrySet()) {
LOG.debug(
"initializeOperators: "
+ entry.getKey()
+ ", children = "
+ entry.getValue().getChildOperators());
}
// this mapper operator is used to initialize all the operators
for (Map.Entry<String, FetchWork> entry : work.getAliasToFetchWork().entrySet()) {
if (entry.getValue() == null) {
continue;
}
JobConf jobClone = new JobConf(job);
TableScanOperator ts = (TableScanOperator) work.getAliasToWork().get(entry.getKey());
// push down projections
ColumnProjectionUtils.appendReadColumns(
jobClone, ts.getNeededColumnIDs(), ts.getNeededColumns());
// push down filters
HiveInputFormat.pushFilters(jobClone, ts);
// create a fetch operator
FetchOperator fetchOp = new FetchOperator(entry.getValue(), jobClone);
fetchOpJobConfMap.put(fetchOp, jobClone);
fetchOperators.put(entry.getKey(), fetchOp);
l4j.info("fetchoperator for " + entry.getKey() + " created");
}
// initialize all forward operator
for (Map.Entry<String, FetchOperator> entry : fetchOperators.entrySet()) {
// get the forward op
String alias = entry.getKey();
Operator<? extends OperatorDesc> forwardOp = work.getAliasToWork().get(alias);
// put the exe context into all the operators
forwardOp.passExecContext(execContext);
// All the operators need to be initialized before process
FetchOperator fetchOp = entry.getValue();
JobConf jobConf = fetchOpJobConfMap.get(fetchOp);
if (jobConf == null) {
jobConf = job;
}
// initialize the forward operator
ObjectInspector objectInspector = fetchOp.getOutputObjectInspector();
forwardOp.initialize(jobConf, new ObjectInspector[] {objectInspector});
l4j.info("fetchoperator for " + entry.getKey() + " initialized");
}
}
