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
9 package org.opendaylight.controller.cluster.datastore;
11 import akka.actor.ActorRef;
12 import akka.actor.ActorSelection;
13 import akka.actor.Cancellable;
14 import akka.actor.Props;
15 import akka.actor.Status;
16 import akka.actor.Status.Failure;
17 import akka.serialization.Serialization;
18 import com.google.common.annotations.VisibleForTesting;
19 import com.google.common.base.Optional;
20 import com.google.common.base.Preconditions;
21 import com.google.common.base.Ticker;
22 import com.google.common.base.Verify;
23 import com.google.common.collect.ImmutableList;
24 import com.google.common.collect.ImmutableMap;
25 import com.google.common.collect.Range;
26 import java.io.IOException;
27 import java.util.Arrays;
28 import java.util.Collection;
29 import java.util.Collections;
31 import java.util.concurrent.TimeUnit;
32 import javax.annotation.Nonnull;
33 import javax.annotation.Nullable;
34 import org.opendaylight.controller.cluster.access.ABIVersion;
35 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
36 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
37 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
38 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
39 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
40 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
41 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
42 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
43 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
44 import org.opendaylight.controller.cluster.access.concepts.Request;
45 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
46 import org.opendaylight.controller.cluster.access.concepts.RequestException;
47 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
48 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
49 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
50 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
51 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
52 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
53 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
54 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
55 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
56 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
57 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
58 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
59 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
60 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
61 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
62 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
63 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
64 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
65 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
66 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
67 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
68 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
69 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
70 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
71 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
72 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
73 import org.opendaylight.controller.cluster.datastore.messages.PersistAbortTransactionPayload;
74 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
75 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
76 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
77 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
78 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
79 import org.opendaylight.controller.cluster.datastore.persisted.AbortTransactionPayload;
80 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
81 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
82 import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
83 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
84 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
85 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
86 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
87 import org.opendaylight.controller.cluster.raft.RaftActor;
88 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
89 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
90 import org.opendaylight.controller.cluster.raft.RaftState;
91 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
92 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
93 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
94 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
95 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
96 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
97 import org.opendaylight.yangtools.concepts.Identifier;
98 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
99 import org.opendaylight.yangtools.yang.data.api.schema.tree.TipProducingDataTree;
100 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
101 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
102 import org.opendaylight.yangtools.yang.model.api.SchemaContextProvider;
103 import scala.concurrent.duration.Duration;
104 import scala.concurrent.duration.FiniteDuration;
107 * A Shard represents a portion of the logical data tree.
110 * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
112 public class Shard extends RaftActor {
115 static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
117 public String toString() {
118 return "txCommitTimeoutCheck";
123 static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
125 public String toString() {
126 return "getShardMBeanMessage";
130 static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
132 public String toString() {
133 return "resumeNextPendingTransaction";
137 // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
138 public static final String DEFAULT_NAME = "default";
140 private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
