use of org.apache.flink.runtime.state.PlaceholderStreamStateHandle in project flink by apache.
the class SchedulerUtilsTest method testSharedStateRegistration.
/**
* Check that a {@link SharedStateRegistryFactory} used by {@link SchedulerUtils} registers
* shared checkpoint state on restore.
*/
@Test
public void testSharedStateRegistration() throws Exception {
UUID backendId = UUID.randomUUID();
StateHandleID key = new StateHandleID("k0");
StreamStateHandle handle = new ByteStreamStateHandle("h0", new byte[] { 1, 2, 3 });
CheckpointRecoveryFactory recoveryFactory = buildRecoveryFactory(buildCheckpoint(buildIncrementalHandle(key, handle, backendId)));
CompletedCheckpointStore checkpointStore = SchedulerUtils.createCompletedCheckpointStore(new Configuration(), recoveryFactory, Executors.directExecutor(), log, new JobID());
SharedStateRegistry sharedStateRegistry = checkpointStore.getSharedStateRegistry();
IncrementalRemoteKeyedStateHandle newHandle = buildIncrementalHandle(key, new PlaceholderStreamStateHandle(1L), backendId);
newHandle.registerSharedStates(sharedStateRegistry, 1L);
assertSame(handle, newHandle.getSharedState().get(key));
}
use of org.apache.flink.runtime.state.PlaceholderStreamStateHandle in project flink by apache.
the class RocksIncrementalSnapshotStrategyTest method testCheckpointIsIncremental.
// Verify the next checkpoint is still incremental after a savepoint completed.
@Test
public void testCheckpointIsIncremental() throws Exception {
try (CloseableRegistry closeableRegistry = new CloseableRegistry();
RocksIncrementalSnapshotStrategy checkpointSnapshotStrategy = createSnapshotStrategy(closeableRegistry)) {
FsCheckpointStreamFactory checkpointStreamFactory = createFsCheckpointStreamFactory();
// make and notify checkpoint with id 1
snapshot(1L, checkpointSnapshotStrategy, checkpointStreamFactory, closeableRegistry);
checkpointSnapshotStrategy.notifyCheckpointComplete(1L);
// notify savepoint with id 2
checkpointSnapshotStrategy.notifyCheckpointComplete(2L);
// make checkpoint with id 3
IncrementalRemoteKeyedStateHandle incrementalRemoteKeyedStateHandle3 = snapshot(3L, checkpointSnapshotStrategy, checkpointStreamFactory, closeableRegistry);
// If 3rd checkpoint's placeholderStateHandleCount > 0,it means 3rd checkpoint is
// incremental.
Map<StateHandleID, StreamStateHandle> sharedState3 = incrementalRemoteKeyedStateHandle3.getSharedState();
long placeholderStateHandleCount = sharedState3.entrySet().stream().filter(e -> e.getValue() instanceof PlaceholderStreamStateHandle).count();
Assert.assertTrue(placeholderStateHandleCount > 0);
}
}
use of org.apache.flink.runtime.state.PlaceholderStreamStateHandle in project flink by apache.
the class CheckpointCoordinatorTest method testSharedStateRegistrationOnRestore.
