@Override
@SuppressWarnings("unchecked")
public Object aggregate(Collection<?> countersList) {
if (countersList.isEmpty()) {
return null;
}
BasicCounters<MutableLong> tempFileCounters =
(BasicCounters<MutableLong>) countersList.iterator().next();
MutableLong globalProcessedFiles =
tempFileCounters.getCounter(FileCounters.GLOBAL_PROCESSED_FILES);
MutableLong globalNumberOfFailures =
tempFileCounters.getCounter(FileCounters.GLOBAL_NUMBER_OF_FAILURES);
MutableLong globalNumberOfRetries =
tempFileCounters.getCounter(FileCounters.GLOBAL_NUMBER_OF_RETRIES);
totalLocalProcessedFiles.setValue(0);
pendingFiles.setValue(0);
totalLocalNumberOfFailures.setValue(0);
totalLocalNumberOfRetries.setValue(0);
for (Object fileCounters : countersList) {
BasicCounters<MutableLong> basicFileCounters = (BasicCounters<MutableLong>) fileCounters;
totalLocalProcessedFiles.add(
basicFileCounters.getCounter(FileCounters.LOCAL_PROCESSED_FILES));
pendingFiles.add(basicFileCounters.getCounter(FileCounters.PENDING_FILES));
totalLocalNumberOfFailures.add(
basicFileCounters.getCounter(FileCounters.LOCAL_NUMBER_OF_FAILURES));
totalLocalNumberOfRetries.add(
basicFileCounters.getCounter(FileCounters.LOCAL_NUMBER_OF_RETRIES));
}
globalProcessedFiles.add(totalLocalProcessedFiles);
globalProcessedFiles.subtract(pendingFiles);
globalNumberOfFailures.add(totalLocalNumberOfFailures);
globalNumberOfRetries.add(totalLocalNumberOfRetries);
BasicCounters<MutableLong> aggregatedCounters =
new BasicCounters<MutableLong>(MutableLong.class);
aggregatedCounters.setCounter(AggregatedFileCounters.PROCESSED_FILES, globalProcessedFiles);
aggregatedCounters.setCounter(AggregatedFileCounters.PENDING_FILES, pendingFiles);
aggregatedCounters.setCounter(
AggregatedFileCounters.NUMBER_OF_ERRORS, totalLocalNumberOfFailures);
aggregatedCounters.setCounter(
AggregatedFileCounters.NUMBER_OF_RETRIES, totalLocalNumberOfRetries);
return aggregatedCounters;
}
@Override
public Collection<Partition<AbstractFileInputOperator<T>>> definePartitions(
Collection<Partition<AbstractFileInputOperator<T>>> partitions, PartitioningContext context) {
lastRepartition = System.currentTimeMillis();
int totalCount = getNewPartitionCount(partitions, context);
LOG.debug("Computed new partitions: {}", totalCount);
if (totalCount == partitions.size()) {
return partitions;
}
AbstractFileInputOperator<T> tempOperator =
partitions.iterator().next().getPartitionedInstance();
MutableLong tempGlobalNumberOfRetries = tempOperator.globalNumberOfRetries;
MutableLong tempGlobalNumberOfFailures = tempOperator.globalNumberOfRetries;
/*
* Build collective state from all instances of the operator.
*/
Set<String> totalProcessedFiles = Sets.newHashSet();
Set<FailedFile> currentFiles = Sets.newHashSet();
List<DirectoryScanner> oldscanners = Lists.newLinkedList();
List<FailedFile> totalFailedFiles = Lists.newLinkedList();
List<String> totalPendingFiles = Lists.newLinkedList();
Set<Integer> deletedOperators = Sets.newHashSet();
for (Partition<AbstractFileInputOperator<T>> partition : partitions) {
AbstractFileInputOperator<T> oper = partition.getPartitionedInstance();
totalProcessedFiles.addAll(oper.processedFiles);
totalFailedFiles.addAll(oper.failedFiles);
totalPendingFiles.addAll(oper.pendingFiles);
currentFiles.addAll(unfinishedFiles);
tempGlobalNumberOfRetries.add(oper.localNumberOfRetries);
tempGlobalNumberOfFailures.add(oper.localNumberOfFailures);
if (oper.currentFile != null) {
currentFiles.add(new FailedFile(oper.currentFile, oper.offset));
}
oldscanners.add(oper.getScanner());
deletedOperators.add(oper.operatorId);
}
/*
* Create partitions of scanners, scanner's partition method will do state
* transfer for DirectoryScanner objects.
*/
List<DirectoryScanner> scanners = scanner.partition(totalCount, oldscanners);
Kryo kryo = new Kryo();
Collection<Partition<AbstractFileInputOperator<T>>> newPartitions =
Lists.newArrayListWithExpectedSize(totalCount);
Collection<IdempotentStorageManager> newManagers =
Lists.newArrayListWithExpectedSize(totalCount);
for (int i = 0; i < scanners.size(); i++) {
// Kryo.copy fails as it attempts to clone transient fields
ByteArrayOutputStream bos = new ByteArrayOutputStream();
Output loutput = new Output(bos);
kryo.writeObject(loutput, this);
loutput.close();
Input lInput = new Input(bos.toByteArray());
@SuppressWarnings("unchecked")
AbstractFileInputOperator<T> oper = kryo.readObject(lInput, this.getClass());
lInput.close();
DirectoryScanner scn = scanners.get(i);
oper.setScanner(scn);
// Do state transfer for processed files.
oper.processedFiles.addAll(totalProcessedFiles);
oper.globalNumberOfFailures = tempGlobalNumberOfRetries;
oper.localNumberOfFailures.setValue(0);
oper.globalNumberOfRetries = tempGlobalNumberOfFailures;
oper.localNumberOfRetries.setValue(0);
/* redistribute unfinished files properly */
oper.unfinishedFiles.clear();
oper.currentFile = null;
oper.offset = 0;
Iterator<FailedFile> unfinishedIter = currentFiles.iterator();
while (unfinishedIter.hasNext()) {
FailedFile unfinishedFile = unfinishedIter.next();
if (scn.acceptFile(unfinishedFile.path)) {
oper.unfinishedFiles.add(unfinishedFile);
unfinishedIter.remove();
}
}
/* transfer failed files */
oper.failedFiles.clear();
Iterator<FailedFile> iter = totalFailedFiles.iterator();
while (iter.hasNext()) {
FailedFile ff = iter.next();
if (scn.acceptFile(ff.path)) {
oper.failedFiles.add(ff);
iter.remove();
}
}
/* redistribute pending files properly */
oper.pendingFiles.clear();
Iterator<String> pendingFilesIterator = totalPendingFiles.iterator();
while (pendingFilesIterator.hasNext()) {
String pathString = pendingFilesIterator.next();
if (scn.acceptFile(pathString)) {
oper.pendingFiles.add(pathString);
pendingFilesIterator.remove();
}
}
newPartitions.add(new DefaultPartition<AbstractFileInputOperator<T>>(oper));
newManagers.add(oper.idempotentStorageManager);
}
idempotentStorageManager.partitioned(newManagers, deletedOperators);
LOG.info("definePartitions called returning {} partitions", newPartitions.size());
return newPartitions;
}
