use of org.apache.tez.dag.app.dag.Vertex in project tez by apache.
the class VertexImpl method checkTasksForCompletion.
// triggered by task_complete
static VertexState checkTasksForCompletion(final VertexImpl vertex) {
// this log helps quickly count the completion count for a vertex.
// grepping and counting for attempts and handling re-tries is time consuming
LOG.info("Task Completion: " + constructCheckTasksForCompletionLog(vertex));
// check for vertex failure first
if (vertex.completedTaskCount > vertex.tasks.size()) {
LOG.error("task completion accounting issue: completedTaskCount > nTasks:" + constructCheckTasksForCompletionLog(vertex));
}
if (vertex.completedTaskCount == vertex.tasks.size()) {
// finished - gather stats
vertex.finalStatistics = vertex.constructStatistics();
// Only succeed if tasks complete successfully and no terminationCause is registered or if failures are below configured threshold.
boolean vertexSucceeded = vertex.succeededTaskCount == vertex.numTasks;
boolean vertexFailuresBelowThreshold = (vertex.succeededTaskCount + vertex.failedTaskCount == vertex.numTasks) && (vertex.failedTaskCount * 100 <= vertex.maxFailuresPercent * vertex.numTasks);
if ((vertexSucceeded || vertexFailuresBelowThreshold) && vertex.terminationCause == null) {
if (vertexSucceeded) {
LOG.info("All tasks have succeeded, vertex:" + vertex.logIdentifier);
} else {
LOG.info("All tasks in the vertex " + vertex.logIdentifier + " have completed and the percentage of failed tasks (failed/total) (" + vertex.failedTaskCount + "/" + vertex.numTasks + ") is less that the threshold of " + vertex.maxFailuresPercent);
vertex.addDiagnostic("Vertex succeeded as percentage of failed tasks (failed/total) (" + vertex.failedTaskCount + "/" + vertex.numTasks + ") is less that the threshold of " + vertex.maxFailuresPercent);
vertex.logSuccessDiagnostics = true;
for (Task task : vertex.tasks.values()) {
if (!task.getState().equals(TaskState.FAILED)) {
continue;
}
// Find the last attempt and mark that as successful
Iterator<TezTaskAttemptID> attempts = task.getAttempts().keySet().iterator();
TezTaskAttemptID lastAttempt = null;
while (attempts.hasNext()) {
TezTaskAttemptID attempt = attempts.next();
if (lastAttempt == null || attempt.getId() > lastAttempt.getId()) {
lastAttempt = attempt;
}
}
LOG.info("Succeeding failed task attempt:" + lastAttempt);
for (Map.Entry<Vertex, Edge> vertexEdge : vertex.targetVertices.entrySet()) {
Vertex destVertex = vertexEdge.getKey();
Edge edge = vertexEdge.getValue();
try {
List<TezEvent> tezEvents = edge.generateEmptyEventsForAttempt(lastAttempt);
// Downstream vertices need to receive a SUCCEEDED completion event for each failed task to ensure num bipartite count is correct
VertexEventTaskAttemptCompleted completionEvent = new VertexEventTaskAttemptCompleted(lastAttempt, TaskAttemptStateInternal.SUCCEEDED);
// Notify all target vertices
vertex.eventHandler.handle(new VertexEventSourceTaskAttemptCompleted(destVertex.getVertexId(), completionEvent));
vertex.eventHandler.handle(new VertexEventRouteEvent(destVertex.getVertexId(), tezEvents));
} catch (Exception e) {
throw new TezUncheckedException(e);
}
}
}
}
if (vertex.commitVertexOutputs && !vertex.committed.getAndSet(true)) {
// start commit if there're commits or just finish if no commits
return commitOrFinish(vertex);
} else {
// just finish because no vertex committing needed
return vertex.finished(VertexState.SUCCEEDED);
}
}
return finishWithTerminationCause(vertex);
}
// return the current state, Vertex not finished yet
return vertex.getInternalState();
}
use of org.apache.tez.dag.app.dag.Vertex in project tez by apache.
the class VertexImpl method setParallelismWrapper.
private void setParallelismWrapper(int parallelism, VertexLocationHint vertexLocationHint, Map<String, EdgeProperty> sourceEdgeProperties, Map<String, InputSpecUpdate> rootInputSpecUpdates, boolean fromVertexManager) throws AMUserCodeException {
Preconditions.checkArgument(parallelism >= 0, "Parallelism must be >=0. Value: " + parallelism + " for vertex: " + logIdentifier);
writeLock.lock();
this.setParallelismCalledFlag = true;
try {
// disallow changing things after a vertex has started
if (!tasksNotYetScheduled) {
String msg = "setParallelism cannot be called after scheduling tasks. Vertex: " + getLogIdentifier();
LOG.info(msg);
throw new TezUncheckedException(msg);
}
if (fromVertexManager && canInitVertex()) {
// vertex is fully defined. setParallelism has been called. VertexManager should have
// informed us about this. Otherwise we would have notified listeners that we are fully
// defined before we are actually fully defined
Preconditions.checkState(vertexToBeReconfiguredByManager, "Vertex is fully configured but still" + " the reconfiguration API has been called. VertexManager must notify the framework using " + " context.vertexReconfigurationPlanned() before re-configuring the vertex." + " vertexId=" + logIdentifier);
}
// Input initializer/Vertex Manager/1-1 split expected to set parallelism.
