Examples with WorkItem com.google.api.services.dataflow.model.WorkItem used on opensource projects

Example 11 with WorkItem

use of com.google.api.services.dataflow.model.WorkItem in project beam by apache.

the class DataflowWorkUnitClient method getWorkItemInternal.

private Optional<WorkItem> getWorkItemInternal(List<String> workItemTypes, List<String> capabilities) throws IOException {
    LeaseWorkItemRequest request = new LeaseWorkItemRequest();
    request.setFactory(Transport.getJsonFactory());
    request.setWorkItemTypes(workItemTypes);
    request.setWorkerCapabilities(capabilities);
    request.setWorkerId(options.getWorkerId());
    request.setCurrentWorkerTime(toCloudTime(DateTime.now()));
    // This shouldn't be necessary, but a valid cloud duration string is
    // required by the Google API parsing framework.  TODO: Fix the framework
    // so that an empty or not-present string can be used as a default value.
    request.setRequestedLeaseDuration(toCloudDuration(Duration.millis(WorkProgressUpdater.DEFAULT_LEASE_DURATION_MILLIS)));
    logger.debug("Leasing work: {}", request);
    LeaseWorkItemResponse response = dataflow.projects().locations().jobs().workItems().lease(options.getProject(), options.getRegion(), options.getJobId(), request).execute();
    logger.debug("Lease work response: {}", response);
    List<WorkItem> workItems = response.getWorkItems();
    if (workItems == null || workItems.isEmpty()) {
        // We didn't lease any work.
        return Optional.absent();
    } else if (workItems.size() > 1) {
        throw new IOException("This version of the SDK expects no more than one work item from the service: " + response);
    }
    WorkItem work = response.getWorkItems().get(0);
    // Looks like the work's a'ight.
    return Optional.of(work);
}

Example 12 with WorkItem

use of com.google.api.services.dataflow.model.WorkItem in project beam by apache.

the class StreamingDataflowWorker method sendWorkerUpdatesToDataflowService.

