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;
34 import java.util.concurrent.TimeUnit;
35 import javax.annotation.Nonnull;
36 import javax.annotation.Nullable;
37 import org.opendaylight.controller.cluster.access.ABIVersion;
38 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
39 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
40 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
41 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
42 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
43 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
44 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
45 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
46 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
47 import org.opendaylight.controller.cluster.access.concepts.Request;
48 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
49 import org.opendaylight.controller.cluster.access.concepts.RequestException;
50 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
51 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
52 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
53 import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
54 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
55 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
56 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
57 import org.opendaylight.controller.cluster.common.actor.Dispatchers;
58 import org.opendaylight.controller.cluster.common.actor.Dispatchers.DispatcherType;
59 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
60 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
61 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
62 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
63 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
64 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardDataTreeListenerInfoMXBeanImpl;
65 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardMBeanFactory;
66 import org.opendaylight.controller.cluster.datastore.jmx.mbeans.shard.ShardStats;
67 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
68 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
69 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
70 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
71 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
72 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
73 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
74 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
75 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
76 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
77 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
78 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
79 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
80 import org.opendaylight.controller.cluster.datastore.messages.PersistAbortTransactionPayload;
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.AbortTransactionPayload;
86 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
87 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
88 import org.opendaylight.controller.cluster.datastore.persisted.DisableTrackingPayload;
89 import org.opendaylight.controller.cluster.datastore.persisted.PurgeTransactionPayload;
90 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
91 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
92 import org.opendaylight.controller.cluster.messaging.SliceOptions;
93 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
94 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
95 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
96 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
97 import org.opendaylight.controller.cluster.raft.RaftActor;
98 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
99 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
100 import org.opendaylight.controller.cluster.raft.RaftState;
101 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
102 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
103 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
104 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
105 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
106 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
107 import org.opendaylight.yangtools.concepts.Identifier;
108 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTree;
109 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
110 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
111 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
112 import org.opendaylight.yangtools.yang.model.api.SchemaContextProvider;
113 import scala.concurrent.duration.FiniteDuration;
116 * A Shard represents a portion of the logical data tree.
119 * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
121 public class Shard extends RaftActor {
124 static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
126 public String toString() {
127 return "txCommitTimeoutCheck";
132 static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
134 public String toString() {
135 return "getShardMBeanMessage";
139 static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
141 public String toString() {
142 return "resumeNextPendingTransaction";
146 // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
147 public static final String DEFAULT_NAME = "default";
149 private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
152 final ABIVersion[] values = ABIVersion.values();
153 final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
154 SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
157 // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
158 private static final int CLIENT_MAX_MESSAGES = 1000;
160 // The state of this Shard
161 private final ShardDataTree store;
163 /// The name of this shard
164 private final String name;
166 private final String shardName;
168 private final ShardStats shardMBean;
170 private final ShardDataTreeListenerInfoMXBeanImpl listenerInfoMXBean;
172 private DatastoreContext datastoreContext;
174 private final ShardCommitCoordinator commitCoordinator;
