2 * Copyright (c) 2014 Cisco Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.cluster.datastore;
10 import static com.google.common.base.Verify.verify;
12 import akka.actor.ActorRef;
13 import akka.actor.ActorSelection;
14 import akka.actor.Cancellable;
15 import akka.actor.ExtendedActorSystem;
16 import akka.actor.Props;
17 import akka.actor.Status;
18 import akka.actor.Status.Failure;
19 import akka.serialization.JavaSerializer;
20 import akka.serialization.Serialization;
21 import com.google.common.annotations.VisibleForTesting;
22 import com.google.common.base.Optional;
23 import com.google.common.base.Preconditions;
24 import com.google.common.base.Ticker;
25 import com.google.common.base.Verify;
26 import com.google.common.collect.ImmutableList;
27 import com.google.common.collect.ImmutableMap;
28 import com.google.common.collect.Range;
29 import java.io.IOException;
30 import java.util.Arrays;
31 import java.util.Collection;
32 import java.util.Collections;
33 import java.util.HashMap;
35 import java.util.concurrent.TimeUnit;
36 import org.eclipse.jdt.annotation.NonNull;
37 import org.eclipse.jdt.annotation.Nullable;
38 import org.opendaylight.controller.cluster.access.ABIVersion;
39 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
40 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
41 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
42 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
43 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
44 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
45 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
46 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
47 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
48 import org.opendaylight.controller.cluster.access.concepts.Request;
49 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
50 import org.opendaylight.controller.cluster.access.concepts.RequestException;
51 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
52 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
53 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
54 import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
55 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
56 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
57 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
58 import org.opendaylight.controller.cluster.common.actor.Dispatchers;
59 import org.opendaylight.controller.cluster.common.actor.Dispatchers.DispatcherType;
60 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
61 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
62 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
63 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
64 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
65 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardDataTreeListenerInfoMXBeanImpl;
66 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
67 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
68 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
69 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
70 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
71 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
72 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
73 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
74 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
75 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
76 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
77 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
78 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
79 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
80 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
81 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
82 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
83 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
84 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
85 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
86 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
87 import org.opendaylight.controller.cluster.datastore.persisted.DisableTrackingPayload;
88 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
89 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
90 import org.opendaylight.controller.cluster.messaging.SliceOptions;
91 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
92 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
93 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
94 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
95 import org.opendaylight.controller.cluster.raft.RaftActor;
96 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
97 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
98 import org.opendaylight.controller.cluster.raft.RaftState;
99 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
100 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
101 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
102 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
103 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
104 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
105 import org.opendaylight.yangtools.concepts.Identifier;
106 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTree;
107 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
108 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
109 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
110 import org.opendaylight.yangtools.yang.model.api.SchemaContextProvider;
111 import scala.concurrent.duration.FiniteDuration;
114 * A Shard represents a portion of the logical data tree.
117 * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
119 public class Shard extends RaftActor {
122 static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
124 public String toString() {
125 return "txCommitTimeoutCheck";
130 static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
132 public String toString() {
133 return "getShardMBeanMessage";
137 static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
139 public String toString() {
140 return "resumeNextPendingTransaction";
144 // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
145 public static final String DEFAULT_NAME = "default";
