opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/Shard.java

   1 /*
   2  * Copyright (c) 2014 Cisco Systems, Inc. and others.  All rights reserved.
   3  *
   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
   7  */
   8 package org.opendaylight.controller.cluster.datastore;
   9
  10 import static com.google.common.base.Preconditions.checkState;
  11 import static com.google.common.base.Verify.verify;
  12 import static com.google.common.base.Verify.verifyNotNull;
  13 import static java.util.Objects.requireNonNull;
  14
  15 import akka.actor.ActorRef;
  16 import akka.actor.ActorSelection;
  17 import akka.actor.Cancellable;
  18 import akka.actor.ExtendedActorSystem;
  19 import akka.actor.PoisonPill;
  20 import akka.actor.Props;
  21 import akka.actor.Status;
  22 import akka.actor.Status.Failure;
  23 import akka.persistence.RecoveryCompleted;
  24 import akka.persistence.SnapshotOffer;
  25 import akka.serialization.JavaSerializer;
  26 import akka.serialization.Serialization;
  27 import com.google.common.annotations.VisibleForTesting;
  28 import com.google.common.base.Ticker;
  29 import com.google.common.collect.ImmutableList;
  30 import com.google.common.collect.ImmutableMap;
  31 import com.google.common.collect.ImmutableSet;
  32 import com.google.common.collect.Range;
  33 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
  34 import java.io.IOException;
  35 import java.util.Arrays;
  36 import java.util.Collection;
  37 import java.util.Collections;
  38 import java.util.HashMap;
  39 import java.util.Map;
  40 import java.util.Optional;
  41 import java.util.OptionalLong;
  42 import java.util.concurrent.TimeUnit;
  43 import org.eclipse.jdt.annotation.NonNull;
  44 import org.eclipse.jdt.annotation.Nullable;
  45 import org.opendaylight.controller.cluster.access.ABIVersion;
  46 import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
  47 import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
  48 import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
  49 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
  50 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
  51 import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
  52 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
  53 import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
  54 import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
  55 import org.opendaylight.controller.cluster.access.concepts.RequestException;
  56 import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
  57 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
  58 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
  59 import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
  60 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
  61 import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
  62 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
  63 import org.opendaylight.controller.cluster.common.actor.Dispatchers;
  64 import org.opendaylight.controller.cluster.common.actor.Dispatchers.DispatcherType;
  65 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
  66 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
  67 import org.opendaylight.controller.cluster.datastore.actors.JsonExportActor;
  68 import org.opendaylight.controller.cluster.datastore.exceptions.NoShardLeaderException;
  69 import org.opendaylight.controller.cluster.datastore.identifiers.ShardIdentifier;
  70 import org.opendaylight.controller.cluster.datastore.messages.AbortTransaction;
  71 import org.opendaylight.controller.cluster.datastore.messages.ActorInitialized;
  72 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
  73 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
  74 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
  75 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
  76 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
  77 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
  78 import org.opendaylight.controller.cluster.datastore.messages.DataTreeChangedReply;
  79 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
  80 import org.opendaylight.controller.cluster.datastore.messages.GetKnownClients;
  81 import org.opendaylight.controller.cluster.datastore.messages.GetKnownClientsReply;
  82 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
  83 import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
  84 import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
  85 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
  86 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
  87 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
  88 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
  89 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
  90 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
  91 import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
  92 import org.opendaylight.controller.cluster.datastore.persisted.DisableTrackingPayload;
  93 import org.opendaylight.controller.cluster.messaging.MessageAssembler;
  94 import org.opendaylight.controller.cluster.messaging.MessageSlicer;
  95 import org.opendaylight.controller.cluster.messaging.SliceOptions;
  96 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
  97 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
  98 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
  99 import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
 100 import org.opendaylight.controller.cluster.raft.RaftActor;
 101 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
 102 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
 103 import org.opendaylight.controller.cluster.raft.RaftState;
 104 import org.opendaylight.controller.cluster.raft.ReplicatedLogEntry;
 105 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
 106 import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
 107 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
 108 import org.opendaylight.controller.cluster.raft.messages.Payload;
 109 import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
 110 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
 111 import org.opendaylight.yang.gen.v1.urn.opendaylight.params.xml.ns.yang.controller.config.distributed.datastore.provider.rev231229.DataStoreProperties.ExportOnRecovery;
 112 import org.opendaylight.yangtools.concepts.Identifier;
 113 import org.opendaylight.yangtools.yang.data.tree.api.DataTree;
 114 import org.opendaylight.yangtools.yang.data.tree.api.DataValidationFailedException;
 115 import org.opendaylight.yangtools.yang.data.tree.api.TreeType;
 116 import org.opendaylight.yangtools.yang.model.api.EffectiveModelContext;
 117 import org.opendaylight.yangtools.yang.model.api.EffectiveModelContextProvider;
 118 import scala.concurrent.duration.FiniteDuration;
 119
 120 /**
 121  * A Shard represents a portion of the logical data tree.
 122  *
 123  * <p>
 124  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
 125  */
 126 // FIXME: non-final for testing?