/**
 * Sends counter updates to Dataflow backend.
 */
private void sendWorkerUpdatesToDataflowService(CounterSet deltaCounters, CounterSet cumulativeCounters) throws IOException {
    // Throttle time is tracked by the windmillServer but is reported to DFE here.
    windmillQuotaThrottling.addValue(windmillServer.getAndResetThrottleTime());
    if (memoryMonitor.isThrashing()) {
        memoryThrashing.addValue(1);
    }
    List<CounterUpdate> counterUpdates = new ArrayList<>(128);
    if (publishCounters) {
        stageInfoMap.values().forEach(s -> counterUpdates.addAll(s.extractCounterUpdates()));
        counterUpdates.addAll(cumulativeCounters.extractUpdates(false, DataflowCounterUpdateExtractor.INSTANCE));
        counterUpdates.addAll(deltaCounters.extractModifiedDeltaUpdates(DataflowCounterUpdateExtractor.INSTANCE));
        if (hasExperiment(options, "beam_fn_api")) {
            Map<Object, List<CounterUpdate>> fnApiCounters = new HashMap<>();
            while (!this.pendingMonitoringInfos.isEmpty()) {
                final CounterUpdate item = this.pendingMonitoringInfos.poll();
                // WorkItem.
                if (item.getCumulative()) {
                    item.setCumulative(false);
                    // Group counterUpdates by counterUpdateKey so they can be aggregated before sending to
                    // dataflow service.
                    fnApiCounters.computeIfAbsent(getCounterUpdateKey(item), k -> new ArrayList<>()).add(item);
                } else {
                    // This is a safety check in case new counter type appears in FnAPI.
                    throw new UnsupportedOperationException("FnApi counters are expected to provide cumulative values." + " Please, update conversion to delta logic" + " if non-cumulative counter type is required.");
                }
            }
            // so we can avoid excessive I/Os for reporting to dataflow service.
            for (List<CounterUpdate> counterUpdateList : fnApiCounters.values()) {
                if (counterUpdateList.isEmpty()) {
                    continue;
                }
                List<CounterUpdate> aggregatedCounterUpdateList = CounterUpdateAggregators.aggregate(counterUpdateList);
                // updates.
                if (aggregatedCounterUpdateList.size() > 10) {
                    CounterUpdate head = aggregatedCounterUpdateList.get(0);
                    this.counterAggregationErrorCount.getAndIncrement();
                    // log warning message only when error count is the power of 2 to avoid spamming.
                    if (this.counterAggregationErrorCount.get() > 10 && Long.bitCount(this.counterAggregationErrorCount.get()) == 1) {
                        LOG.warn("Found non-aggregated counter updates of size {} with kind {}, this will likely " + "cause performance degradation and excessive GC if size is large.", counterUpdateList.size(), MoreObjects.firstNonNull(head.getNameAndKind(), head.getStructuredNameAndMetadata()));
                    }
                }
                counterUpdates.addAll(aggregatedCounterUpdateList);
            }
        }
    }
    // Handle duplicate counters from different stages. Store all the counters in a multi-map and
    // send the counters that appear multiple times in separate RPCs. Same logical counter could
    // appear in multiple stages if a step runs in multiple stages (as with flatten-unzipped stages)
    // especially if the counter definition does not set execution_step_name.
    ListMultimap<Object, CounterUpdate> counterMultimap = MultimapBuilder.hashKeys(counterUpdates.size()).linkedListValues().build();
    boolean hasDuplicates = false;
    for (CounterUpdate c : counterUpdates) {
        Object key = getCounterUpdateKey(c);
        if (counterMultimap.containsKey(key)) {
            hasDuplicates = true;
        }
        counterMultimap.put(key, c);
    }
    // Clears counterUpdates and enqueues unique counters from counterMultimap. If a counter
    // appears more than once, one of them is extracted leaving the remaining in the map.
    Runnable extractUniqueCounters = () -> {
        counterUpdates.clear();
        for (Iterator<Object> iter = counterMultimap.keySet().iterator(); iter.hasNext(); ) {
            List<CounterUpdate> counters = counterMultimap.get(iter.next());
            counterUpdates.add(counters.get(0));
            if (counters.size() == 1) {
                // There is single value. Remove the entry through the iterator.
                iter.remove();
            } else {
                // Otherwise remove the first value.
                counters.remove(0);
            }
        }
    };
    if (hasDuplicates) {
        extractUniqueCounters.run();
    } else {
        // Common case: no duplicates. We can just send counterUpdates, empty the multimap.
        counterMultimap.clear();
    }
    List<Status> errors;
    synchronized (pendingFailuresToReport) {
        errors = new ArrayList<>(pendingFailuresToReport.size());
        for (String stackTrace : pendingFailuresToReport) {
            errors.add(new Status().setCode(// rpc.Code.UNKNOWN
            2).setMessage(stackTrace));
        }
        // Best effort only, no need to wait till successfully sent.
        pendingFailuresToReport.clear();
    }
    WorkItemStatus workItemStatus = new WorkItemStatus().setWorkItemId(WINDMILL_COUNTER_UPDATE_WORK_ID).setErrors(errors).setCounterUpdates(counterUpdates);
    workUnitClient.reportWorkItemStatus(workItemStatus);
    // Send any counters appearing more than once in subsequent RPCs:
    while (!counterMultimap.isEmpty()) {
        extractUniqueCounters.run();
        workUnitClient.reportWorkItemStatus(new WorkItemStatus().setWorkItemId(WINDMILL_COUNTER_UPDATE_WORK_ID).setCounterUpdates(counterUpdates));
    }
}

Example 13 with WorkItem

use of com.google.api.services.dataflow.model.WorkItem in project beam by apache.

the class StreamingDataflowWorker method process.