@Test
public void testSharedStateRegistrationOnRestore() throws Exception {
JobVertexID jobVertexID1 = new JobVertexID();
int parallelism1 = 2;
int maxParallelism1 = 4;
ExecutionGraph graph = new CheckpointCoordinatorTestingUtils.CheckpointExecutionGraphBuilder().addJobVertex(jobVertexID1, parallelism1, maxParallelism1).build();
ExecutionJobVertex jobVertex1 = graph.getJobVertex(jobVertexID1);
List<CompletedCheckpoint> checkpoints = Collections.emptyList();
SharedStateRegistry firstInstance = SharedStateRegistry.DEFAULT_FACTORY.create(org.apache.flink.util.concurrent.Executors.directExecutor(), checkpoints);
final EmbeddedCompletedCheckpointStore store = new EmbeddedCompletedCheckpointStore(10, checkpoints, firstInstance);
// set up the coordinator and validate the initial state
final CheckpointCoordinatorBuilder coordinatorBuilder = new CheckpointCoordinatorBuilder().setExecutionGraph(graph).setTimer(manuallyTriggeredScheduledExecutor);
final CheckpointCoordinator coordinator = coordinatorBuilder.setCompletedCheckpointStore(store).build();
final int numCheckpoints = 3;
List<KeyGroupRange> keyGroupPartitions1 = StateAssignmentOperation.createKeyGroupPartitions(maxParallelism1, parallelism1);
for (int i = 0; i < numCheckpoints; ++i) {
performIncrementalCheckpoint(graph.getJobID(), coordinator, jobVertex1, keyGroupPartitions1, i);
}
List<CompletedCheckpoint> completedCheckpoints = coordinator.getSuccessfulCheckpoints();
assertEquals(numCheckpoints, completedCheckpoints.size());
int sharedHandleCount = 0;
List<Map<StateHandleID, StreamStateHandle>> sharedHandlesByCheckpoint = new ArrayList<>(numCheckpoints);
for (int i = 0; i < numCheckpoints; ++i) {
sharedHandlesByCheckpoint.add(new HashMap<>(2));
}
int cp = 0;
for (CompletedCheckpoint completedCheckpoint : completedCheckpoints) {
for (OperatorState taskState : completedCheckpoint.getOperatorStates().values()) {
for (OperatorSubtaskState subtaskState : taskState.getStates()) {
for (KeyedStateHandle keyedStateHandle : subtaskState.getManagedKeyedState()) {
// test we are once registered with the current registry
verify(keyedStateHandle, times(1)).registerSharedStates(firstInstance, completedCheckpoint.getCheckpointID());
IncrementalRemoteKeyedStateHandle incrementalKeyedStateHandle = (IncrementalRemoteKeyedStateHandle) keyedStateHandle;
sharedHandlesByCheckpoint.get(cp).putAll(incrementalKeyedStateHandle.getSharedState());
for (StreamStateHandle streamStateHandle : incrementalKeyedStateHandle.getSharedState().values()) {
assertTrue(!(streamStateHandle instanceof PlaceholderStreamStateHandle));
verify(streamStateHandle, never()).discardState();
++sharedHandleCount;
}
for (StreamStateHandle streamStateHandle : incrementalKeyedStateHandle.getPrivateState().values()) {
verify(streamStateHandle, never()).discardState();
}
verify(incrementalKeyedStateHandle.getMetaStateHandle(), never()).discardState();
}
verify(subtaskState, never()).discardState();
}
}
++cp;
}
// 2 (parallelism) x (1 (CP0) + 2 (CP1) + 2 (CP2)) = 10
assertEquals(10, sharedHandleCount);
// discard CP0
store.removeOldestCheckpoint();
// CP1
for (Map<StateHandleID, StreamStateHandle> cpList : sharedHandlesByCheckpoint) {
for (StreamStateHandle streamStateHandle : cpList.values()) {
verify(streamStateHandle, never()).discardState();
}
}
// shutdown the store
store.shutdown(JobStatus.SUSPENDED, new CheckpointsCleaner());
// restore the store
Set<ExecutionJobVertex> tasks = new HashSet<>();
tasks.add(jobVertex1);
assertEquals(JobStatus.SUSPENDED, store.getShutdownStatus().orElse(null));
SharedStateRegistry secondInstance = SharedStateRegistry.DEFAULT_FACTORY.create(org.apache.flink.util.concurrent.Executors.directExecutor(), store.getAllCheckpoints());
final EmbeddedCompletedCheckpointStore secondStore = new EmbeddedCompletedCheckpointStore(10, store.getAllCheckpoints(), secondInstance);
final CheckpointCoordinator secondCoordinator = coordinatorBuilder.setCompletedCheckpointStore(secondStore).build();
assertTrue(secondCoordinator.restoreLatestCheckpointedStateToAll(tasks, false));
// validate that all shared states are registered again after the recovery.
