@Override
public void releaseSlot() {
// try to transition to the CANCELED state. That state marks
// that the releasing is in progress
if (markCancelled()) {
// kill all tasks currently running in this slot
Execution exec = this.executedTask;
if (exec != null && !exec.isFinished()) {
exec.fail(
new Exception(
"The slot in which the task was executed has been released. Probably loss of TaskManager "
+ getInstance()));
}
// release directly (if we are directly allocated),
// otherwise release through the parent shared slot
if (getParent() == null) {
// we have to give back the slot to the owning instance
getInstance().returnAllocatedSlot(this);
} else {
// we have to ask our parent to dispose us
getParent().releaseChild(this);
}
}
}
@Test
public void testRegistrationOfExecutionsFailingFinalize() {
try {
final JobVertexID jid1 = new JobVertexID();
final JobVertexID jid2 = new JobVertexID();
JobVertex v1 = new FailingFinalizeJobVertex("v1", jid1);
JobVertex v2 = new JobVertex("v2", jid2);
Map<ExecutionAttemptID, Execution> executions = setupExecution(v1, 6, v2, 4);
List<Execution> execList = new ArrayList<Execution>();
execList.addAll(executions.values());
// sort executions by job vertex. Failing job vertex first
Collections.sort(
execList,
new Comparator<Execution>() {
@Override
public int compare(Execution o1, Execution o2) {
return o1.getVertex().getSimpleName().compareTo(o2.getVertex().getSimpleName());
}
});
int cnt = 0;
for (Execution e : execList) {
cnt++;
e.markFinished();
if (cnt <= 6) {
// the last execution of the first job vertex triggers the failing finalize hook
assertEquals(ExecutionState.FINISHED, e.getState());
} else {
// all following executions should be canceled
assertEquals(ExecutionState.CANCELED, e.getState());
}
}
assertEquals(0, executions.size());
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
@Test
public void testRegistrationOfExecutionsFinishing() {
try {
final JobVertexID jid1 = new JobVertexID();
final JobVertexID jid2 = new JobVertexID();
JobVertex v1 = new JobVertex("v1", jid1);
JobVertex v2 = new JobVertex("v2", jid2);
Map<ExecutionAttemptID, Execution> executions = setupExecution(v1, 7650, v2, 2350);
for (Execution e : executions.values()) {
e.markFinished();
}
assertEquals(0, executions.size());
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
@Test
public void testRegistrationOfExecutionsFailedExternally() {
try {
final JobVertexID jid1 = new JobVertexID();
final JobVertexID jid2 = new JobVertexID();
JobVertex v1 = new JobVertex("v1", jid1);
JobVertex v2 = new JobVertex("v2", jid2);
Map<ExecutionAttemptID, Execution> executions = setupExecution(v1, 7, v2, 6);
for (Execution e : executions.values()) {
e.fail(null);
}
assertEquals(0, executions.size());
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
@Test
public void testRegistrationOfExecutionsCanceled() {
try {
final JobVertexID jid1 = new JobVertexID();
final JobVertexID jid2 = new JobVertexID();
JobVertex v1 = new JobVertex("v1", jid1);
JobVertex v2 = new JobVertex("v2", jid2);
Map<ExecutionAttemptID, Execution> executions = setupExecution(v1, 19, v2, 37);
for (Execution e : executions.values()) {
e.cancel();
e.cancelingComplete();
}
assertEquals(0, executions.size());
} catch (Exception e) {
e.printStackTrace();
fail(e.getMessage());
}
}
void scheduleOrUpdateConsumers(List<List<ExecutionEdge>> allConsumers) {
final int numConsumers = allConsumers.size();
if (numConsumers > 1) {
fail(
new IllegalStateException(
"Currently, only a single consumer group per partition is supported."));
} else if (numConsumers == 0) {
return;
}
for (ExecutionEdge edge : allConsumers.get(0)) {
final ExecutionVertex consumerVertex = edge.getTarget();
final Execution consumer = consumerVertex.getCurrentExecutionAttempt();
final ExecutionState consumerState = consumer.getState();
final IntermediateResultPartition partition = edge.getSource();
// ----------------------------------------------------------------
// Consumer is created => try to deploy and cache input channel
// descriptors if there is a deployment race
// ----------------------------------------------------------------
if (consumerState == CREATED) {
final Execution partitionExecution = partition.getProducer().getCurrentExecutionAttempt();
consumerVertex.cachePartitionInfo(
PartialInputChannelDeploymentDescriptor.fromEdge(partition, partitionExecution));
// When deploying a consuming task, its task deployment descriptor will contain all
// deployment information available at the respective time. It is possible that some
// of the partitions to be consumed have not been created yet. These are updated
// runtime via the update messages.