147 private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
150 final ABIVersion[] values = ABIVersion.values();
151 final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
152 SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
155 // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
156 private static final int CLIENT_MAX_MESSAGES = 1000;
158 // The state of this Shard
159 private final ShardDataTree store;
161 /// The name of this shard
162 private final String name;
164 private final String shardName;
166 private final ShardStats shardMBean;
168 private final ShardDataTreeListenerInfoMXBeanImpl listenerInfoMXBean;
170 private DatastoreContext datastoreContext;
172 private final ShardCommitCoordinator commitCoordinator;
174 private long transactionCommitTimeout;
176 private Cancellable txCommitTimeoutCheckSchedule;
178 private final Optional<ActorRef> roleChangeNotifier;
180 private final MessageTracker appendEntriesReplyTracker;
182 private final ShardTransactionActorFactory transactionActorFactory;
184 private final ShardSnapshotCohort snapshotCohort;
186 private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
188 private ShardSnapshot restoreFromSnapshot;
190 private final ShardTransactionMessageRetrySupport messageRetrySupport;
192 private final FrontendMetadata frontendMetadata;
193 private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
194 private boolean paused;
196 private final MessageSlicer responseMessageSlicer;
197 private final Dispatchers dispatchers;
199 private final MessageAssembler requestMessageAssembler;
201 protected Shard(final AbstractBuilder<?, ?> builder) {
202 super(builder.getId().toString(), builder.getPeerAddresses(),
203 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
205 this.name = builder.getId().toString();
206 this.shardName = builder.getId().getShardName();
207 this.datastoreContext = builder.getDatastoreContext();
208 this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
209 this.frontendMetadata = new FrontendMetadata(name);
211 setPersistence(datastoreContext.isPersistent());
213 LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
215 ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
216 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
217 if (builder.getDataTree() != null) {
218 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
219 treeChangeListenerPublisher, name, frontendMetadata);
221 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
222 builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher, name, frontendMetadata);
225 shardMBean = ShardMBeanFactory.getShardStatsMBean(name, datastoreContext.getDataStoreMXBeanType(), this);
227 if (isMetricsCaptureEnabled()) {
228 getContext().become(new MeteringBehavior(this));
231 commitCoordinator = new ShardCommitCoordinator(store, LOG, this.name);
233 setTransactionCommitTimeout();
235 // create a notifier actor for each cluster member
236 roleChangeNotifier = createRoleChangeNotifier(name);
238 appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
239 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
241 dispatchers = new Dispatchers(context().system().dispatchers());
242 transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
243 dispatchers.getDispatcherPath(Dispatchers.DispatcherType.Transaction),
244 self(), getContext(), shardMBean, builder.getId().getShardName());
246 snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
249 messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
251 responseMessageSlicer = MessageSlicer.builder().logContext(this.name)
252 .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
253 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
254 .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
256 requestMessageAssembler = MessageAssembler.builder().logContext(this.name)
257 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
258 .assembledMessageCallback((message, sender) -> self().tell(message, sender))
259 .expireStateAfterInactivity(datastoreContext.getRequestTimeout(), TimeUnit.NANOSECONDS).build();
261 listenerInfoMXBean = new ShardDataTreeListenerInfoMXBeanImpl(name, datastoreContext.getDataStoreMXBeanType(),
263 listenerInfoMXBean.register();
266 private void setTransactionCommitTimeout() {
267 transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
268 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
271 private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
272 ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
273 RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
274 return Optional.of(shardRoleChangeNotifier);
278 public void postStop() {
279 LOG.info("Stopping Shard {}", persistenceId());
283 messageRetrySupport.close();
285 if (txCommitTimeoutCheckSchedule != null) {
286 txCommitTimeoutCheckSchedule.cancel();
289 commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
291 shardMBean.unregisterMBean();
292 listenerInfoMXBean.unregister();
296 protected void handleRecover(final Object message) {
297 LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
300 super.handleRecover(message);
301 if (LOG.isTraceEnabled()) {
302 appendEntriesReplyTracker.begin();
307 protected void handleNonRaftCommand(final Object message) {
308 try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
309 final Optional<Error> maybeError = context.error();
310 if (maybeError.isPresent()) {
311 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
315 store.resetTransactionBatch();
317 if (message instanceof RequestEnvelope) {
318 handleRequestEnvelope((RequestEnvelope)message);
319 } else if (MessageAssembler.isHandledMessage(message)) {
320 handleRequestAssemblerMessage(message);
321 } else if (message instanceof ConnectClientRequest) {
322 handleConnectClient((ConnectClientRequest)message);
323 } else if (CreateTransaction.isSerializedType(message)) {
324 handleCreateTransaction(message);
325 } else if (message instanceof BatchedModifications) {
326 handleBatchedModifications((BatchedModifications)message);
327 } else if (message instanceof ForwardedReadyTransaction) {
328 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
329 } else if (message instanceof ReadyLocalTransaction) {
330 handleReadyLocalTransaction((ReadyLocalTransaction)message);