private void process(final SdkWorkerHarness worker, final ComputationState computationState, final Instant inputDataWatermark, @Nullable final Instant outputDataWatermark, @Nullable final Instant synchronizedProcessingTime, final Work work) {
    final Windmill.WorkItem workItem = work.getWorkItem();
    final String computationId = computationState.getComputationId();
    final ByteString key = workItem.getKey();
    work.setState(State.PROCESSING);
    {
        StringBuilder workIdBuilder = new StringBuilder(33);
        workIdBuilder.append(Long.toHexString(workItem.getShardingKey()));
        workIdBuilder.append('-');
        workIdBuilder.append(Long.toHexString(workItem.getWorkToken()));
        DataflowWorkerLoggingMDC.setWorkId(workIdBuilder.toString());
    }
    DataflowWorkerLoggingMDC.setStageName(computationId);
    LOG.debug("Starting processing for {}:\n{}", computationId, work);
    Windmill.WorkItemCommitRequest.Builder outputBuilder = initializeOutputBuilder(key, workItem);
    // Before any processing starts, call any pending OnCommit callbacks.  Nothing that requires
    // cleanup should be done before this, since we might exit early here.
    callFinalizeCallbacks(workItem);
    if (workItem.getSourceState().getOnlyFinalize()) {
        outputBuilder.setSourceStateUpdates(Windmill.SourceState.newBuilder().setOnlyFinalize(true));
        work.setState(State.COMMIT_QUEUED);
        commitQueue.put(new Commit(outputBuilder.build(), computationState, work));
        return;
    }
    long processingStartTimeNanos = System.nanoTime();
    final MapTask mapTask = computationState.getMapTask();
    StageInfo stageInfo = stageInfoMap.computeIfAbsent(mapTask.getStageName(), s -> new StageInfo(s, mapTask.getSystemName(), this));
    ExecutionState executionState = null;
    try {
        executionState = computationState.getExecutionStateQueue(worker).poll();
        if (executionState == null) {
            MutableNetwork<Node, Edge> mapTaskNetwork = mapTaskToNetwork.apply(mapTask);
            if (LOG.isDebugEnabled()) {
                LOG.debug("Network as Graphviz .dot: {}", Networks.toDot(mapTaskNetwork));
            }
            ParallelInstructionNode readNode = (ParallelInstructionNode) Iterables.find(mapTaskNetwork.nodes(), node -> node instanceof ParallelInstructionNode && ((ParallelInstructionNode) node).getParallelInstruction().getRead() != null);
            InstructionOutputNode readOutputNode = (InstructionOutputNode) Iterables.getOnlyElement(mapTaskNetwork.successors(readNode));
            DataflowExecutionContext.DataflowExecutionStateTracker executionStateTracker = new DataflowExecutionContext.DataflowExecutionStateTracker(ExecutionStateSampler.instance(), stageInfo.executionStateRegistry.getState(NameContext.forStage(mapTask.getStageName()), "other", null, ScopedProfiler.INSTANCE.emptyScope()), stageInfo.deltaCounters, options, computationId);
            StreamingModeExecutionContext context = new StreamingModeExecutionContext(pendingDeltaCounters, computationId, readerCache, !computationState.getTransformUserNameToStateFamily().isEmpty() ? computationState.getTransformUserNameToStateFamily() : stateNameMap, stateCache.forComputation(computationId), stageInfo.metricsContainerRegistry, executionStateTracker, stageInfo.executionStateRegistry, maxSinkBytes);
            DataflowMapTaskExecutor mapTaskExecutor = mapTaskExecutorFactory.create(worker.getControlClientHandler(), worker.getGrpcDataFnServer(), sdkHarnessRegistry.beamFnDataApiServiceDescriptor(), worker.getGrpcStateFnServer(), mapTaskNetwork, options, mapTask.getStageName(), readerRegistry, sinkRegistry, context, pendingDeltaCounters, idGenerator);
            ReadOperation readOperation = mapTaskExecutor.getReadOperation();
            // Disable progress updates since its results are unused  for streaming
            // and involves starting a thread.
            readOperation.setProgressUpdatePeriodMs(ReadOperation.DONT_UPDATE_PERIODICALLY);
            Preconditions.checkState(mapTaskExecutor.supportsRestart(), "Streaming runner requires all operations support restart.");
            Coder<?> readCoder;
            readCoder = CloudObjects.coderFromCloudObject(CloudObject.fromSpec(readOutputNode.getInstructionOutput().getCodec()));
            Coder<?> keyCoder = extractKeyCoder(readCoder);
            // If using a custom source, count bytes read for autoscaling.
            if (CustomSources.class.getName().equals(readNode.getParallelInstruction().getRead().getSource().getSpec().get("@type"))) {
                NameContext nameContext = NameContext.create(mapTask.getStageName(), readNode.getParallelInstruction().getOriginalName(), readNode.getParallelInstruction().getSystemName(), readNode.getParallelInstruction().getName());
                readOperation.receivers[0].addOutputCounter(new OutputObjectAndByteCounter(new IntrinsicMapTaskExecutorFactory.ElementByteSizeObservableCoder<>(readCoder), mapTaskExecutor.getOutputCounters(), nameContext).setSamplingPeriod(100).countBytes("dataflow_input_size-" + mapTask.getSystemName()));
            }
            executionState = new ExecutionState(mapTaskExecutor, context, keyCoder, executionStateTracker);
        }
        WindmillStateReader stateReader = new WindmillStateReader(metricTrackingWindmillServer, computationId, key, workItem.getShardingKey(), workItem.getWorkToken());
        StateFetcher localStateFetcher = stateFetcher.byteTrackingView();
        // If the read output KVs, then we can decode Windmill's byte key into a userland
        // key object and provide it to the execution context for use with per-key state.
        // Otherwise, we pass null.
        // 
        // The coder type that will be present is:
        // WindowedValueCoder(TimerOrElementCoder(KvCoder))
        @Nullable Coder<?> keyCoder = executionState.getKeyCoder();
        @Nullable Object executionKey = keyCoder == null ? null : keyCoder.decode(key.newInput(), Coder.Context.OUTER);
        if (workItem.hasHotKeyInfo()) {
            Windmill.HotKeyInfo hotKeyInfo = workItem.getHotKeyInfo();
            Duration hotKeyAge = Duration.millis(hotKeyInfo.getHotKeyAgeUsec() / 1000);
            // The MapTask instruction is ordered by dependencies, such that the first element is
            // always going to be the shuffle task.
            String stepName = computationState.getMapTask().getInstructions().get(0).getName();