//
// TODO The current approach may send many update messages even though the consuming
// task has already been deployed with all necessary information. We have to check
// whether this is a problem and fix it, if it is.
future(
new Callable<Boolean>() {
@Override
public Boolean call() throws Exception {
try {
consumerVertex.scheduleForExecution(
consumerVertex.getExecutionGraph().getScheduler(),
consumerVertex.getExecutionGraph().isQueuedSchedulingAllowed());
} catch (Throwable t) {
fail(
new IllegalStateException(
"Could not schedule consumer " + "vertex " + consumerVertex, t));
}
return true;
}
},
executionContext);
// double check to resolve race conditions
if (consumerVertex.getExecutionState() == RUNNING) {
consumerVertex.sendPartitionInfos();
}
}
// ----------------------------------------------------------------
// Consumer is running => send update message now
// ----------------------------------------------------------------
else {
if (consumerState == RUNNING) {
final SimpleSlot consumerSlot = consumer.getAssignedResource();
if (consumerSlot == null) {
// The consumer has been reset concurrently
continue;
}
final Instance consumerInstance = consumerSlot.getInstance();
final ResultPartitionID partitionId =
new ResultPartitionID(partition.getPartitionId(), attemptId);
final Instance partitionInstance =
partition.getProducer().getCurrentAssignedResource().getInstance();
final ResultPartitionLocation partitionLocation;
if (consumerInstance.equals(partitionInstance)) {
// Consuming task is deployed to the same instance as the partition => local
partitionLocation = ResultPartitionLocation.createLocal();
} else {
// Different instances => remote
final ConnectionID connectionId =
new ConnectionID(
partitionInstance.getInstanceConnectionInfo(),
partition.getIntermediateResult().getConnectionIndex());
partitionLocation = ResultPartitionLocation.createRemote(connectionId);
}
final InputChannelDeploymentDescriptor descriptor =
new InputChannelDeploymentDescriptor(partitionId, partitionLocation);
final UpdatePartitionInfo updateTaskMessage =
new UpdateTaskSinglePartitionInfo(
consumer.getAttemptId(), partition.getIntermediateResult().getId(), descriptor);
sendUpdatePartitionInfoRpcCall(consumerSlot, updateTaskMessage);
}
// ----------------------------------------------------------------
// Consumer is scheduled or deploying => cache input channel
// deployment descriptors and send update message later
// ----------------------------------------------------------------
else if (consumerState == SCHEDULED || consumerState == DEPLOYING) {
final Execution partitionExecution = partition.getProducer().getCurrentExecutionAttempt();
consumerVertex.cachePartitionInfo(
PartialInputChannelDeploymentDescriptor.fromEdge(partition, partitionExecution));
// double check to resolve race conditions
if (consumerVertex.getExecutionState() == RUNNING) {
consumerVertex.sendPartitionInfos();
}
}
}
}
}
public boolean restoreLatestCheckpointedState(
Map<JobVertexID, ExecutionJobVertex> tasks,
boolean errorIfNoCheckpoint,
boolean allOrNothingState)
throws Exception {
synchronized (lock) {
if (shutdown) {
throw new IllegalStateException("CheckpointCoordinator is shut down");
}
// Recover the checkpoints
completedCheckpointStore.recover();
// restore from the latest checkpoint
CompletedCheckpoint latest = completedCheckpointStore.getLatestCheckpoint();
if (latest == null) {
if (errorIfNoCheckpoint) {
throw new IllegalStateException("No completed checkpoint available");
} else {
return false;
}
}
long recoveryTimestamp = System.currentTimeMillis();
for (Map.Entry<JobVertexID, TaskState> taskGroupStateEntry :
latest.getTaskStates().entrySet()) {
TaskState taskState = taskGroupStateEntry.getValue();
ExecutionJobVertex executionJobVertex = tasks.get(taskGroupStateEntry.getKey());
if (executionJobVertex != null) {
// check that we only restore the state if the parallelism has not been changed
if (taskState.getParallelism() != executionJobVertex.getParallelism()) {
throw new RuntimeException(
"Cannot restore the latest checkpoint because "
+ "the parallelism changed. The operator"
+ executionJobVertex.getJobVertexId()
+ " has parallelism "
+ executionJobVertex.getParallelism()
+ " whereas the corresponding"
+ "state object has a parallelism of "
+ taskState.getParallelism());
}
int counter = 0;
List<Set<Integer>> keyGroupPartitions =
createKeyGroupPartitions(numberKeyGroups, executionJobVertex.getParallelism());