cp = 0;
for (CompletedCheckpoint completedCheckpoint : completedCheckpoints) {
for (OperatorState taskState : completedCheckpoint.getOperatorStates().values()) {
for (OperatorSubtaskState subtaskState : taskState.getStates()) {
for (KeyedStateHandle keyedStateHandle : subtaskState.getManagedKeyedState()) {
VerificationMode verificationMode;
// test we are once registered with the new registry
if (cp > 0) {
verificationMode = times(1);
} else {
verificationMode = never();
}
// check that all are registered with the new registry
verify(keyedStateHandle, verificationMode).registerSharedStates(secondInstance, completedCheckpoint.getCheckpointID());
}
}
}
++cp;
}
// discard CP1
secondStore.removeOldestCheckpoint();
// we expect that all shared state from CP0 is no longer referenced and discarded. CP2 is
// still live and also
// references the state from CP1, so we expect they are not discarded.
verifyDiscard(sharedHandlesByCheckpoint, cpId -> cpId == 0 ? times(1) : never());
// discard CP2
secondStore.removeOldestCheckpoint();
// still expect shared state not to be discarded because it may be used in later checkpoints
verifyDiscard(sharedHandlesByCheckpoint, cpId -> cpId == 1 ? never() : atLeast(0));
}
use of org.apache.flink.runtime.state.PlaceholderStreamStateHandle in project flink by apache.
the class CheckpointCoordinatorTest method performIncrementalCheckpoint.
private void performIncrementalCheckpoint(JobID jobId, CheckpointCoordinator checkpointCoordinator, ExecutionJobVertex jobVertex1, List<KeyGroupRange> keyGroupPartitions1, int cpSequenceNumber) throws Exception {
// trigger the checkpoint
checkpointCoordinator.triggerCheckpoint(false);
manuallyTriggeredScheduledExecutor.triggerAll();
assertEquals(1, checkpointCoordinator.getPendingCheckpoints().size());
long checkpointId = Iterables.getOnlyElement(checkpointCoordinator.getPendingCheckpoints().keySet());
for (int index = 0; index < jobVertex1.getParallelism(); index++) {
KeyGroupRange keyGroupRange = keyGroupPartitions1.get(index);
Map<StateHandleID, StreamStateHandle> privateState = new HashMap<>();
privateState.put(new StateHandleID("private-1"), spy(new ByteStreamStateHandle("private-1", new byte[] { 'p' })));
Map<StateHandleID, StreamStateHandle> sharedState = new HashMap<>();
// let all but the first CP overlap by one shared state.
if (cpSequenceNumber > 0) {
sharedState.put(new StateHandleID("shared-" + (cpSequenceNumber - 1)), spy(new PlaceholderStreamStateHandle(1L)));
}
sharedState.put(new StateHandleID("shared-" + cpSequenceNumber), spy(new ByteStreamStateHandle("shared-" + cpSequenceNumber + "-" + keyGroupRange, new byte[] { 's' })));
IncrementalRemoteKeyedStateHandle managedState = spy(new IncrementalRemoteKeyedStateHandle(new UUID(42L, 42L), keyGroupRange, checkpointId, sharedState, privateState, spy(new ByteStreamStateHandle("meta", new byte[] { 'm' }))));
OperatorSubtaskState operatorSubtaskState = spy(OperatorSubtaskState.builder().setManagedKeyedState(managedState).build());
Map<OperatorID, OperatorSubtaskState> opStates = new HashMap<>();
opStates.put(jobVertex1.getOperatorIDs().get(0).getGeneratedOperatorID(), operatorSubtaskState);
TaskStateSnapshot taskStateSnapshot = new TaskStateSnapshot(opStates);
AcknowledgeCheckpoint acknowledgeCheckpoint = new AcknowledgeCheckpoint(jobId, jobVertex1.getTaskVertices()[index].getCurrentExecutionAttempt().getAttemptId(), checkpointId, new CheckpointMetrics(), taskStateSnapshot);
checkpointCoordinator.receiveAcknowledgeMessage(acknowledgeCheckpoint, TASK_MANAGER_LOCATION_INFO);
}
}
Aggregations