176 private long transactionCommitTimeout;
178 private Cancellable txCommitTimeoutCheckSchedule;
180 private final Optional<ActorRef> roleChangeNotifier;
182 private final MessageTracker appendEntriesReplyTracker;
184 private final ShardTransactionActorFactory transactionActorFactory;
186 private final ShardSnapshotCohort snapshotCohort;
188 private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
190 private ShardSnapshot restoreFromSnapshot;
192 private final ShardTransactionMessageRetrySupport messageRetrySupport;
194 private final FrontendMetadata frontendMetadata;
195 private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
196 private boolean paused;
198 private final MessageSlicer responseMessageSlicer;
199 private final Dispatchers dispatchers;
201 private final MessageAssembler requestMessageAssembler;
203 protected Shard(final AbstractBuilder<?, ?> builder) {
204 super(builder.getId().toString(), builder.getPeerAddresses(),
205 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
207 this.name = builder.getId().toString();
208 this.shardName = builder.getId().getShardName();
209 this.datastoreContext = builder.getDatastoreContext();
210 this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
211 this.frontendMetadata = new FrontendMetadata(name);
213 setPersistence(datastoreContext.isPersistent());
215 LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
217 ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
218 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
219 if (builder.getDataTree() != null) {
220 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
221 treeChangeListenerPublisher, name, frontendMetadata);
223 store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
224 builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher, name, frontendMetadata);
227 shardMBean = ShardMBeanFactory.getShardStatsMBean(name, datastoreContext.getDataStoreMXBeanType(), this);
229 if (isMetricsCaptureEnabled()) {
230 getContext().become(new MeteringBehavior(this));
233 commitCoordinator = new ShardCommitCoordinator(store, LOG, this.name);
235 setTransactionCommitTimeout();
237 // create a notifier actor for each cluster member
238 roleChangeNotifier = createRoleChangeNotifier(name);
240 appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
241 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
243 dispatchers = new Dispatchers(context().system().dispatchers());
244 transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
245 dispatchers.getDispatcherPath(Dispatchers.DispatcherType.Transaction),
246 self(), getContext(), shardMBean, builder.getId().getShardName());
248 snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
251 messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
253 responseMessageSlicer = MessageSlicer.builder().logContext(this.name)
254 .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
255 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
256 .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
258 requestMessageAssembler = MessageAssembler.builder().logContext(this.name)
259 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
260 .assembledMessageCallback((message, sender) -> self().tell(message, sender))
261 .expireStateAfterInactivity(datastoreContext.getRequestTimeout(), TimeUnit.NANOSECONDS).build();
263 listenerInfoMXBean = new ShardDataTreeListenerInfoMXBeanImpl(name, datastoreContext.getDataStoreMXBeanType(),
265 listenerInfoMXBean.register();
268 private void setTransactionCommitTimeout() {
269 transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
270 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
273 private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
274 ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
275 RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
276 return Optional.of(shardRoleChangeNotifier);
280 public void postStop() {
281 LOG.info("Stopping Shard {}", persistenceId());
285 messageRetrySupport.close();
287 if (txCommitTimeoutCheckSchedule != null) {
288 txCommitTimeoutCheckSchedule.cancel();
291 commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
293 shardMBean.unregisterMBean();
294 listenerInfoMXBean.unregister();
298 protected void handleRecover(final Object message) {
299 LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
302 super.handleRecover(message);
303 if (LOG.isTraceEnabled()) {
304 appendEntriesReplyTracker.begin();
309 protected void handleNonRaftCommand(final Object message) {
310 try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
311 final Optional<Error> maybeError = context.error();
312 if (maybeError.isPresent()) {
313 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
317 store.resetTransactionBatch();
319 if (message instanceof RequestEnvelope) {
320 handleRequestEnvelope((RequestEnvelope)message);
321 } else if (MessageAssembler.isHandledMessage(message)) {
322 handleRequestAssemblerMessage(message);
323 } else if (message instanceof ConnectClientRequest) {
324 handleConnectClient((ConnectClientRequest)message);
325 } else if (CreateTransaction.isSerializedType(message)) {
326 handleCreateTransaction(message);
327 } else if (message instanceof BatchedModifications) {
328 handleBatchedModifications((BatchedModifications)message);
329 } else if (message instanceof ForwardedReadyTransaction) {
330 handleForwardedReadyTransaction((ForwardedReadyTransaction) message);
331 } else if (message instanceof ReadyLocalTransaction) {
332 handleReadyLocalTransaction((ReadyLocalTransaction)message);
333 } else if (CanCommitTransaction.isSerializedType(message)) {
334 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
335 } else if (CommitTransaction.isSerializedType(message)) {
336 handleCommitTransaction(CommitTransaction.fromSerializable(message));
337 } else if (AbortTransaction.isSerializedType(message)) {
338 handleAbortTransaction(AbortTransaction.fromSerializable(message));
339 } else if (CloseTransactionChain.isSerializedType(message)) {
340 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
341 } else if (message instanceof RegisterDataTreeChangeListener) {
