public void addKnownInput(
String hostName,
int port,
InputAttemptIdentifier srcAttemptIdentifier,
int srcPhysicalIndex) {
String identifier = InputHost.createIdentifier(hostName, port);
InputHost host = knownSrcHosts.get(identifier);
if (host == null) {
host = new InputHost(hostName, port, inputContext.getApplicationId(), srcPhysicalIndex);
assert identifier.equals(host.getIdentifier());
InputHost old = knownSrcHosts.putIfAbsent(identifier, host);
if (old != null) {
host = old;
}
}
if (LOG.isDebugEnabled()) {
LOG.debug("Adding input: " + srcAttemptIdentifier + ", to host: " + host);
}
host.addKnownInput(srcAttemptIdentifier);
lock.lock();
try {
boolean added = pendingHosts.offer(host);
if (!added) {
String errorMessage = "Unable to add host: " + host.getIdentifier() + " to pending queue";
LOG.error(errorMessage);
throw new TezUncheckedException(errorMessage);
}
wakeLoop.signal();
} finally {
lock.unlock();
}
}
private Fetcher constructFetcherForHost(InputHost inputHost) {
FetcherBuilder fetcherBuilder =
new FetcherBuilder(
ShuffleManager.this,
httpConnectionParams,
inputManager,
inputContext.getApplicationId(),
shuffleSecret,
srcNameTrimmed);
if (codec != null) {
fetcherBuilder.setCompressionParameters(codec);
}
fetcherBuilder.setIFileParams(ifileReadAhead, ifileReadAheadLength);
// Remove obsolete inputs from the list being given to the fetcher. Also
// remove from the obsolete list.
List<InputAttemptIdentifier> pendingInputsForHost = inputHost.clearAndGetPendingInputs();
for (Iterator<InputAttemptIdentifier> inputIter = pendingInputsForHost.iterator();
inputIter.hasNext(); ) {
InputAttemptIdentifier input = inputIter.next();
// Avoid adding attempts which have already completed.
if (completedInputSet.contains(input.getInputIdentifier())) {
inputIter.remove();
continue;
}
// Avoid adding attempts which have been marked as OBSOLETE
if (obsoletedInputs.contains(input)) {
inputIter.remove();
}
}
// TODO NEWTEZ Maybe limit the number of inputs being given to a single
// fetcher, especially in the case where #hosts < #fetchers
fetcherBuilder.assignWork(
inputHost.getHost(),
inputHost.getPort(),
inputHost.getSrcPhysicalIndex(),
pendingInputsForHost);
LOG.info(
"Created Fetcher for host: "
+ inputHost.getHost()
+ ", with inputs: "
+ pendingInputsForHost);
return fetcherBuilder.build();
}
@Override
public void onSuccess(FetchResult result) {
fetcher.shutdown();
if (isShutdown.get()) {
LOG.info("Already shutdown. Ignoring event from fetcher");
} else {
Iterable<InputAttemptIdentifier> pendingInputs = result.getPendingInputs();
if (pendingInputs != null && pendingInputs.iterator().hasNext()) {
InputHost inputHost =
knownSrcHosts.get(InputHost.createIdentifier(result.getHost(), result.getPort()));
assert inputHost != null;
for (InputAttemptIdentifier input : pendingInputs) {
inputHost.addKnownInput(input);
}
pendingHosts.add(inputHost);
}
doBookKeepingForFetcherComplete();
}
}
@Override
public Void call() throws Exception {
while (!isShutdown.get() && numCompletedInputs.get() < numInputs) {
lock.lock();
try {
if (runningFetchers.size() >= numFetchers || pendingHosts.isEmpty()) {
if (numCompletedInputs.get() < numInputs) {
wakeLoop.await();
}
}
} finally {
lock.unlock();
}
if (shuffleError != null) {
// InputContext has already been informed of a fatal error. Relying on
// tez to kill the task.
break;
}
if (LOG.isDebugEnabled()) {
LOG.debug("NumCompletedInputs: " + numCompletedInputs);
}
if (numCompletedInputs.get() < numInputs && !isShutdown.get()) {
lock.lock();
try {
int maxFetchersToRun = numFetchers - runningFetchers.size();
int count = 0;
while (pendingHosts.peek() != null && !isShutdown.get()) {
InputHost inputHost = null;
try {
inputHost = pendingHosts.take();
} catch (InterruptedException e) {
if (isShutdown.get()) {
LOG.info(
"Interrupted and hasBeenShutdown, Breaking out of ShuffleScheduler Loop");
break;
} else {
throw e;
}
}
if (LOG.isDebugEnabled()) {
LOG.debug("Processing pending host: " + inputHost.toDetailedString());
}
if (inputHost.getNumPendingInputs() > 0 && !isShutdown.get()) {
LOG.info("Scheduling fetch for inputHost: " + inputHost.getIdentifier());
Fetcher fetcher = constructFetcherForHost(inputHost);
runningFetchers.add(fetcher);
if (isShutdown.get()) {
LOG.info("hasBeenShutdown, Breaking out of ShuffleScheduler Loop");
}
ListenableFuture<FetchResult> future = fetcherExecutor.submit(fetcher);
Futures.addCallback(future, new FetchFutureCallback(fetcher));
if (++count >= maxFetchersToRun) {
break;
}
} else {
if (LOG.isDebugEnabled()) {
LOG.debug(
"Skipping host: "
+ inputHost.getIdentifier()
+ " since it has no inputs to process");
}
}
}
} finally {
lock.unlock();
}
}
}
LOG.info(
"Shutting down FetchScheduler, Was Interrupted: "
+ Thread.currentThread().isInterrupted());
// TODO NEWTEZ Maybe clean up inputs.
if (!fetcherExecutor.isShutdown()) {
fetcherExecutor.shutdownNow();
}
return null;
}