143 final ABIVersion[] values = ABIVersion.values();
144 final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
145 SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
148 // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
149 private static final int CLIENT_MAX_MESSAGES = 1000;
151 // The state of this Shard
152 private final ShardDataTree store;
154 /// The name of this shard
155 private final String name;
157 private final ShardStats shardMBean;
159 private DatastoreContext datastoreContext;
161 private final ShardCommitCoordinator commitCoordinator;
163 private long transactionCommitTimeout;
165 private Cancellable txCommitTimeoutCheckSchedule;
167 private final Optional<ActorRef> roleChangeNotifier;
169 private final MessageTracker appendEntriesReplyTracker;
171 private final ShardTransactionActorFactory transactionActorFactory;
173 private final ShardSnapshotCohort snapshotCohort;
175 private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
176 private final DataChangeListenerSupport changeSupport = new DataChangeListenerSupport(this);
179 private ShardSnapshot restoreFromSnapshot;
181 private final ShardTransactionMessageRetrySupport messageRetrySupport;
183 private final FrontendMetadata frontendMetadata;
184 private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
185 private boolean paused;
187 protected Shard(final AbstractBuilder<?, ?> builder) {
188 super(builder.getId().toString(), builder.getPeerAddresses(),
189 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
191 this.name = builder.getId().toString();
192 this.datastoreContext = builder.getDatastoreContext();
193 this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
194 this.frontendMetadata = new FrontendMetadata(name);
196 setPersistence(datastoreContext.isPersistent());
198 LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
200 ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
201 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
202 ShardDataChangeListenerPublisherActorProxy dataChangeListenerPublisher =
203 new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher", name);
204 if (builder.getDataTree() != null) {
205 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
206 treeChangeListenerPublisher, dataChangeListenerPublisher, name, frontendMetadata);
208 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
209 builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher,
210 dataChangeListenerPublisher, name, frontendMetadata);
213 shardMBean = ShardMBeanFactory.getShardStatsMBean(name, datastoreContext.getDataStoreMXBeanType(), this);
215 if (isMetricsCaptureEnabled()) {
216 getContext().become(new MeteringBehavior(this));
219 commitCoordinator = new ShardCommitCoordinator(store, LOG, this.name);
221 setTransactionCommitTimeout();
223 // create a notifier actor for each cluster member
224 roleChangeNotifier = createRoleChangeNotifier(name);
226 appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
227 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
229 transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
230 new Dispatchers(context().system().dispatchers()).getDispatcherPath(Dispatchers.DispatcherType.Transaction),
231 self(), getContext(), shardMBean, builder.getId().getShardName());
233 snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
236 messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
239 private void setTransactionCommitTimeout() {
240 transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
241 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
244 private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
245 ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
246 RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
247 return Optional.of(shardRoleChangeNotifier);
251 public void postStop() {
252 LOG.info("Stopping Shard {}", persistenceId());
256 messageRetrySupport.close();
258 if (txCommitTimeoutCheckSchedule != null) {
259 txCommitTimeoutCheckSchedule.cancel();
262 commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
264 shardMBean.unregisterMBean();
268 protected void handleRecover(final Object message) {
269 LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
272 super.handleRecover(message);
273 if (LOG.isTraceEnabled()) {
274 appendEntriesReplyTracker.begin();
278 @SuppressWarnings("checkstyle:IllegalCatch")
280 protected void handleNonRaftCommand(final Object message) {
281 try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
282 final Optional<Error> maybeError = context.error();
283 if (maybeError.isPresent()) {
284 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
288 store.resetTransactionBatch();
290 if (message instanceof RequestEnvelope) {
291 final long now = ticker().read();
292 final RequestEnvelope envelope = (RequestEnvelope)message;
295 final RequestSuccess<?, ?> success = handleRequest(envelope, now);
296 if (success != null) {
297 envelope.sendSuccess(success, ticker().read() - now);
299 } catch (RequestException e) {
300 LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
301 envelope.sendFailure(e, ticker().read() - now);
302 } catch (Exception e) {
303 LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
304 envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
305 ticker().read() - now);
307 } else if (message instanceof ConnectClientRequest) {
308 handleConnectClient((ConnectClientRequest)message);
309 } else if (CreateTransaction.isSerializedType(message)) {
310 handleCreateTransaction(message);
311 } else if (message instanceof BatchedModifications) {
312 handleBatchedModifications((BatchedModifications)message);
313 } else if (message instanceof ForwardedReadyTransaction) {
314 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
315 } else if (message instanceof ReadyLocalTransaction) {
316 handleReadyLocalTransaction((ReadyLocalTransaction)message);
317 } else if (CanCommitTransaction.isSerializedType(message)) {