342 treeChangeSupport.onMessage((RegisterDataTreeChangeListener) message, isLeader(), hasLeader());
343 } else if (message instanceof UpdateSchemaContext) {
344 updateSchemaContext((UpdateSchemaContext) message);
345 } else if (message instanceof PeerAddressResolved) {
346 PeerAddressResolved resolved = (PeerAddressResolved) message;
347 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
348 } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
349 commitTimeoutCheck();
350 } else if (message instanceof DatastoreContext) {
351 onDatastoreContext((DatastoreContext)message);
352 } else if (message instanceof RegisterRoleChangeListener) {
353 roleChangeNotifier.get().forward(message, context());
354 } else if (message instanceof FollowerInitialSyncUpStatus) {
355 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
356 context().parent().tell(message, self());
357 } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
358 sender().tell(getShardMBean(), self());
359 } else if (message instanceof GetShardDataTree) {
360 sender().tell(store.getDataTree(), self());
361 } else if (message instanceof ServerRemoved) {
362 context().parent().forward(message, context());
363 } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
364 messageRetrySupport.onTimerMessage(message);
365 } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
366 store.processCohortRegistryCommand(getSender(),
367 (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
368 } else if (message instanceof PersistAbortTransactionPayload) {
369 final TransactionIdentifier txId = ((PersistAbortTransactionPayload) message).getTransactionId();
370 persistPayload(txId, AbortTransactionPayload.create(txId,
371 datastoreContext.getInitialPayloadSerializedBufferCapacity()), true);
372 persistPayload(txId, PurgeTransactionPayload.create(txId,
373 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
374 } else if (message instanceof MakeLeaderLocal) {
376 } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
377 store.resumeNextPendingTransaction();
378 } else if (!responseMessageSlicer.handleMessage(message)) {
379 super.handleNonRaftCommand(message);
384 private void handleRequestAssemblerMessage(final Object message) {
385 dispatchers.getDispatcher(DispatcherType.Serialization).execute(() -> {
386 JavaSerializer.currentSystem().value_$eq((ExtendedActorSystem) context().system());
387 requestMessageAssembler.handleMessage(message, self());
391 @SuppressWarnings("checkstyle:IllegalCatch")
392 private void handleRequestEnvelope(final RequestEnvelope envelope) {
393 final long now = ticker().read();
395 final RequestSuccess<?, ?> success = handleRequest(envelope, now);
396 if (success != null) {
397 final long executionTimeNanos = ticker().read() - now;
398 if (success instanceof SliceableMessage) {
399 dispatchers.getDispatcher(DispatcherType.Serialization).execute(() ->
400 responseMessageSlicer.slice(SliceOptions.builder().identifier(success.getTarget())
401 .message(envelope.newSuccessEnvelope(success, executionTimeNanos))
402 .sendTo(envelope.getMessage().getReplyTo()).replyTo(self())
403 .onFailureCallback(t -> LOG.warn("Error slicing response {}", success, t)).build()));
405 envelope.sendSuccess(success, executionTimeNanos);
408 } catch (RequestException e) {
409 LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
410 envelope.sendFailure(e, ticker().read() - now);
411 } catch (Exception e) {
412 LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
413 envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
414 ticker().read() - now);
418 private void commitTimeoutCheck() {
419 store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
420 commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
421 requestMessageAssembler.checkExpiredAssembledMessageState();
424 private Optional<Long> updateAccess(final SimpleShardDataTreeCohort cohort) {
425 final FrontendIdentifier frontend = cohort.getIdentifier().getHistoryId().getClientId().getFrontendId();
426 final LeaderFrontendState state = knownFrontends.get(frontend);
428 // Not tell-based protocol, do nothing
429 return Optional.absent();
432 if (isIsolatedLeader()) {
433 // We are isolated and no new request can come through until we emerge from it. We are still updating
434 // liveness of frontend when we see it attempting to communicate. Use the last access timer.
435 return Optional.of(state.getLastSeenTicks());
438 // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
439 return Optional.of(state.getLastConnectTicks());
442 private void disableTracking(final DisableTrackingPayload payload) {
443 final ClientIdentifier clientId = payload.getIdentifier();
444 LOG.debug("{}: disabling tracking of {}", persistenceId(), clientId);
445 frontendMetadata.disableTracking(clientId);
448 final FrontendIdentifier frontendId = clientId.getFrontendId();
449 final LeaderFrontendState frontend = knownFrontends.get(frontendId);
450 if (frontend != null) {
451 if (clientId.equals(frontend.getIdentifier())) {
452 if (!(frontend instanceof LeaderFrontendState.Disabled)) {
453 verify(knownFrontends.replace(frontendId, frontend,
454 new LeaderFrontendState.Disabled(persistenceId(), clientId, store)));
455 LOG.debug("{}: leader state for {} disabled", persistenceId(), clientId);
457 LOG.debug("{}: leader state {} is already disabled", persistenceId(), frontend);
460 LOG.debug("{}: leader state {} does not match {}", persistenceId(), frontend, clientId);
463 LOG.debug("{}: leader state for {} not found", persistenceId(), clientId);
468 private void onMakeLeaderLocal() {
469 LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
471 getSender().tell(new Status.Success(null), getSelf());
475 final ActorSelection leader = getLeader();
477 if (leader == null) {
478 // Leader is not present. The cluster is most likely trying to
479 // elect a leader and we should let that run its normal course
481 // TODO we can wait for the election to complete and retry the
482 // request. We can also let the caller retry by sending a flag
483 // in the response indicating the request is "reTryable".