            if (options.isHotKeyLoggingEnabled() && keyCoder != null) {
                hotKeyLogger.logHotKeyDetection(stepName, hotKeyAge, executionKey);
            } else {
                hotKeyLogger.logHotKeyDetection(stepName, hotKeyAge);
            }
        }
        executionState.getContext().start(executionKey, workItem, inputDataWatermark, outputDataWatermark, synchronizedProcessingTime, stateReader, localStateFetcher, outputBuilder);
        // Blocks while executing work.
        executionState.getWorkExecutor().execute();
        Iterables.addAll(this.pendingMonitoringInfos, executionState.getWorkExecutor().extractMetricUpdates());
        commitCallbacks.putAll(executionState.getContext().flushState());
        // Release the execution state for another thread to use.
        computationState.getExecutionStateQueue(worker).offer(executionState);
        executionState = null;
        // Add the output to the commit queue.
        work.setState(State.COMMIT_QUEUED);
        WorkItemCommitRequest commitRequest = outputBuilder.build();
        int byteLimit = maxWorkItemCommitBytes;
        int commitSize = commitRequest.getSerializedSize();
        int estimatedCommitSize = commitSize < 0 ? Integer.MAX_VALUE : commitSize;
        // Detect overflow of integer serialized size or if the byte limit was exceeded.
        windmillMaxObservedWorkItemCommitBytes.addValue(estimatedCommitSize);
        if (commitSize < 0 || commitSize > byteLimit) {
            KeyCommitTooLargeException e = KeyCommitTooLargeException.causedBy(computationId, byteLimit, commitRequest);
            reportFailure(computationId, workItem, e);
            LOG.error(e.toString());
            // Drop the current request in favor of a new, minimal one requesting truncation.
            // Messages, timers, counters, and other commit content will not be used by the service
            // so we're purposefully dropping them here
            commitRequest = buildWorkItemTruncationRequest(key, workItem, estimatedCommitSize);
        }
        commitQueue.put(new Commit(commitRequest, computationState, work));
        // Compute shuffle and state byte statistics these will be flushed asynchronously.
        long stateBytesWritten = outputBuilder.clearOutputMessages().build().getSerializedSize();
        long shuffleBytesRead = 0;
        for (Windmill.InputMessageBundle bundle : workItem.getMessageBundlesList()) {
            for (Windmill.Message message : bundle.getMessagesList()) {
                shuffleBytesRead += message.getSerializedSize();
            }
        }
        long stateBytesRead = stateReader.getBytesRead() + localStateFetcher.getBytesRead();
        windmillShuffleBytesRead.addValue(shuffleBytesRead);
        windmillStateBytesRead.addValue(stateBytesRead);
        windmillStateBytesWritten.addValue(stateBytesWritten);
        LOG.debug("Processing done for work token: {}", workItem.getWorkToken());
    } catch (Throwable t) {
        if (executionState != null) {
            try {
                executionState.getContext().invalidateCache();
                executionState.getWorkExecutor().close();
            } catch (Exception e) {
                LOG.warn("Failed to close map task executor: ", e);
            } finally {
                // Release references to potentially large objects early.
                executionState = null;
            }
        }
        t = t instanceof UserCodeException ? t.getCause() : t;
        boolean retryLocally = false;
        if (KeyTokenInvalidException.isKeyTokenInvalidException(t)) {
            LOG.debug("Execution of work for computation '{}' on key '{}' failed due to token expiration. " + "Work will not be retried locally.", computationId, key.toStringUtf8());
        } else {
            LastExceptionDataProvider.reportException(t);
            LOG.debug("Failed work: {}", work);
            Duration elapsedTimeSinceStart = new Duration(Instant.now(), work.getStartTime());
            if (!reportFailure(computationId, workItem, t)) {
                LOG.error("Execution of work for computation '{}' on key '{}' failed with uncaught exception, " + "and Windmill indicated not to retry locally.", computationId, key.toStringUtf8(), t);
            } else if (isOutOfMemoryError(t)) {
                File heapDump = memoryMonitor.tryToDumpHeap();
                LOG.error("Execution of work for computation '{}' for key '{}' failed with out-of-memory. " + "Work will not be retried locally. Heap dump {}.", computationId, key.toStringUtf8(), heapDump == null ? "not written" : ("written to '" + heapDump + "'"), t);
            } else if (elapsedTimeSinceStart.isLongerThan(MAX_LOCAL_PROCESSING_RETRY_DURATION)) {
                LOG.error("Execution of work for computation '{}' for key '{}' failed with uncaught exception, " + "and it will not be retried locally because the elapsed time since start {} " + "exceeds {}.", computationId, key.toStringUtf8(), elapsedTimeSinceStart, MAX_LOCAL_PROCESSING_RETRY_DURATION, t);
            } else {
                LOG.error("Execution of work for computation '{}' on key '{}' failed with uncaught exception. " + "Work will be retried locally.", computationId, key.toStringUtf8(), t);
                retryLocally = true;
            }
        }
        if (retryLocally) {
            // Try again after some delay and at the end of the queue to avoid a tight loop.
            sleep(retryLocallyDelayMs);
            workUnitExecutor.forceExecute(work, work.getWorkItem().getSerializedSize());
        } else {
            // Consider the item invalid. It will eventually be retried by Windmill if it still needs to
            // be processed.
            computationState.completeWork(ShardedKey.create(key, workItem.getShardingKey()), workItem.getWorkToken());
        }
    } finally {
        // Update total processing time counters. Updating in finally clause ensures that
        // work items causing exceptions are also accounted in time spent.
        long processingTimeMsecs = TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - processingStartTimeNanos);
        stageInfo.totalProcessingMsecs.addValue(processingTimeMsecs);
        // either here or in DFE.
        if (work.getWorkItem().hasTimers()) {
            stageInfo.timerProcessingMsecs.addValue(processingTimeMsecs);
        }
        DataflowWorkerLoggingMDC.setWorkId(null);
        DataflowWorkerLoggingMDC.setStageName(null);
    }
}