331 } else if (CanCommitTransaction.isSerializedType(message)) {
332 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
333 } else if (CommitTransaction.isSerializedType(message)) {
334 handleCommitTransaction(CommitTransaction.fromSerializable(message));
335 } else if (AbortTransaction.isSerializedType(message)) {
336 handleAbortTransaction(AbortTransaction.fromSerializable(message));
337 } else if (CloseTransactionChain.isSerializedType(message)) {
338 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
339 } else if (message instanceof RegisterDataTreeChangeListener) {
340 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
341 } else if (message instanceof UpdateSchemaContext) {
342 updateSchemaContext((UpdateSchemaContext) message);
343 } else if (message instanceof PeerAddressResolved) {
344 PeerAddressResolved resolved = (PeerAddressResolved) message;
345 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
346 } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
347 commitTimeoutCheck();
348 } else if (message instanceof DatastoreContext) {
349 onDatastoreContext((DatastoreContext)message);
350 } else if (message instanceof RegisterRoleChangeListener) {
351 roleChangeNotifier.get().forward(message, context());
352 } else if (message instanceof FollowerInitialSyncUpStatus) {
353 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
354 context().parent().tell(message, self());
355 } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
356 sender().tell(getShardMBean(), self());
357 } else if (message instanceof GetShardDataTree) {
358 sender().tell(store.getDataTree(), self());
359 } else if (message instanceof ServerRemoved) {
360 context().parent().forward(message, context());
361 } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
362 messageRetrySupport.onTimerMessage(message);
363 } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
364 store.processCohortRegistryCommand(getSender(),
365 (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
366 } else if (message instanceof MakeLeaderLocal) {
368 } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
369 store.resumeNextPendingTransaction();
370 } else if (!responseMessageSlicer.handleMessage(message)) {
371 super.handleNonRaftCommand(message);
376 private void handleRequestAssemblerMessage(final Object message) {
377 dispatchers.getDispatcher(DispatcherType.Serialization).execute(() -> {
378 JavaSerializer.currentSystem().value_$eq((ExtendedActorSystem) context().system());
379 requestMessageAssembler.handleMessage(message, self());
383 @SuppressWarnings("checkstyle:IllegalCatch")
384 private void handleRequestEnvelope(final RequestEnvelope envelope) {
385 final long now = ticker().read();
387 final RequestSuccess<?, ?> success = handleRequest(envelope, now);
388 if (success != null) {
389 final long executionTimeNanos = ticker().read() - now;
390 if (success instanceof SliceableMessage) {
391 dispatchers.getDispatcher(DispatcherType.Serialization).execute(() ->
392 responseMessageSlicer.slice(SliceOptions.builder().identifier(success.getTarget())
393 .message(envelope.newSuccessEnvelope(success, executionTimeNanos))
394 .sendTo(envelope.getMessage().getReplyTo()).replyTo(self())
395 .onFailureCallback(t -> LOG.warn("Error slicing response {}", success, t)).build()));
397 envelope.sendSuccess(success, executionTimeNanos);
400 } catch (RequestException e) {
401 LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
402 envelope.sendFailure(e, ticker().read() - now);
403 } catch (Exception e) {
404 LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
405 envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
406 ticker().read() - now);
410 private void commitTimeoutCheck() {
411 store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
412 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
413 requestMessageAssembler.checkExpiredAssembledMessageState();
416 private Optional<Long> updateAccess(final SimpleShardDataTreeCohort cohort) {
417 final FrontendIdentifier frontend = cohort.getIdentifier().getHistoryId().getClientId().getFrontendId();
418 final LeaderFrontendState state = knownFrontends.get(frontend);
420 // Not tell-based protocol, do nothing
421 return Optional.absent();
424 if (isIsolatedLeader()) {
425 // We are isolated and no new request can come through until we emerge from it. We are still updating
426 // liveness of frontend when we see it attempting to communicate. Use the last access timer.
427 return Optional.of(state.getLastSeenTicks());
430 // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
431 return Optional.of(state.getLastConnectTicks());
434 private void disableTracking(final DisableTrackingPayload payload) {
435 final ClientIdentifier clientId = payload.getIdentifier();
436 LOG.debug("{}: disabling tracking of {}", persistenceId(), clientId);
437 frontendMetadata.disableTracking(clientId);
440 final FrontendIdentifier frontendId = clientId.getFrontendId();
441 final LeaderFrontendState frontend = knownFrontends.get(frontendId);
442 if (frontend != null) {
443 if (clientId.equals(frontend.getIdentifier())) {
444 if (!(frontend instanceof LeaderFrontendState.Disabled)) {
445 verify(knownFrontends.replace(frontendId, frontend,
446 new LeaderFrontendState.Disabled(persistenceId(), clientId, store)));
447 LOG.debug("{}: leader state for {} disabled", persistenceId(), clientId);
449 LOG.debug("{}: leader state {} is already disabled", persistenceId(), frontend);
452 LOG.debug("{}: leader state {} does not match {}", persistenceId(), frontend, clientId);
455 LOG.debug("{}: leader state for {} not found", persistenceId(), clientId);
456 knownFrontends.put(frontendId, new LeaderFrontendState.Disabled(persistenceId(), clientId,
462 private void onMakeLeaderLocal() {
463 LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
465 getSender().tell(new Status.Success(null), getSelf());
469 final ActorSelection leader = getLeader();
471 if (leader == null) {
472 // Leader is not present. The cluster is most likely trying to
473 // elect a leader and we should let that run its normal course
475 // TODO we can wait for the election to complete and retry the
476 // request. We can also let the caller retry by sending a flag
477 // in the response indicating the request is "reTryable".