for (int i = 0; i < executionJobVertex.getParallelism(); i++) {
SubtaskState subtaskState = taskState.getState(i);
SerializedValue<StateHandle<?>> state = null;
if (subtaskState != null) {
// count the number of executions for which we set a state
counter++;
state = subtaskState.getState();
}
Map<Integer, SerializedValue<StateHandle<?>>> kvStateForTaskMap =
taskState.getUnwrappedKvStates(keyGroupPartitions.get(i));
Execution currentExecutionAttempt =
executionJobVertex.getTaskVertices()[i].getCurrentExecutionAttempt();
currentExecutionAttempt.setInitialState(state, kvStateForTaskMap, recoveryTimestamp);
}
if (allOrNothingState && counter > 0 && counter < executionJobVertex.getParallelism()) {
throw new IllegalStateException(
"The checkpoint contained state only for "
+ "a subset of tasks for vertex "
+ executionJobVertex);
}
} else {
throw new IllegalStateException(
"There is no execution job vertex for the job"
+ " vertex ID "
+ taskGroupStateEntry.getKey());
}
}
return true;
}
}
/**
* Receives an AcknowledgeCheckpoint message and returns whether the message was associated with a
* pending checkpoint.
*
* @param message Checkpoint ack from the task manager
* @return Flag indicating whether the ack'd checkpoint was associated with a pending checkpoint.
* @throws Exception If the checkpoint cannot be added to the completed checkpoint store.
*/
public boolean receiveAcknowledgeMessage(AcknowledgeCheckpoint message) throws Exception {
if (shutdown || message == null) {
return false;
}
if (!job.equals(message.getJob())) {
LOG.error("Received AcknowledgeCheckpoint message for wrong job: {}", message);
return false;
}
final long checkpointId = message.getCheckpointId();
CompletedCheckpoint completed = null;
PendingCheckpoint checkpoint;
// Flag indicating whether the ack message was for a known pending
// checkpoint.
boolean isPendingCheckpoint;
synchronized (lock) {
// we need to check inside the lock for being shutdown as well, otherwise we
// get races and invalid error log messages
if (shutdown) {
return false;
}
checkpoint = pendingCheckpoints.get(checkpointId);
if (checkpoint != null && !checkpoint.isDiscarded()) {
isPendingCheckpoint = true;
if (checkpoint.acknowledgeTask(
message.getTaskExecutionId(),
message.getState(),
message.getStateSize(),
null)) { // TODO: Give KV-state to the acknowledgeTask method
if (checkpoint.isFullyAcknowledged()) {
completed = checkpoint.toCompletedCheckpoint();
completedCheckpointStore.addCheckpoint(completed);
LOG.info(
"Completed checkpoint "
+ checkpointId
+ " (in "
+ completed.getDuration()
+ " ms)");
if (LOG.isDebugEnabled()) {
StringBuilder builder = new StringBuilder();
for (Map.Entry<JobVertexID, TaskState> entry : completed.getTaskStates().entrySet()) {
builder
.append("JobVertexID: ")
.append(entry.getKey())
.append(" {")
.append(entry.getValue())
.append("}");
}
LOG.debug(builder.toString());
}
pendingCheckpoints.remove(checkpointId);
rememberRecentCheckpointId(checkpointId);
dropSubsumedCheckpoints(completed.getTimestamp());
onFullyAcknowledgedCheckpoint(completed);
triggerQueuedRequests();
}
} else {
// checkpoint did not accept message
LOG.error(
"Received duplicate or invalid acknowledge message for checkpoint "
+ checkpointId
+ " , task "
+ message.getTaskExecutionId());
}
} else if (checkpoint != null) {
// this should not happen
throw new IllegalStateException(
"Received message for discarded but non-removed checkpoint " + checkpointId);
} else {
// message is for an unknown checkpoint, or comes too late (checkpoint disposed)
if (recentPendingCheckpoints.contains(checkpointId)) {
isPendingCheckpoint = true;
LOG.warn("Received late message for now expired checkpoint attempt " + checkpointId);
} else {
isPendingCheckpoint = false;
}
}
}
// send the confirmation messages to the necessary targets. we do this here
// to be outside the lock scope
if (completed != null) {
final long timestamp = completed.getTimestamp();
for (ExecutionVertex ev : tasksToCommitTo) {
Execution ee = ev.getCurrentExecutionAttempt();
if (ee != null) {
ExecutionAttemptID attemptId = ee.getAttemptId();
NotifyCheckpointComplete notifyMessage =
new NotifyCheckpointComplete(job, attemptId, checkpointId, timestamp);
ev.sendMessageToCurrentExecution(notifyMessage, ee.getAttemptId());
}
}
statsTracker.onCompletedCheckpoint(completed);
}
return isPendingCheckpoint;
}
/**
* Triggers a new checkpoint and uses the given timestamp as the checkpoint timestamp.