484 getSender().tell(new Failure(
485 new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
486 + "Currently there is no leader for " + persistenceId())),
491 leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
494 // Acquire our frontend tracking handle and verify generation matches
496 private LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
497 final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
498 if (existing != null) {
499 final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
505 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
506 throw new RetiredGenerationException(clientId.getGeneration(),
507 existing.getIdentifier().getGeneration());
510 LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
512 knownFrontends.remove(clientId.getFrontendId());
514 LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
520 private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
521 final LeaderFrontendState ret = findFrontend(clientId);
526 // TODO: a dedicated exception would be better, but this is technically true, too
527 throw new OutOfSequenceEnvelopeException(0);
531 private static ABIVersion selectVersion(final ConnectClientRequest message) {
532 final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
533 for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
534 if (clientRange.contains(v)) {
539 throw new IllegalArgumentException(String.format(
540 "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
544 @SuppressWarnings("checkstyle:IllegalCatch")
545 private void handleConnectClient(final ConnectClientRequest message) {
547 final ClientIdentifier clientId = message.getTarget();
548 final LeaderFrontendState existing = findFrontend(clientId);
549 if (existing != null) {
553 if (!isLeader() || !isLeaderActive()) {
554 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
555 + "isLeadershipTransferInProgress: {}.",
556 persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
557 throw new NotLeaderException(getSelf());
560 final ABIVersion selectedVersion = selectVersion(message);
561 final LeaderFrontendState frontend;
562 if (existing == null) {
563 frontend = new LeaderFrontendState.Enabled(persistenceId(), clientId, store);
564 knownFrontends.put(clientId.getFrontendId(), frontend);
565 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
570 frontend.reconnect();
571 message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
572 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
573 ActorRef.noSender());
574 } catch (RequestException | RuntimeException e) {
575 message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
580 private RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
581 throws RequestException {
582 // We are not the leader, hence we want to fail-fast.
583 if (!isLeader() || paused || !isLeaderActive()) {
584 LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
585 + "isLeadershipTransferInProgress: {}, paused: {}",
586 persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
587 throw new NotLeaderException(getSelf());
590 final Request<?, ?> request = envelope.getMessage();
591 if (request instanceof TransactionRequest) {
592 final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
593 final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
594 return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
595 } else if (request instanceof LocalHistoryRequest) {
596 final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
597 final ClientIdentifier clientId = lhReq.getTarget().getClientId();
598 return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
600 LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
601 throw new UnsupportedRequestException(request);
605 private boolean hasLeader() {
606 return getLeaderId() != null;
609 public int getPendingTxCommitQueueSize() {
610 return store.getQueueSize();
613 public int getCohortCacheSize() {
614 return commitCoordinator.getCohortCacheSize();
618 protected Optional<ActorRef> getRoleChangeNotifier() {
619 return roleChangeNotifier;
622 String getShardName() {
627 protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
628 final short leaderPayloadVersion) {
629 return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
630 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
633 protected void onDatastoreContext(final DatastoreContext context) {
634 datastoreContext = context;
636 setTransactionCommitTimeout();
638 setPersistence(datastoreContext.isPersistent());
640 updateConfigParams(datastoreContext.getShardRaftConfig());
643 // applyState() will be invoked once consensus is reached on the payload
644 void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
645 boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
646 if (canSkipPayload) {
647 applyState(self(), id, payload);
649 // We are faking the sender
650 persistData(self(), id, payload, batchHint);
654 private void handleCommitTransaction(final CommitTransaction commit) {
655 final TransactionIdentifier txId = commit.getTransactionId();
657 askProtocolEncountered(txId);
658 commitCoordinator.handleCommit(txId, getSender(), this);
660 ActorSelection leader = getLeader();
661 if (leader == null) {
662 messageRetrySupport.addMessageToRetry(commit, getSender(), "Could not commit transaction " + txId);
664 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
665 leader.forward(commit, getContext());
670 private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
671 final TransactionIdentifier txId = canCommit.getTransactionId();
672 LOG.debug("{}: Can committing transaction {}", persistenceId(), txId);
675 askProtocolEncountered(txId);
676 commitCoordinator.handleCanCommit(txId, getSender(), this);
678 ActorSelection leader = getLeader();
679 if (leader == null) {
680 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
681 "Could not canCommit transaction " + txId);
683 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
684 leader.forward(canCommit, getContext());
689 @SuppressWarnings("checkstyle:IllegalCatch")
690 protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
691 askProtocolEncountered(batched.getTransactionId());
694 commitCoordinator.handleBatchedModifications(batched, sender, this);
695 } catch (Exception e) {
696 LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
697 batched.getTransactionId(), e);
698 sender.tell(new Failure(e), getSelf());
702 private void handleBatchedModifications(final BatchedModifications batched) {
703 // This message is sent to prepare the modifications transaction directly on the Shard as an
704 // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
705 // BatchedModifications message, the caller sets the ready flag in the message indicating
706 // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
707 // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
708 // ReadyTransaction message.