478 getSender().tell(new Failure(
479 new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
480 + "Currently there is no leader for " + persistenceId())),
485 leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
488 // Acquire our frontend tracking handle and verify generation matches
489 private @Nullable LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
490 final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
491 if (existing != null) {
492 final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
498 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
499 throw new RetiredGenerationException(clientId.getGeneration(),
500 existing.getIdentifier().getGeneration());
503 LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
505 knownFrontends.remove(clientId.getFrontendId());
507 LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
513 private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
514 final LeaderFrontendState ret = findFrontend(clientId);
519 // TODO: a dedicated exception would be better, but this is technically true, too
520 throw new OutOfSequenceEnvelopeException(0);
523 private static @NonNull ABIVersion selectVersion(final ConnectClientRequest message) {
524 final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
525 for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
526 if (clientRange.contains(v)) {
531 throw new IllegalArgumentException(String.format(
532 "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
536 @SuppressWarnings("checkstyle:IllegalCatch")
537 private void handleConnectClient(final ConnectClientRequest message) {
539 final ClientIdentifier clientId = message.getTarget();
540 final LeaderFrontendState existing = findFrontend(clientId);
541 if (existing != null) {
545 if (!isLeader() || !isLeaderActive()) {
546 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
547 + "isLeadershipTransferInProgress: {}.",
548 persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
549 throw new NotLeaderException(getSelf());
552 final ABIVersion selectedVersion = selectVersion(message);
553 final LeaderFrontendState frontend;
554 if (existing == null) {
555 frontend = new LeaderFrontendState.Enabled(persistenceId(), clientId, store);
556 knownFrontends.put(clientId.getFrontendId(), frontend);
557 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
562 frontend.reconnect();
563 message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
564 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
565 ActorRef.noSender());
566 } catch (RequestException | RuntimeException e) {
567 message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
571 private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
572 throws RequestException {
573 // We are not the leader, hence we want to fail-fast.
574 if (!isLeader() || paused || !isLeaderActive()) {
575 LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
576 + "isLeadershipTransferInProgress: {}, paused: {}",
577 persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
578 throw new NotLeaderException(getSelf());
581 final Request<?, ?> request = envelope.getMessage();
582 if (request instanceof TransactionRequest) {
583 final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
584 final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
585 return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
586 } else if (request instanceof LocalHistoryRequest) {
587 final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
588 final ClientIdentifier clientId = lhReq.getTarget().getClientId();
589 return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
591 LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
592 throw new UnsupportedRequestException(request);
596 private boolean hasLeader() {
597 return getLeaderId() != null;
600 public int getPendingTxCommitQueueSize() {
601 return store.getQueueSize();
604 public int getCohortCacheSize() {
605 return commitCoordinator.getCohortCacheSize();
609 protected Optional<ActorRef> getRoleChangeNotifier() {
610 return roleChangeNotifier;
613 String getShardName() {
618 protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
619 final short leaderPayloadVersion) {
620 return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
621 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
624 protected void onDatastoreContext(final DatastoreContext context) {
625 datastoreContext = context;
627 setTransactionCommitTimeout();
629 setPersistence(datastoreContext.isPersistent());
631 updateConfigParams(datastoreContext.getShardRaftConfig());
634 // applyState() will be invoked once consensus is reached on the payload
635 void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
636 boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
637 if (canSkipPayload) {
638 applyState(self(), id, payload);
640 // We are faking the sender
641 persistData(self(), id, payload, batchHint);
645 private void handleCommitTransaction(final CommitTransaction commit) {
646 final TransactionIdentifier txId = commit.getTransactionId();
648 askProtocolEncountered(txId);
649 commitCoordinator.handleCommit(txId, getSender(), this);
651 ActorSelection leader = getLeader();
652 if (leader == null) {
653 messageRetrySupport.addMessageToRetry(commit, getSender(), "Could not commit transaction " + txId);
655 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
656 leader.forward(commit, getContext());
661 private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
662 final TransactionIdentifier txId = canCommit.getTransactionId();
663 LOG.debug("{}: Can committing transaction {}", persistenceId(), txId);
666 askProtocolEncountered(txId);
667 commitCoordinator.handleCanCommit(txId, getSender(), this);
669 ActorSelection leader = getLeader();
670 if (leader == null) {
671 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
672 "Could not canCommit transaction " + txId);
674 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
675 leader.forward(canCommit, getContext());
680 @SuppressWarnings("checkstyle:IllegalCatch")
681 protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
682 askProtocolEncountered(batched.getTransactionId());
685 commitCoordinator.handleBatchedModifications(batched, sender, this);
686 } catch (Exception e) {
687 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
688 batched.getTransactionId(), e);
689 sender.tell(new Failure(e), getSelf());
693 private void handleBatchedModifications(final BatchedModifications batched) {
694 // This message is sent to prepare the modifications transaction directly on the Shard as an
695 // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
696 // BatchedModifications message, the caller sets the ready flag in the message indicating
697 // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
698 // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
699 // ReadyTransaction message.