318 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
319 } else if (CommitTransaction.isSerializedType(message)) {
320 handleCommitTransaction(CommitTransaction.fromSerializable(message));
321 } else if (AbortTransaction.isSerializedType(message)) {
322 handleAbortTransaction(AbortTransaction.fromSerializable(message));
323 } else if (CloseTransactionChain.isSerializedType(message)) {
324 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
325 } else if (message instanceof RegisterChangeListener) {
326 changeSupport.onMessage((RegisterChangeListener) message, isLeader(), hasLeader());
327 } else if (message instanceof RegisterDataTreeChangeListener) {
328 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
329 } else if (message instanceof UpdateSchemaContext) {
330 updateSchemaContext((UpdateSchemaContext) message);
331 } else if (message instanceof PeerAddressResolved) {
332 PeerAddressResolved resolved = (PeerAddressResolved) message;
333 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
334 } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
335 store.checkForExpiredTransactions(transactionCommitTimeout);
336 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
337 } else if (message instanceof DatastoreContext) {
338 onDatastoreContext((DatastoreContext)message);
339 } else if (message instanceof RegisterRoleChangeListener) {
340 roleChangeNotifier.get().forward(message, context());
341 } else if (message instanceof FollowerInitialSyncUpStatus) {
342 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
343 context().parent().tell(message, self());
344 } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
345 sender().tell(getShardMBean(), self());
346 } else if (message instanceof GetShardDataTree) {
347 sender().tell(store.getDataTree(), self());
348 } else if (message instanceof ServerRemoved) {
349 context().parent().forward(message, context());
350 } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
351 messageRetrySupport.onTimerMessage(message);
352 } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
353 store.processCohortRegistryCommand(getSender(),
354 (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
355 } else if (message instanceof PersistAbortTransactionPayload) {
356 final TransactionIdentifier txId = ((PersistAbortTransactionPayload) message).getTransactionId();
357 persistPayload(txId, AbortTransactionPayload.create(txId), true);
358 } else if (message instanceof MakeLeaderLocal) {
360 } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
361 store.resumeNextPendingTransaction();
363 super.handleNonRaftCommand(message);
368 private void onMakeLeaderLocal() {
369 LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
371 getSender().tell(new Status.Success(null), getSelf());
375 final ActorSelection leader = getLeader();
377 if (leader == null) {
378 // Leader is not present. The cluster is most likely trying to
379 // elect a leader and we should let that run its normal course
381 // TODO we can wait for the election to complete and retry the
382 // request. We can also let the caller retry by sending a flag
383 // in the response indicating the request is "reTryable".
384 getSender().tell(new Failure(
385 new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
386 + "Currently there is no leader for " + persistenceId())),
391 leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
394 // Acquire our frontend tracking handle and verify generation matches
396 private LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
397 final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
398 if (existing != null) {
399 final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
405 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
406 throw new RetiredGenerationException(existing.getIdentifier().getGeneration());
409 LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
411 knownFrontends.remove(clientId.getFrontendId());
413 LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
419 private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
420 final LeaderFrontendState ret = findFrontend(clientId);
425 // TODO: a dedicated exception would be better, but this is technically true, too
426 throw new OutOfSequenceEnvelopeException(0);
429 private static @Nonnull ABIVersion selectVersion(final ConnectClientRequest message) {
430 final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
431 for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
432 if (clientRange.contains(v)) {
437 throw new IllegalArgumentException(String.format(
438 "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
442 @SuppressWarnings("checkstyle:IllegalCatch")
443 private void handleConnectClient(final ConnectClientRequest message) {
445 final ClientIdentifier clientId = message.getTarget();
446 final LeaderFrontendState existing = findFrontend(clientId);
447 if (existing != null) {
451 if (!isLeader() || !isLeaderActive()) {
452 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
453 + "isLeadershipTransferInProgress: {}.",
454 persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
455 throw new NotLeaderException(getSelf());
458 final ABIVersion selectedVersion = selectVersion(message);
459 final LeaderFrontendState frontend;
460 if (existing == null) {
461 frontend = new LeaderFrontendState(persistenceId(), clientId, store);
462 knownFrontends.put(clientId.getFrontendId(), frontend);
463 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
468 frontend.reconnect();
469 message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
470 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
471 ActorRef.noSender());
472 } catch (RequestException | RuntimeException e) {
473 message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
477 private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
478 throws RequestException {
479 // We are not the leader, hence we want to fail-fast.