710 // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
711 // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
712 // the primary/leader shard. However with timing and caching on the front-end, there's a small
713 // window where it could have a stale leader during leadership transitions.
715 boolean isLeaderActive = isLeaderActive();
716 if (isLeader() && isLeaderActive) {
717 handleBatchedModificationsLocal(batched, getSender());
719 ActorSelection leader = getLeader();
720 if (!isLeaderActive || leader == null) {
721 messageRetrySupport.addMessageToRetry(batched, getSender(),
722 "Could not process BatchedModifications " + batched.getTransactionId());
724 // If this is not the first batch and leadership changed in between batched messages,
725 // we need to reconstruct previous BatchedModifications from the transaction
726 // DataTreeModification, honoring the max batched modification count, and forward all the
727 // previous BatchedModifications to the new leader.
728 Collection<BatchedModifications> newModifications = commitCoordinator
729 .createForwardedBatchedModifications(batched,
730 datastoreContext.getShardBatchedModificationCount());
732 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
733 newModifications.size(), leader);
735 for (BatchedModifications bm : newModifications) {
736 leader.forward(bm, getContext());
742 private boolean failIfIsolatedLeader(final ActorRef sender) {
743 if (isIsolatedLeader()) {
744 sender.tell(new Failure(new NoShardLeaderException(String.format(
745 "Shard %s was the leader but has lost contact with all of its followers. Either all"
746 + " other follower nodes are down or this node is isolated by a network partition.",
747 persistenceId()))), getSelf());
754 protected boolean isIsolatedLeader() {
755 return getRaftState() == RaftState.IsolatedLeader;
758 @SuppressWarnings("checkstyle:IllegalCatch")
759 private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
760 LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
762 boolean isLeaderActive = isLeaderActive();
763 if (isLeader() && isLeaderActive) {
765 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
766 } catch (Exception e) {
767 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
768 message.getTransactionId(), e);
769 getSender().tell(new Failure(e), getSelf());
772 ActorSelection leader = getLeader();
773 if (!isLeaderActive || leader == null) {
774 messageRetrySupport.addMessageToRetry(message, getSender(),
775 "Could not process ready local transaction " + message.getTransactionId());
777 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
778 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
779 leader.forward(message, getContext());
784 private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
785 LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
787 boolean isLeaderActive = isLeaderActive();
788 if (isLeader() && isLeaderActive) {
789 askProtocolEncountered(forwardedReady.getTransactionId());
790 commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
792 ActorSelection leader = getLeader();
793 if (!isLeaderActive || leader == null) {
794 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
795 "Could not process forwarded ready transaction " + forwardedReady.getTransactionId());
797 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
799 ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
800 forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit(),
801 forwardedReady.getParticipatingShardNames());
802 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
803 leader.forward(readyLocal, getContext());
808 private void handleAbortTransaction(final AbortTransaction abort) {
809 final TransactionIdentifier transactionId = abort.getTransactionId();
810 askProtocolEncountered(transactionId);
811 doAbortTransaction(transactionId, getSender());
814 void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
815 commitCoordinator.handleAbort(transactionID, sender, this);
818 private void handleCreateTransaction(final Object message) {
820 createTransaction(CreateTransaction.fromSerializable(message));
821 } else if (getLeader() != null) {
822 getLeader().forward(message, getContext());
824 getSender().tell(new Failure(new NoShardLeaderException(
825 "Could not create a shard transaction", persistenceId())), getSelf());
829 private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
831 final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
832 askProtocolEncountered(id.getClientId());
834 // FIXME: CONTROLLER-1628: stage purge once no transactions are present
835 store.closeTransactionChain(id, null);