701 // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
702 // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
703 // the primary/leader shard. However with timing and caching on the front-end, there's a small
704 // window where it could have a stale leader during leadership transitions.
706 boolean isLeaderActive = isLeaderActive();
707 if (isLeader() && isLeaderActive) {
708 handleBatchedModificationsLocal(batched, getSender());
710 ActorSelection leader = getLeader();
711 if (!isLeaderActive || leader == null) {
712 messageRetrySupport.addMessageToRetry(batched, getSender(),
713 "Could not process BatchedModifications " + batched.getTransactionId());
715 // If this is not the first batch and leadership changed in between batched messages,
716 // we need to reconstruct previous BatchedModifications from the transaction
717 // DataTreeModification, honoring the max batched modification count, and forward all the
718 // previous BatchedModifications to the new leader.
719 Collection<BatchedModifications> newModifications = commitCoordinator
720 .createForwardedBatchedModifications(batched,
721 datastoreContext.getShardBatchedModificationCount());
723 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
724 newModifications.size(), leader);
726 for (BatchedModifications bm : newModifications) {
727 leader.forward(bm, getContext());
733 private boolean failIfIsolatedLeader(final ActorRef sender) {
734 if (isIsolatedLeader()) {
735 sender.tell(new Failure(new NoShardLeaderException(String.format(
736 "Shard %s was the leader but has lost contact with all of its followers. Either all"
737 + " other follower nodes are down or this node is isolated by a network partition.",
738 persistenceId()))), getSelf());
745 protected boolean isIsolatedLeader() {
746 return getRaftState() == RaftState.IsolatedLeader;
749 @SuppressWarnings("checkstyle:IllegalCatch")
750 private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
751 LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
753 boolean isLeaderActive = isLeaderActive();
754 if (isLeader() && isLeaderActive) {
756 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
757 } catch (Exception e) {
758 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
759 message.getTransactionId(), e);
760 getSender().tell(new Failure(e), getSelf());
763 ActorSelection leader = getLeader();
764 if (!isLeaderActive || leader == null) {
765 messageRetrySupport.addMessageToRetry(message, getSender(),
766 "Could not process ready local transaction " + message.getTransactionId());
768 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
769 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
770 leader.forward(message, getContext());
775 private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
776 LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
778 boolean isLeaderActive = isLeaderActive();
779 if (isLeader() && isLeaderActive) {
780 askProtocolEncountered(forwardedReady.getTransactionId());
781 commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
783 ActorSelection leader = getLeader();
784 if (!isLeaderActive || leader == null) {
785 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
786 "Could not process forwarded ready transaction " + forwardedReady.getTransactionId());
788 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
790 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
791 forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit(),
792 forwardedReady.getParticipatingShardNames());
793 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
794 leader.forward(readyLocal, getContext());
799 private void handleAbortTransaction(final AbortTransaction abort) {
800 final TransactionIdentifier transactionId = abort.getTransactionId();
801 askProtocolEncountered(transactionId);
802 doAbortTransaction(transactionId, getSender());
805 void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
806 commitCoordinator.handleAbort(transactionID, sender, this);
809 private void handleCreateTransaction(final Object message) {
811 createTransaction(CreateTransaction.fromSerializable(message));
812 } else if (getLeader() != null) {
813 getLeader().forward(message, getContext());
815 getSender().tell(new Failure(new NoShardLeaderException(
816 "Could not create a shard transaction", persistenceId())), getSelf());
820 private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
822 final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
823 askProtocolEncountered(id.getClientId());
824 store.closeTransactionChain(id);
825 } else if (getLeader() != null) {
826 getLeader().forward(closeTransactionChain, getContext());