480 if (!isLeader() || paused || !isLeaderActive()) {
481 LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
482 + "isLeadershipTransferInProgress: {}, paused: {}",
483 persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
484 throw new NotLeaderException(getSelf());
487 final Request<?, ?> request = envelope.getMessage();
488 if (request instanceof TransactionRequest) {
489 final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
490 final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
491 return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
492 } else if (request instanceof LocalHistoryRequest) {
493 final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
494 final ClientIdentifier clientId = lhReq.getTarget().getClientId();
495 return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
497 LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
498 throw new UnsupportedRequestException(request);
502 private boolean hasLeader() {
503 return getLeaderId() != null;
506 public int getPendingTxCommitQueueSize() {
507 return store.getQueueSize();
510 public int getCohortCacheSize() {
511 return commitCoordinator.getCohortCacheSize();
515 protected Optional<ActorRef> getRoleChangeNotifier() {
516 return roleChangeNotifier;
520 protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
521 final short leaderPayloadVersion) {
522 return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
523 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
526 protected void onDatastoreContext(final DatastoreContext context) {
527 datastoreContext = context;
529 setTransactionCommitTimeout();
531 setPersistence(datastoreContext.isPersistent());
533 updateConfigParams(datastoreContext.getShardRaftConfig());
536 // applyState() will be invoked once consensus is reached on the payload
537 void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
538 boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
539 if (canSkipPayload) {
540 applyState(self(), id, payload);
542 // We are faking the sender
543 persistData(self(), id, payload, batchHint);
547 private void handleCommitTransaction(final CommitTransaction commit) {
549 commitCoordinator.handleCommit(commit.getTransactionId(), getSender(), this);
551 ActorSelection leader = getLeader();
552 if (leader == null) {
553 messageRetrySupport.addMessageToRetry(commit, getSender(),
554 "Could not commit transaction " + commit.getTransactionId());
556 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
557 leader.forward(commit, getContext());
562 private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
563 LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionId());
566 commitCoordinator.handleCanCommit(canCommit.getTransactionId(), getSender(), this);
568 ActorSelection leader = getLeader();
569 if (leader == null) {
570 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
571 "Could not canCommit transaction " + canCommit.getTransactionId());
573 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
574 leader.forward(canCommit, getContext());
579 @SuppressWarnings("checkstyle:IllegalCatch")
580 protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
582 commitCoordinator.handleBatchedModifications(batched, sender, this);
583 } catch (Exception e) {
584 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
585 batched.getTransactionId(), e);
586 sender.tell(new Failure(e), getSelf());
590 private void handleBatchedModifications(final BatchedModifications batched) {
591 // This message is sent to prepare the modifications transaction directly on the Shard as an
592 // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
593 // BatchedModifications message, the caller sets the ready flag in the message indicating
594 // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
595 // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
596 // ReadyTransaction message.
598 // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
599 // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
600 // the primary/leader shard. However with timing and caching on the front-end, there's a small
601 // window where it could have a stale leader during leadership transitions.
603 boolean isLeaderActive = isLeaderActive();
604 if (isLeader() && isLeaderActive) {
605 handleBatchedModificationsLocal(batched, getSender());
607 ActorSelection leader = getLeader();
608 if (!isLeaderActive || leader == null) {
609 messageRetrySupport.addMessageToRetry(batched, getSender(),
610 "Could not commit transaction " + batched.getTransactionId());
612 // If this is not the first batch and leadership changed in between batched messages,
613 // we need to reconstruct previous BatchedModifications from the transaction
614 // DataTreeModification, honoring the max batched modification count, and forward all the
615 // previous BatchedModifications to the new leader.