836 store.purgeTransactionChain(id, null);
837 } else if (getLeader() != null) {
838 getLeader().forward(closeTransactionChain, getContext());
840 LOG.warn("{}: Could not close transaction {}", persistenceId(), closeTransactionChain.getIdentifier());
844 @SuppressWarnings("checkstyle:IllegalCatch")
845 private void createTransaction(final CreateTransaction createTransaction) {
846 askProtocolEncountered(createTransaction.getTransactionId());
849 if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
850 && failIfIsolatedLeader(getSender())) {
854 ActorRef transactionActor = createTransaction(createTransaction.getTransactionType(),
855 createTransaction.getTransactionId());
857 getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
858 createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
859 } catch (Exception e) {
860 getSender().tell(new Failure(e), getSelf());
864 private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
865 LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
866 return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
870 // Called on leader only
871 private void askProtocolEncountered(final TransactionIdentifier transactionId) {
872 askProtocolEncountered(transactionId.getHistoryId().getClientId());
875 // Called on leader only
876 private void askProtocolEncountered(final ClientIdentifier clientId) {
877 final LeaderFrontendState state = knownFrontends.get(clientId.getFrontendId());
878 if (state instanceof LeaderFrontendState.Enabled) {
879 LOG.debug("{}: encountered ask-based client {}, disabling transaction tracking", persistenceId(), clientId);
880 persistPayload(clientId, DisableTrackingPayload.create(clientId,
881 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
885 private void updateSchemaContext(final UpdateSchemaContext message) {
886 updateSchemaContext(message.getSchemaContext());
890 void updateSchemaContext(final SchemaContext schemaContext) {
891 store.updateSchemaContext(schemaContext);
894 private boolean isMetricsCaptureEnabled() {
895 CommonConfig config = new CommonConfig(getContext().system().settings().config());
896 return config.isMetricCaptureEnabled();
901 public RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
902 return snapshotCohort;
907 protected RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
908 if (restoreFromSnapshot == null) {
909 return ShardRecoveryCoordinator.create(store, persistenceId(), LOG);
912 return ShardRecoveryCoordinator.forSnapshot(store, persistenceId(), LOG, restoreFromSnapshot.getSnapshot());
916 protected void onRecoveryComplete() {
917 restoreFromSnapshot = null;
919 //notify shard manager
920 getContext().parent().tell(new ActorInitialized(), getSelf());
922 // Being paranoid here - this method should only be called once but just in case...
923 if (txCommitTimeoutCheckSchedule == null) {
924 // Schedule a message to be periodically sent to check if the current in-progress
925 // transaction should be expired and aborted.
926 FiniteDuration period = FiniteDuration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
927 txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
928 period, period, getSelf(),
929 TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
934 protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
935 if (data instanceof Payload) {
936 if (data instanceof DisableTrackingPayload) {
937 disableTracking((DisableTrackingPayload) data);
942 store.applyReplicatedPayload(identifier, (Payload)data);
943 } catch (DataValidationFailedException | IOException e) {
944 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
947 LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
952 protected void onStateChanged() {
953 boolean isLeader = isLeader();
954 boolean hasLeader = hasLeader();
955 treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
957 // If this actor is no longer the leader close all the transaction chains
959 if (LOG.isDebugEnabled()) {
961 "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
962 persistenceId(), getId());
966 store.purgeLeaderState();
969 if (hasLeader && !isIsolatedLeader()) {
970 messageRetrySupport.retryMessages();
975 protected void onLeaderChanged(final String oldLeader, final String newLeader) {
976 shardMBean.incrementLeadershipChangeCount();
980 if (!knownFrontends.isEmpty()) {
981 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
982 knownFrontends = ImmutableMap.of();
985 requestMessageAssembler.close();
988 // No leader anywhere, nothing else to do
992 // Another leader was elected. If we were the previous leader and had pending transactions, convert
993 // them to transaction messages and send to the new leader.