828 LOG.warn("{}: Could not close transaction {}", persistenceId(), closeTransactionChain.getIdentifier());
832 @SuppressWarnings("checkstyle:IllegalCatch")
833 private void createTransaction(final CreateTransaction createTransaction) {
834 askProtocolEncountered(createTransaction.getTransactionId());
837 if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
838 && failIfIsolatedLeader(getSender())) {
842 ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
843 createTransaction.getTransactionId());
845 getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
846 createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
847 } catch (Exception e) {
848 getSender().tell(new Failure(e), getSelf());
852 private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
853 LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
854 return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
858 // Called on leader only
859 private void askProtocolEncountered(final TransactionIdentifier transactionId) {
860 askProtocolEncountered(transactionId.getHistoryId().getClientId());
863 // Called on leader only
864 private void askProtocolEncountered(final ClientIdentifier clientId) {
865 final FrontendIdentifier frontend = clientId.getFrontendId();
866 final LeaderFrontendState state = knownFrontends.get(frontend);
867 if (!(state instanceof LeaderFrontendState.Disabled)) {
868 LOG.debug("{}: encountered ask-based client {}, disabling transaction tracking", persistenceId(), clientId);
869 if (knownFrontends.isEmpty()) {
870 knownFrontends = new HashMap<>();
872 knownFrontends.put(frontend, new LeaderFrontendState.Disabled(persistenceId(), clientId, getDataStore()));
874 persistPayload(clientId, DisableTrackingPayload.create(clientId,
875 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
879 private void updateSchemaContext(final UpdateSchemaContext message) {
880 updateSchemaContext(message.getSchemaContext());
884 void updateSchemaContext(final SchemaContext schemaContext) {
885 store.updateSchemaContext(schemaContext);
888 private boolean isMetricsCaptureEnabled() {
889 CommonConfig config = new CommonConfig(getContext().system().settings().config());
890 return config.isMetricCaptureEnabled();
895 public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
896 return snapshotCohort;
900 protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
901 if (restoreFromSnapshot == null) {
902 return ShardRecoveryCoordinator.create(store, persistenceId(), LOG);
905 return ShardRecoveryCoordinator.forSnapshot(store, persistenceId(), LOG, restoreFromSnapshot.getSnapshot());
909 protected void onRecoveryComplete() {
910 restoreFromSnapshot = null;
912 //notify shard manager
913 getContext().parent().tell(new ActorInitialized(), getSelf());
915 // Being paranoid here - this method should only be called once but just in case...
916 if (txCommitTimeoutCheckSchedule == null) {
917 // Schedule a message to be periodically sent to check if the current in-progress
918 // transaction should be expired and aborted.
919 FiniteDuration period = FiniteDuration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
920 txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
921 period, period, getSelf(),
922 TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
927 protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
928 if (data instanceof Payload) {
929 if (data instanceof DisableTrackingPayload) {
930 disableTracking((DisableTrackingPayload) data);
935 store.applyReplicatedPayload(identifier, (Payload)data);
936 } catch (DataValidationFailedException | IOException e) {
937 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
940 LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
945 protected void onStateChanged() {
946 boolean isLeader = isLeader();
947 boolean hasLeader = hasLeader();
948 treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
950 // If this actor is no longer the leader close all the transaction chains
952 if (LOG.isDebugEnabled()) {
954 "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
955 persistenceId(), getId());
959 store.purgeLeaderState();
962 if (hasLeader && !isIsolatedLeader()) {
963 messageRetrySupport.retryMessages();
968 protected void onLeaderChanged(final String oldLeader, final String newLeader) {
969 shardMBean.incrementLeadershipChangeCount();
973 if (!knownFrontends.isEmpty()) {
974 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
975 knownFrontends = ImmutableMap.of();
978 requestMessageAssembler.close();
981 // No leader anywhere, nothing else to do
985 // Another leader was elected. If we were the previous leader and had pending transactions, convert
986 // them to transaction messages and send to the new leader.
