Пример #1
0
  @Override
  public void fetchFailed(
      String host, InputAttemptIdentifier srcAttemptIdentifier, boolean connectFailed) {
    // TODO NEWTEZ. Implement logic to report fetch failures after a threshold.
    // For now, reporting immediately.
    LOG.info(
        "Fetch failed for src: "
            + srcAttemptIdentifier
            + "InputIdentifier: "
            + srcAttemptIdentifier
            + ", connectFailed: "
            + connectFailed);
    failedShufflesCounter.increment(1);
    if (srcAttemptIdentifier == null) {
      String message = "Received fetchFailure for an unknown src (null)";
      LOG.fatal(message);
      inputContext.fatalError(null, message);
    } else {
      InputReadErrorEvent readError =
          new InputReadErrorEvent(
              "Fetch failure while fetching from "
                  + TezRuntimeUtils.getTaskAttemptIdentifier(
                      inputContext.getSourceVertexName(),
                      srcAttemptIdentifier.getInputIdentifier().getInputIndex(),
                      srcAttemptIdentifier.getAttemptNumber()),
              srcAttemptIdentifier.getInputIdentifier().getInputIndex(),
              srcAttemptIdentifier.getAttemptNumber());

      List<Event> failedEvents = Lists.newArrayListWithCapacity(1);
      failedEvents.add(readError);
      inputContext.sendEvents(failedEvents);
    }
  }
Пример #2
0
  @Override
  public void fetchSucceeded(
      String host,
      InputAttemptIdentifier srcAttemptIdentifier,
      FetchedInput fetchedInput,
      long fetchedBytes,
      long decompressedLength,
      long copyDuration)
      throws IOException {
    InputIdentifier inputIdentifier = srcAttemptIdentifier.getInputIdentifier();

    LOG.info(
        "Completed fetch for attempt: " + srcAttemptIdentifier + " to " + fetchedInput.getType());

    // Count irrespective of whether this is a copy of an already fetched input
    lock.lock();
    try {
      lastProgressTime = System.currentTimeMillis();
    } finally {
      lock.unlock();
    }

    boolean committed = false;
    if (!completedInputSet.contains(inputIdentifier)) {
      synchronized (completedInputSet) {
        if (!completedInputSet.contains(inputIdentifier)) {
          fetchedInput.commit();
          committed = true;

          // Processing counters for completed and commit fetches only. Need
          // additional counters for excessive fetches - which primarily comes
          // in after speculation or retries.
          shuffledInputsCounter.increment(1);
          bytesShuffledCounter.increment(fetchedBytes);
          if (fetchedInput.getType() == Type.MEMORY) {
            bytesShuffledToMemCounter.increment(fetchedBytes);
          } else {
            bytesShuffledToDiskCounter.increment(fetchedBytes);
          }
          decompressedDataSizeCounter.increment(decompressedLength);

          registerCompletedInput(fetchedInput);
        }
      }
    }
    if (!committed) {
      fetchedInput.abort(); // If this fails, the fetcher may attempt another abort.
    } else {
      lock.lock();
      try {
        // Signal the wakeLoop to check for termination.
        wakeLoop.signal();
      } finally {
        lock.unlock();
      }
    }
    // TODO NEWTEZ Maybe inform fetchers, in case they have an alternate attempt of the same task in
    // their queue.
  }
Пример #3
0
  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();
  }
  /**
   * Create the map-output-url. This will contain all the map ids separated by commas
   *
   * @param host
   * @param maps
   * @return
   * @throws MalformedURLException
   */
  private URL getMapOutputURL(MapHost host, List<InputAttemptIdentifier> srcAttempts)
      throws MalformedURLException {
    // Get the base url
    StringBuffer url = new StringBuffer(host.getBaseUrl());

    boolean first = true;
    for (InputAttemptIdentifier mapId : srcAttempts) {
      if (!first) {
        url.append(",");
      }
      url.append(mapId.getPathComponent());
      first = false;
    }

    if (LOG.isDebugEnabled()) {
      LOG.debug("MapOutput URL for " + host + " -> " + url.toString());
    }
    return new URL(url.toString());
  }
Пример #5
0
  public void addCompletedInputWithNoData(InputAttemptIdentifier srcAttemptIdentifier) {
    InputIdentifier inputIdentifier = srcAttemptIdentifier.getInputIdentifier();
    LOG.info("No input data exists for SrcTask: " + inputIdentifier + ". Marking as complete.");

    if (!completedInputSet.contains(inputIdentifier)) {
      synchronized (completedInputSet) {
        if (!completedInputSet.contains(inputIdentifier)) {
          registerCompletedInput(new NullFetchedInput(srcAttemptIdentifier));
        }
      }
    }

    // Awake the loop to check for termination.
    lock.lock();
    try {
      wakeLoop.signal();
    } finally {
      lock.unlock();
    }
  }
Пример #6
0
  public void addCompletedInputWithData(
      InputAttemptIdentifier srcAttemptIdentifier, FetchedInput fetchedInput) throws IOException {
    InputIdentifier inputIdentifier = srcAttemptIdentifier.getInputIdentifier();

    LOG.info("Received Data via Event: " + srcAttemptIdentifier + " to " + fetchedInput.getType());
    // Count irrespective of whether this is a copy of an already fetched input
    lock.lock();
    try {
      lastProgressTime = System.currentTimeMillis();
    } finally {
      lock.unlock();
    }

    boolean committed = false;
    if (!completedInputSet.contains(inputIdentifier)) {
      synchronized (completedInputSet) {
        if (!completedInputSet.contains(inputIdentifier)) {
          fetchedInput.commit();
          committed = true;
          registerCompletedInput(fetchedInput);
        }
      }
    }
    if (!committed) {
      fetchedInput.abort(); // If this fails, the fetcher may attempt another
      // abort.
    } else {
      lock.lock();
      try {
        // Signal the wakeLoop to check for termination.
        wakeLoop.signal();
      } finally {
        lock.unlock();
      }
    }
  }