616 Collection<BatchedModifications> newModifications = commitCoordinator
617 .createForwardedBatchedModifications(batched,
618 datastoreContext.getShardBatchedModificationCount());
620 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
621 newModifications.size(), leader);
623 for (BatchedModifications bm : newModifications) {
624 leader.forward(bm, getContext());
630 private boolean failIfIsolatedLeader(final ActorRef sender) {
631 if (isIsolatedLeader()) {
632 sender.tell(new Failure(new NoShardLeaderException(String.format(
633 "Shard %s was the leader but has lost contact with all of its followers. Either all"
634 + " other follower nodes are down or this node is isolated by a network partition.",
635 persistenceId()))), getSelf());
642 protected boolean isIsolatedLeader() {
643 return getRaftState() == RaftState.IsolatedLeader;
646 @SuppressWarnings("checkstyle:IllegalCatch")
647 private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
648 LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
650 boolean isLeaderActive = isLeaderActive();
651 if (isLeader() && isLeaderActive) {
653 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
654 } catch (Exception e) {
655 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
656 message.getTransactionId(), e);
657 getSender().tell(new Failure(e), getSelf());
660 ActorSelection leader = getLeader();
661 if (!isLeaderActive || leader == null) {
662 messageRetrySupport.addMessageToRetry(message, getSender(),
663 "Could not commit transaction " + message.getTransactionId());
665 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
666 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
667 leader.forward(message, getContext());
672 private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
673 LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
675 boolean isLeaderActive = isLeaderActive();
676 if (isLeader() && isLeaderActive) {
677 commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
679 ActorSelection leader = getLeader();
680 if (!isLeaderActive || leader == null) {
681 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
682 "Could not commit transaction " + forwardedReady.getTransactionId());
684 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
686 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
687 forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
688 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
689 leader.forward(readyLocal, getContext());
694 private void handleAbortTransaction(final AbortTransaction abort) {
695 doAbortTransaction(abort.getTransactionId(), getSender());
698 void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
699 commitCoordinator.handleAbort(transactionID, sender, this);
702 private void handleCreateTransaction(final Object message) {
704 createTransaction(CreateTransaction.fromSerializable(message));
705 } else if (getLeader() != null) {
706 getLeader().forward(message, getContext());
708 getSender().tell(new Failure(new NoShardLeaderException(
709 "Could not create a shard transaction", persistenceId())), getSelf());
713 private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
714 final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
715 store.closeTransactionChain(id, null);
716 store.purgeTransactionChain(id, null);
719 @SuppressWarnings("checkstyle:IllegalCatch")
720 private void createTransaction(final CreateTransaction createTransaction) {
722 if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
723 && failIfIsolatedLeader(getSender())) {
727 ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
728 createTransaction.getTransactionId());
730 getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
731 createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
732 } catch (Exception e) {
733 getSender().tell(new Failure(e), getSelf());
737 private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
738 LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
739 return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
743 private void updateSchemaContext(final UpdateSchemaContext message) {
744 updateSchemaContext(message.getSchemaContext());
748 void updateSchemaContext(final SchemaContext schemaContext) {
749 store.updateSchemaContext(schemaContext);
752 private boolean isMetricsCaptureEnabled() {
753 CommonConfig config = new CommonConfig(getContext().system().settings().config());
754 return config.isMetricCaptureEnabled();
759 public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
760 return snapshotCohort;
765 protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
766 return new ShardRecoveryCoordinator(store,
767 restoreFromSnapshot != null ? restoreFromSnapshot.getSnapshot() : null, persistenceId(), LOG);
771 protected void onRecoveryComplete() {
772 restoreFromSnapshot = null;
774 //notify shard manager
775 getContext().parent().tell(new ActorInitialized(), getSelf());
777 // Being paranoid here - this method should only be called once but just in case...
778 if (txCommitTimeoutCheckSchedule == null) {
779 // Schedule a message to be periodically sent to check if the current in-progress
780 // transaction should be expired and aborted.
781 FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
782 txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
783 period, period, getSelf(),
784 TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
789 protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
790 if (data instanceof Payload) {
792 store.applyReplicatedPayload(identifier, (Payload)data);
793 } catch (DataValidationFailedException | IOException e) {
794 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
797 LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
802 protected void onStateChanged() {
803 boolean isLeader = isLeader();
804 boolean hasLeader = hasLeader();
805 changeSupport.onLeadershipChange(isLeader, hasLeader);
806 treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
808 // If this actor is no longer the leader close all the transaction chains
810 if (LOG.isDebugEnabled()) {
812 "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
813 persistenceId(), getId());
817 store.purgeLeaderState();
820 if (hasLeader && !isIsolatedLeader()) {
821 messageRetrySupport.retryMessages();
826 protected void onLeaderChanged(final String oldLeader, final String newLeader) {
827 shardMBean.incrementLeadershipChangeCount();
831 if (!knownFrontends.isEmpty()) {
832 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
833 knownFrontends = ImmutableMap.of();
837 // No leader anywhere, nothing else to do
841 // Another leader was elected. If we were the previous leader and had pending transactions, convert
842 // them to transaction messages and send to the new leader.