Example 14 with WorkItem

use of com.google.api.services.dataflow.model.WorkItem in project beam by apache.

the class StreamingDataflowWorker method streamingDispatchLoop.

void streamingDispatchLoop() {
    while (running.get()) {
        GetWorkStream stream = windmillServer.getWorkStream(Windmill.GetWorkRequest.newBuilder().setClientId(clientId).setMaxItems(chooseMaximumBundlesOutstanding()).setMaxBytes(MAX_GET_WORK_FETCH_BYTES).build(), (String computation, Instant inputDataWatermark, Instant synchronizedProcessingTime, Windmill.WorkItem workItem) -> {
            memoryMonitor.waitForResources("GetWork");
            scheduleWorkItem(getComputationState(computation), inputDataWatermark, synchronizedProcessingTime, workItem);
        });
        try {
            // we half-close the stream after some time and create a new one.
            if (!stream.awaitTermination(GET_WORK_STREAM_TIMEOUT_MINUTES, TimeUnit.MINUTES)) {
                stream.close();
            }
        } catch (InterruptedException e) {
        // Continue processing until !running.get()
        }
    }
}

Example 15 with WorkItem

use of com.google.api.services.dataflow.model.WorkItem in project beam by apache.

the class DataflowWorkUnitClientTest method testCloudServiceCallMapTaskStagePropagation.

@Test
public void testCloudServiceCallMapTaskStagePropagation() throws Exception {
    WorkUnitClient client = new DataflowWorkUnitClient(pipelineOptions, LOG);
    // Publish and acquire a map task work item, and verify we're now processing that stage.
    final String stageName = "test_stage_name";
    MapTask mapTask = new MapTask();
    mapTask.setStageName(stageName);
    WorkItem workItem = createWorkItem(PROJECT_ID, JOB_ID);
    workItem.setMapTask(mapTask);
    when(request.execute()).thenReturn(generateMockResponse(workItem));
    assertEquals(Optional.of(workItem), client.getWorkItem());
    assertEquals(stageName, DataflowWorkerLoggingMDC.getStageName());
}