987 ActorSelection leader = getLeader();
988 if (leader != null) {
989 Collection<?> messagesToForward = convertPendingTransactionsToMessages();
991 if (!messagesToForward.isEmpty()) {
992 LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
993 messagesToForward.size(), leader);
995 for (Object message : messagesToForward) {
996 LOG.debug("{}: Forwarding pending transaction message {}", persistenceId(), message);
998 leader.tell(message, self());
1002 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
1003 + "change and the leader address isn't available.", this);
1006 // We have become the leader, we need to reconstruct frontend state
1007 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
1008 LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
1011 if (!isIsolatedLeader()) {
1012 messageRetrySupport.retryMessages();
1017 * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
1019 * @return the converted messages
1021 public Collection<?> convertPendingTransactionsToMessages() {
1022 return commitCoordinator.convertPendingTransactionsToMessages(
1023 datastoreContext.getShardBatchedModificationCount());
1027 protected void pauseLeader(final Runnable operation) {
1028 LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
1031 // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
1032 if (datastoreContext.isUseTellBasedProtocol()) {
1033 knownFrontends.values().forEach(LeaderFrontendState::retire);
1034 knownFrontends = ImmutableMap.of();
1037 store.setRunOnPendingTransactionsComplete(operation);
1041 protected void unpauseLeader() {
1042 LOG.debug("{}: In unpauseLeader", persistenceId());
1045 store.setRunOnPendingTransactionsComplete(null);
1047 // Restore tell-based protocol state as if we were becoming the leader
1048 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
1052 protected OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
1053 return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
1054 .commitCohortActors(store.getCohortActors());
1058 public String persistenceId() {
1063 ShardCommitCoordinator getCommitCoordinator() {
1064 return commitCoordinator;
1067 public DatastoreContext getDatastoreContext() {
1068 return datastoreContext;
1072 public ShardDataTree getDataStore() {
1077 ShardStats getShardMBean() {
1081 public static Builder builder() {
1082 return new Builder();
1085 public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
1086 private final Class<S> shardClass;
1087 private ShardIdentifier id;
1088 private Map<String, String> peerAddresses = Collections.emptyMap();
1089 private DatastoreContext datastoreContext;
1090 private SchemaContextProvider schemaContextProvider;
1091 private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
1092 private DataTree dataTree;
1093 private volatile boolean sealed;
1095 protected AbstractBuilder(final Class<S> shardClass) {
1096 this.shardClass = shardClass;
1099 protected void checkSealed() {
1100 Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
1103 @SuppressWarnings("unchecked")
1108 public T id(final ShardIdentifier newId) {
1114 public T peerAddresses(final Map<String, String> newPeerAddresses) {
1116 this.peerAddresses = newPeerAddresses;
1120 public T datastoreContext(final DatastoreContext newDatastoreContext) {
1122 this.datastoreContext = newDatastoreContext;
1126 public T schemaContextProvider(final SchemaContextProvider newSchemaContextProvider) {
1128 this.schemaContextProvider = Preconditions.checkNotNull(newSchemaContextProvider);
1132 public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
1134 this.restoreFromSnapshot = newRestoreFromSnapshot;
1138 public T dataTree(final DataTree newDataTree) {
1140 this.dataTree = newDataTree;
1144 public ShardIdentifier getId() {
1148 public Map<String, String> getPeerAddresses() {
1149 return peerAddresses;
1152 public DatastoreContext getDatastoreContext() {
1153 return datastoreContext;
1156 public SchemaContext getSchemaContext() {
1157 return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
1160 public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1161 return restoreFromSnapshot;
1164 public DataTree getDataTree() {
1168 public TreeType getTreeType() {
1169 switch (datastoreContext.getLogicalStoreType()) {
1171 return TreeType.CONFIGURATION;
1173 return TreeType.OPERATIONAL;
1175 throw new IllegalStateException("Unhandled logical store type "
1176 + datastoreContext.getLogicalStoreType());
1180 protected void verify() {
1181 Preconditions.checkNotNull(id, "id should not be null");
1182 Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
1183 Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
1184 Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
1187 public Props props() {
1190 return Props.create(shardClass, this);
1194 public static class Builder extends AbstractBuilder<Builder, Shard> {
1201 return Ticker.systemTicker();
1204 void scheduleNextPendingTransaction() {
1205 self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());