843 ActorSelection leader = getLeader();
844 if (leader != null) {
845 Collection<?> messagesToForward = convertPendingTransactionsToMessages();
847 if (!messagesToForward.isEmpty()) {
848 LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
849 messagesToForward.size(), leader);
851 for (Object message : messagesToForward) {
852 leader.tell(message, self());
856 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
857 + "change and the leader address isn't available.", this);
860 // We have become the leader, we need to reconstruct frontend state
861 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
862 LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
865 if (!isIsolatedLeader()) {
866 messageRetrySupport.retryMessages();
871 * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
873 * @return the converted messages
875 public Collection<?> convertPendingTransactionsToMessages() {
876 return commitCoordinator.convertPendingTransactionsToMessages(
877 datastoreContext.getShardBatchedModificationCount());
881 protected void pauseLeader(final Runnable operation) {
882 LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
885 // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
886 knownFrontends.values().forEach(LeaderFrontendState::retire);
887 knownFrontends = ImmutableMap.of();
889 store.setRunOnPendingTransactionsComplete(operation);
893 protected void unpauseLeader() {
894 LOG.debug("{}: In unpauseLeader", persistenceId());
897 store.setRunOnPendingTransactionsComplete(null);
899 // Restore tell-based protocol state as if we were becoming the leader
900 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
904 protected OnDemandRaftState.AbstractBuilder<?> newOnDemandRaftStateBuilder() {
905 return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
906 .dataChangeListenerActors(changeSupport.getListenerActors())
907 .commitCohortActors(store.getCohortActors());
911 public String persistenceId() {
916 ShardCommitCoordinator getCommitCoordinator() {
917 return commitCoordinator;
920 public DatastoreContext getDatastoreContext() {
921 return datastoreContext;
925 public ShardDataTree getDataStore() {
930 ShardStats getShardMBean() {
934 public static Builder builder() {
935 return new Builder();
938 public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
939 private final Class<S> shardClass;
940 private ShardIdentifier id;
941 private Map<String, String> peerAddresses = Collections.emptyMap();
942 private DatastoreContext datastoreContext;
943 private SchemaContextProvider schemaContextProvider;
944 private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
945 private TipProducingDataTree dataTree;
946 private volatile boolean sealed;
948 protected AbstractBuilder(final Class<S> shardClass) {
949 this.shardClass = shardClass;
952 protected void checkSealed() {
953 Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
956 @SuppressWarnings("unchecked")
961 public T id(final ShardIdentifier newId) {
967 public T peerAddresses(final Map<String, String> newPeerAddresses) {
969 this.peerAddresses = newPeerAddresses;
973 public T datastoreContext(final DatastoreContext newDatastoreContext) {
975 this.datastoreContext = newDatastoreContext;
979 public T schemaContextProvider(final SchemaContextProvider schemaContextProvider) {
981 this.schemaContextProvider = Preconditions.checkNotNull(schemaContextProvider);
985 public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
987 this.restoreFromSnapshot = newRestoreFromSnapshot;
991 public T dataTree(final TipProducingDataTree newDataTree) {
993 this.dataTree = newDataTree;
997 public ShardIdentifier getId() {
1001 public Map<String, String> getPeerAddresses() {
1002 return peerAddresses;
1005 public DatastoreContext getDatastoreContext() {
1006 return datastoreContext;
1009 public SchemaContext getSchemaContext() {
1010 return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
1013 public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1014 return restoreFromSnapshot;
1017 public TipProducingDataTree getDataTree() {
1021 public TreeType getTreeType() {
1022 switch (datastoreContext.getLogicalStoreType()) {
1024 return TreeType.CONFIGURATION;
1026 return TreeType.OPERATIONAL;
1028 throw new IllegalStateException("Unhandled logical store type "
1029 + datastoreContext.getLogicalStoreType());
1033 protected void verify() {
1034 Preconditions.checkNotNull(id, "id should not be null");
1035 Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
1036 Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
1037 Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
1040 public Props props() {
1043 return Props.create(shardClass, this);
1047 public static class Builder extends AbstractBuilder<Builder, Shard> {
1054 return Ticker.systemTicker();
1057 void scheduleNextPendingTransaction() {
1058 self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());