if (numTasks == -1) {
if (getState() != VertexState.INITIALIZING) {
throw new TezUncheckedException("Vertex state is not Initializing. Value: " + getState() + " for vertex: " + logIdentifier);
}
if (sourceEdgeProperties != null) {
for (Map.Entry<String, EdgeProperty> entry : sourceEdgeProperties.entrySet()) {
LOG.info("Replacing edge manager for source:" + entry.getKey() + " destination: " + getLogIdentifier());
Vertex sourceVertex = appContext.getCurrentDAG().getVertex(entry.getKey());
Edge edge = sourceVertices.get(sourceVertex);
try {
edge.setEdgeProperty(entry.getValue());
} catch (Exception e) {
throw new TezUncheckedException("Fail to update EdgeProperty for Edge," + "sourceVertex:" + edge.getSourceVertexName() + "destinationVertex:" + edge.getDestinationVertexName(), e);
}
}
}
if (rootInputSpecUpdates != null) {
LOG.info("Got updated RootInputsSpecs: " + rootInputSpecUpdates.toString());
// Sanity check for correct number of updates.
for (Entry<String, InputSpecUpdate> rootInputSpecUpdateEntry : rootInputSpecUpdates.entrySet()) {
Preconditions.checkState(rootInputSpecUpdateEntry.getValue().isForAllWorkUnits() || (rootInputSpecUpdateEntry.getValue().getAllNumPhysicalInputs() != null && rootInputSpecUpdateEntry.getValue().getAllNumPhysicalInputs().size() == parallelism), "Not enough input spec updates for root input named " + rootInputSpecUpdateEntry.getKey());
}
this.rootInputSpecs.putAll(rootInputSpecUpdates);
}
int oldNumTasks = numTasks;
this.numTasks = parallelism;
stateChangeNotifier.stateChanged(vertexId, new VertexStateUpdateParallelismUpdated(vertexName, numTasks, oldNumTasks));
this.createTasks();
setVertexLocationHint(vertexLocationHint);
LOG.info("Vertex " + getLogIdentifier() + " parallelism set to " + parallelism);
if (canInitVertex()) {
getEventHandler().handle(new VertexEvent(getVertexId(), VertexEventType.V_READY_TO_INIT));
}
} else {
// This is an artificial restriction since there's no way of knowing whether a VertexManager
// will attempt to update root input specs. When parallelism has not been initialized, the
// Vertex will not be in started state so it's safe to update the specifications.
// TODO TEZ-937 - add e mechanism to query vertex managers, or for VMs to indicate readines
// for a vertex to start.
Preconditions.checkState(rootInputSpecUpdates == null, "Root Input specs can only be updated when the vertex is configured with -1 tasks");
int oldNumTasks = numTasks;
// start buffering incoming events so that we can re-route existing events
for (Edge edge : sourceVertices.values()) {
edge.startEventBuffering();
}
if (parallelism == numTasks) {
LOG.info("setParallelism same as current value: " + parallelism + " for vertex: " + logIdentifier);
Preconditions.checkArgument(sourceEdgeProperties != null, "Source edge managers or RootInputSpecs must be set when not changing parallelism");
} else {
LOG.info("Resetting vertex location hints due to change in parallelism for vertex: " + logIdentifier);
vertexLocationHint = null;
if (parallelism > numTasks) {
addTasks((parallelism));
} else if (parallelism < numTasks) {
removeTasks(parallelism);
}
}
Preconditions.checkState(this.numTasks == parallelism, getLogIdentifier());
// set new vertex location hints
setVertexLocationHint(vertexLocationHint);
LOG.info("Vertex " + getLogIdentifier() + " parallelism set to " + parallelism + " from " + oldNumTasks);
// notify listeners
stateChangeNotifier.stateChanged(vertexId, new VertexStateUpdateParallelismUpdated(vertexName, numTasks, oldNumTasks));
assert tasks.size() == numTasks;
// set new edge managers
if (sourceEdgeProperties != null) {
for (Map.Entry<String, EdgeProperty> entry : sourceEdgeProperties.entrySet()) {
LOG.info("Replacing edge manager for source:" + entry.getKey() + " destination: " + getLogIdentifier());
Vertex sourceVertex = appContext.getCurrentDAG().getVertex(entry.getKey());
Edge edge = sourceVertices.get(sourceVertex);
try {
edge.setEdgeProperty(entry.getValue());
} catch (Exception e) {
throw new TezUncheckedException(e);
}
}
}
// stop buffering events
for (Edge edge : sourceVertices.values()) {
edge.stopEventBuffering();
}
}
} finally {
writeLock.unlock();
}
}
use of org.apache.tez.dag.app.dag.Vertex in project tez by apache.