994 ActorSelection leader = getLeader();
995 if (leader != null) {
996 Collection<?> messagesToForward = convertPendingTransactionsToMessages();
998 if (!messagesToForward.isEmpty()) {
999 LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
1000 messagesToForward.size(), leader);
1002 for (Object message : messagesToForward) {
1003 LOG.debug("{}: Forwarding pending transaction message {}", persistenceId(), message);
1005 leader.tell(message, self());
1009 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
1010 + "change and the leader address isn't available.", this);
1013 // We have become the leader, we need to reconstruct frontend state
1014 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
1015 LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
1018 if (!isIsolatedLeader()) {
1019 messageRetrySupport.retryMessages();
1024 * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
1026 * @return the converted messages
1028 public Collection<?> convertPendingTransactionsToMessages() {
1029 return commitCoordinator.convertPendingTransactionsToMessages(
1030 datastoreContext.getShardBatchedModificationCount());
1034 protected void pauseLeader(final Runnable operation) {
1035 LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
1038 // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
1039 knownFrontends.values().forEach(LeaderFrontendState::retire);
1040 knownFrontends = ImmutableMap.of();
1042 store.setRunOnPendingTransactionsComplete(operation);
1046 protected void unpauseLeader() {
1047 LOG.debug("{}: In unpauseLeader", persistenceId());
1050 store.setRunOnPendingTransactionsComplete(null);
1052 // Restore tell-based protocol state as if we were becoming the leader
1053 knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
1057 protected OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
1058 return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
1059 .commitCohortActors(store.getCohortActors());
1063 public String persistenceId() {
1068 ShardCommitCoordinator getCommitCoordinator() {
1069 return commitCoordinator;
1072 public DatastoreContext getDatastoreContext() {
1073 return datastoreContext;
1077 public ShardDataTree getDataStore() {
1082 ShardStats getShardMBean() {
1086 public static Builder builder() {
1087 return new Builder();
1090 public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
1091 private final Class<S> shardClass;
1092 private ShardIdentifier id;
1093 private Map<String, String> peerAddresses = Collections.emptyMap();
1094 private DatastoreContext datastoreContext;
1095 private SchemaContextProvider schemaContextProvider;
1096 private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
1097 private DataTree dataTree;
1098 private volatile boolean sealed;
1100 protected AbstractBuilder(final Class<S> shardClass) {
1101 this.shardClass = shardClass;
1104 protected void checkSealed() {
1105 Preconditions.checkState(!sealed, "Builder isalready sealed - further modifications are not allowed");
1108 @SuppressWarnings("unchecked")
1113 public T id(final ShardIdentifier newId) {
1119 public T peerAddresses(final Map<String, String> newPeerAddresses) {
1121 this.peerAddresses = newPeerAddresses;
1125 public T datastoreContext(final DatastoreContext newDatastoreContext) {
1127 this.datastoreContext = newDatastoreContext;
1131 public T schemaContextProvider(final SchemaContextProvider newSchemaContextProvider) {
1133 this.schemaContextProvider = Preconditions.checkNotNull(newSchemaContextProvider);
1137 public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
1139 this.restoreFromSnapshot = newRestoreFromSnapshot;
1143 public T dataTree(final DataTree newDataTree) {
1145 this.dataTree = newDataTree;
1149 public ShardIdentifier getId() {
1153 public Map<String, String> getPeerAddresses() {
1154 return peerAddresses;
1157 public DatastoreContext getDatastoreContext() {
1158 return datastoreContext;
1161 public SchemaContext getSchemaContext() {
1162 return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
1165 public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1166 return restoreFromSnapshot;
1169 public DataTree getDataTree() {
1173 public TreeType getTreeType() {
1174 switch (datastoreContext.getLogicalStoreType()) {
1176 return TreeType.CONFIGURATION;
1178 return TreeType.OPERATIONAL;
1180 throw new IllegalStateException("Unhandled logical store type "
1181 + datastoreContext.getLogicalStoreType());
1185 protected void verify() {
1186 Preconditions.checkNotNull(id, "id should not be null");
1187 Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
1188 Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
1189 Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
1192 public Props props() {
1195 return Props.create(shardClass, this);
1199 public static class Builder extends AbstractBuilder<Builder, Shard> {
1206 return Ticker.systemTicker();
1209 void scheduleNextPendingTransaction() {
1210 self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());