*
* @param timestamp The timestamp for the checkpoint.
* @param nextCheckpointId The checkpoint ID to use for this checkpoint or <code>-1</code> if the
* checkpoint ID counter should be queried.
*/
public boolean triggerCheckpoint(long timestamp, long nextCheckpointId) throws Exception {
// make some eager pre-checks
synchronized (lock) {
// abort if the coordinator has been shutdown in the meantime
if (shutdown) {
return false;
}
// sanity check: there should never be more than one trigger request queued
if (triggerRequestQueued) {
LOG.warn("Trying to trigger another checkpoint while one was queued already");
return false;
}
// if too many checkpoints are currently in progress, we need to mark that a request is queued
if (pendingCheckpoints.size() >= maxConcurrentCheckpointAttempts) {
triggerRequestQueued = true;
if (currentPeriodicTrigger != null) {
currentPeriodicTrigger.cancel();
currentPeriodicTrigger = null;
}
return false;
}
// make sure the minimum interval between checkpoints has passed
if (lastTriggeredCheckpoint + minPauseBetweenCheckpoints > timestamp) {
if (currentPeriodicTrigger != null) {
currentPeriodicTrigger.cancel();
currentPeriodicTrigger = null;
}
ScheduledTrigger trigger = new ScheduledTrigger();
timer.scheduleAtFixedRate(trigger, minPauseBetweenCheckpoints, baseInterval);
return false;
}
}
// first check if all tasks that we need to trigger are running.
// if not, abort the checkpoint
ExecutionAttemptID[] triggerIDs = new ExecutionAttemptID[tasksToTrigger.length];
for (int i = 0; i < tasksToTrigger.length; i++) {
Execution ee = tasksToTrigger[i].getCurrentExecutionAttempt();
if (ee != null && ee.getState() == ExecutionState.RUNNING) {
triggerIDs[i] = ee.getAttemptId();
} else {
LOG.info(
"Checkpoint triggering task {} is not being executed at the moment. Aborting checkpoint.",
tasksToTrigger[i].getSimpleName());
return false;
}
}
// next, check if all tasks that need to acknowledge the checkpoint are running.
// if not, abort the checkpoint
Map<ExecutionAttemptID, ExecutionVertex> ackTasks = new HashMap<>(tasksToWaitFor.length);
for (ExecutionVertex ev : tasksToWaitFor) {
Execution ee = ev.getCurrentExecutionAttempt();
if (ee != null) {
ackTasks.put(ee.getAttemptId(), ev);
} else {
LOG.info(
"Checkpoint acknowledging task {} is not being executed at the moment. Aborting checkpoint.",
ev.getSimpleName());
return false;
}
}
// we will actually trigger this checkpoint!
lastTriggeredCheckpoint = timestamp;
final long checkpointID;
if (nextCheckpointId < 0) {
try {
// this must happen outside the locked scope, because it communicates
// with external services (in HA mode) and may block for a while.