the class VertexImpl method logVertexConfigurationDoneEvent.
void logVertexConfigurationDoneEvent() {
if (recoveryData == null || !recoveryData.shouldSkipInit()) {
Map<String, EdgeProperty> sourceEdgeProperties = new HashMap<String, EdgeProperty>();
for (Map.Entry<Vertex, Edge> entry : this.sourceVertices.entrySet()) {
sourceEdgeProperties.put(entry.getKey().getName(), entry.getValue().getEdgeProperty());
}
VertexConfigurationDoneEvent reconfigureDoneEvent = new VertexConfigurationDoneEvent(vertexId, clock.getTime(), numTasks, taskLocationHints == null ? null : VertexLocationHint.create(Lists.newArrayList(taskLocationHints)), sourceEdgeProperties, rootInputSpecs, setParallelismCalledFlag);
this.appContext.getHistoryHandler().handle(new DAGHistoryEvent(getDAGId(), reconfigureDoneEvent));
}
}
use of org.apache.tez.dag.app.dag.Vertex in project tez by apache.
the class VertexImpl method getInputSpecList.
@Override
public List<InputSpec> getInputSpecList(int taskIndex) throws AMUserCodeException {
// For locking strategy, please refer to getOutputSpecList()
readLock.lock();
List<InputSpec> inputSpecList = null;
try {
inputSpecList = new ArrayList<InputSpec>(this.getInputVerticesCount() + (rootInputDescriptors == null ? 0 : rootInputDescriptors.size()));
if (rootInputDescriptors != null) {
for (Entry<String, RootInputLeafOutput<InputDescriptor, InputInitializerDescriptor>> rootInputDescriptorEntry : rootInputDescriptors.entrySet()) {
inputSpecList.add(new InputSpec(rootInputDescriptorEntry.getKey(), rootInputDescriptorEntry.getValue().getIODescriptor(), rootInputSpecs.get(rootInputDescriptorEntry.getKey()).getNumPhysicalInputsForWorkUnit(taskIndex)));
}
}
} finally {
readLock.unlock();
}
for (Vertex vertex : getInputVertices().keySet()) {
/**
* It is possible that setParallelism is in the middle of processing in target vertex with
* its write lock. So we need to get inputspec by acquiring read lock in target vertex to
* get consistent view.
* Refer TEZ-2251
*/
InputSpec inputSpec = ((VertexImpl) vertex).getDestinationSpecFor(this, taskIndex);
// TODO DAGAM This should be based on the edge type.
inputSpecList.add(inputSpec);
}
return inputSpecList;
}
use of org.apache.tez.dag.app.dag.Vertex in project tez by apache.
the class VertexImpl method getOutputSpecList.
@Override
public List<OutputSpec> getOutputSpecList(int taskIndex) throws AMUserCodeException {
/**
* Ref: TEZ-3297
* Locking entire method could introduce a nested lock and
* could lead to deadlock in corner cases. Example of deadlock with nested lock here:
* 1. In thread#1, Downstream vertex is in the middle of processing setParallelism and gets
* writeLock.
* 2. In thread#2, currentVertex acquires read lock
* 3. In thread#3, central dispatcher tries to process an event for current vertex,
* so tries to acquire write lock.
*
* In further processing,
* 4. In thread#1, it tries to acquire readLock on current vertex for setting edges. But
* this would be blocked as #3 already requested for write lock
* 5. In thread#2, getting readLock on downstream vertex would be blocked as writeLock
* is held by thread#1.
* 6. thread#3 is anyways blocked due to thread#2's read lock on current vertex.
*/
List<OutputSpec> outputSpecList = null;
readLock.lock();
try {
outputSpecList = new ArrayList<OutputSpec>(this.getOutputVerticesCount() + this.additionalOutputSpecs.size());
outputSpecList.addAll(additionalOutputSpecs);
} finally {
readLock.unlock();
}
for (Vertex vertex : targetVertices.keySet()) {
/**
* It is possible that setParallelism (which could change numTasks) is in the middle of
* processing in target vertex with its write lock. So we need to get outputspec by
* acquiring read lock in target vertex to get consistent view.
* Refer TEZ-2251
*/
OutputSpec outputSpec = ((VertexImpl) vertex).getSourceSpecFor(this, taskIndex);
outputSpecList.add(outputSpec);
}
return outputSpecList;
}
Aggregations