checkpointID = checkpointIdCounter.getAndIncrement();
} catch (Throwable t) {
int numUnsuccessful = ++numUnsuccessfulCheckpointsTriggers;
LOG.warn(
"Failed to trigger checkpoint ("
+ numUnsuccessful
+ " consecutive failed attempts so far)",
t);
return false;
}
} else {
checkpointID = nextCheckpointId;
}
LOG.info("Triggering checkpoint " + checkpointID + " @ " + timestamp);
final PendingCheckpoint checkpoint =
new PendingCheckpoint(job, checkpointID, timestamp, ackTasks);
// schedule the timer that will clean up the expired checkpoints
TimerTask canceller =
new TimerTask() {
@Override
public void run() {
try {
synchronized (lock) {
// only do the work if the checkpoint is not discarded anyways
// note that checkpoint completion discards the pending checkpoint object
if (!checkpoint.isDiscarded()) {
LOG.info("Checkpoint " + checkpointID + " expired before completing.");
checkpoint.discard(userClassLoader);
pendingCheckpoints.remove(checkpointID);
rememberRecentCheckpointId(checkpointID);
onCancelCheckpoint(checkpointID);
triggerQueuedRequests();
}
}
} catch (Throwable t) {
LOG.error("Exception while handling checkpoint timeout", t);
}
}
};
try {
// re-acquire the lock
synchronized (lock) {
// since we released the lock in the meantime, we need to re-check
// that the conditions still hold. this is clumsy, but it allows us to
// release the lock in the meantime while calls to external services are
// blocking progress, and still gives us early checks that skip work
// if no checkpoint can happen anyways
if (shutdown) {
return false;
} else if (triggerRequestQueued) {
LOG.warn("Trying to trigger another checkpoint while one was queued already");
return false;
} else if (pendingCheckpoints.size() >= maxConcurrentCheckpointAttempts) {
triggerRequestQueued = true;
if (currentPeriodicTrigger != null) {
currentPeriodicTrigger.cancel();
currentPeriodicTrigger = null;
}
return false;
}
pendingCheckpoints.put(checkpointID, checkpoint);
timer.schedule(canceller, checkpointTimeout);
}
// end of lock scope
// send the messages to the tasks that trigger their checkpoint
for (int i = 0; i < tasksToTrigger.length; i++) {
ExecutionAttemptID id = triggerIDs[i];
TriggerCheckpoint message = new TriggerCheckpoint(job, id, checkpointID, timestamp);
tasksToTrigger[i].sendMessageToCurrentExecution(message, id);
}
numUnsuccessfulCheckpointsTriggers = 0;
return true;
} catch (Throwable t) {
// guard the map against concurrent modifications
synchronized (lock) {
pendingCheckpoints.remove(checkpointID);
}
int numUnsuccessful = ++numUnsuccessfulCheckpointsTriggers;
LOG.warn(
"Failed to trigger checkpoint ("
+ numUnsuccessful
+ " consecutive failed attempts so far)",
t);
if (!checkpoint.isDiscarded()) {
checkpoint.discard(userClassLoader);
}
return false;
}
}
public void restoreLatestCheckpointedState(
Map<JobVertexID, ExecutionJobVertex> tasks,
boolean errorIfNoCheckpoint,
boolean allOrNothingState)
throws Exception {
synchronized (lock) {
if (shutdown) {
throw new IllegalStateException("CheckpointCoordinator is shut down");
}
// Recover the checkpoints
completedCheckpointStore.recover();
// restore from the latest checkpoint
CompletedCheckpoint latest = completedCheckpointStore.getLatestCheckpoint();
if (latest == null) {
if (errorIfNoCheckpoint) {
throw new IllegalStateException("No completed checkpoint available");
} else {
return;
}
}
long recoveryTimestamp = System.currentTimeMillis();
if (allOrNothingState) {
Map<ExecutionJobVertex, Integer> stateCounts = new HashMap<ExecutionJobVertex, Integer>();
for (StateForTask state : latest.getStates()) {
ExecutionJobVertex vertex = tasks.get(state.getOperatorId());
Execution exec =
vertex.getTaskVertices()[state.getSubtask()].getCurrentExecutionAttempt();
exec.setInitialState(state.getState(), recoveryTimestamp);
Integer count = stateCounts.get(vertex);
if (count != null) {
stateCounts.put(vertex, count + 1);
} else {
stateCounts.put(vertex, 1);
}
}
// validate that either all task vertices have state, or none
for (Map.Entry<ExecutionJobVertex, Integer> entry : stateCounts.entrySet()) {
ExecutionJobVertex vertex = entry.getKey();
if (entry.getValue() != vertex.getParallelism()) {
throw new IllegalStateException(
"The checkpoint contained state only for a subset of tasks for vertex " + vertex);
}
}
} else {
for (StateForTask state : latest.getStates()) {
ExecutionJobVertex vertex = tasks.get(state.getOperatorId());
Execution exec =
vertex.getTaskVertices()[state.getSubtask()].getCurrentExecutionAttempt();
exec.setInitialState(state.getState(), recoveryTimestamp);
}
}
}
}