 127 public class Shard extends RaftActor {
 128
 129     @VisibleForTesting
 130     static final Object TX_COMMIT_TIMEOUT_CHECK_MESSAGE = new Object() {
 131         @Override
 132         public String toString() {
 133             return "txCommitTimeoutCheck";
 134         }
 135     };
 136
 137     @VisibleForTesting
 138     static final Object GET_SHARD_MBEAN_MESSAGE = new Object() {
 139         @Override
 140         public String toString() {
 141             return "getShardMBeanMessage";
 142         }
 143     };
 144
 145     static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
 146         @Override
 147         public String toString() {
 148             return "resumeNextPendingTransaction";
 149         }
 150     };
 151
 152     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
 153     public static final String DEFAULT_NAME = "default";
 154
 155     private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
 156
 157     // Make sure to keep this in sync with the journal configuration in factory-akka.conf
 158     public static final String NON_PERSISTENT_JOURNAL_ID = "akka.persistence.non-persistent.journal";
 159
 160     static {
 161         final ABIVersion[] values = ABIVersion.values();
 162         final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
 163         SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
 164     }
 165
 166     // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
 167     private static final int CLIENT_MAX_MESSAGES = 1000;
 168
 169     // The state of this Shard
 170     private final ShardDataTree store;
 171
 172     /// The name of this shard
 173     private final String name;
 174
 175     private final String shardName;
 176
 177     private final ShardStats shardMBean;
 178
 179     private final ShardDataTreeListenerInfoMXBeanImpl listenerInfoMXBean;
 180
 181     private DatastoreContext datastoreContext;
 182
 183     @Deprecated(since = "9.0.0", forRemoval = true)
 184     private final ShardCommitCoordinator commitCoordinator;
 185
 186     private long transactionCommitTimeout;
 187
 188     private Cancellable txCommitTimeoutCheckSchedule;
 189
 190     private final Optional<ActorRef> roleChangeNotifier;
 191
 192     private final MessageTracker appendEntriesReplyTracker;
 193
 194     @Deprecated(since = "9.0.0", forRemoval = true)
 195     private final ShardTransactionActorFactory transactionActorFactory;
 196
 197     private final ShardSnapshotCohort snapshotCohort;
 198
 199     private final DataTreeChangeListenerSupport treeChangeSupport = new DataTreeChangeListenerSupport(this);
 200
 201     private ShardSnapshot restoreFromSnapshot;
 202
 203     @Deprecated(since = "9.0.0", forRemoval = true)
 204     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 205
 206     @VisibleForTesting
 207     final FrontendMetadata frontendMetadata;
 208
 209     private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
 210     private boolean paused;
 211
 212     private final MessageSlicer responseMessageSlicer;
 213     private final Dispatchers dispatchers;
 214
 215     private final MessageAssembler requestMessageAssembler;
 216
 217     private final ExportOnRecovery exportOnRecovery;
 218
 219     private final ActorRef exportActor;
 220
 221     @SuppressFBWarnings(value = "MC_OVERRIDABLE_METHOD_CALL_IN_CONSTRUCTOR", justification = "Akka class design")
 222     Shard(final AbstractBuilder<?, ?> builder) {
 223         super(builder.getId().toString(), builder.getPeerAddresses(),
 224                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 225
 226         name = builder.getId().toString();
 227         shardName = builder.getId().getShardName();
 228         datastoreContext = builder.getDatastoreContext();
 229         restoreFromSnapshot = builder.getRestoreFromSnapshot();
 230         frontendMetadata = new FrontendMetadata(name);
 231         exportOnRecovery = datastoreContext.getExportOnRecovery();
 232
 233         exportActor = switch (exportOnRecovery) {
 234             case Json -> getContext().actorOf(JsonExportActor.props(builder.getSchemaContext(),
 235                 datastoreContext.getRecoveryExportBaseDir()));
 236             case Off -> null;
 237         };
 238
 239         setPersistence(datastoreContext.isPersistent());
 240
 241         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 242
 243         ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
 244                 new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
 245         if (builder.getDataTree() != null) {
 246             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
 247                     treeChangeListenerPublisher, name,
 248                     frontendMetadata);
 249         } else {
 250             store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
 251                     builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher, name,
 252                     frontendMetadata);
 253         }
 254
 255         shardMBean = ShardStats.create(name, datastoreContext.getDataStoreMXBeanType(), this);
 256
 257         if (isMetricsCaptureEnabled()) {
 258             getContext().become(new MeteringBehavior(this));
 259         }
 260
 261         commitCoordinator = new ShardCommitCoordinator(store, LOG, name);
 262
 263         setTransactionCommitTimeout();
 264
 265         // create a notifier actor for each cluster member
 266         roleChangeNotifier = createRoleChangeNotifier(name);
 267
 268         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
 269                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 270
 271         dispatchers = new Dispatchers(context().system().dispatchers());
 272         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
 273             dispatchers.getDispatcherPath(Dispatchers.DispatcherType.Transaction),
 274                 self(), getContext(), shardMBean, builder.getId().getShardName());
 275
 276         snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
 277             name, datastoreContext);
 278
 279         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
 280
 281         responseMessageSlicer = MessageSlicer.builder().logContext(name)
 282                 .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
 283                 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
 284                 .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
 285
 286         requestMessageAssembler = MessageAssembler.builder().logContext(name)
 287                 .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
 288                 .assembledMessageCallback((message, sender) -> self().tell(message, sender))
 289                 .expireStateAfterInactivity(datastoreContext.getRequestTimeout(), TimeUnit.NANOSECONDS).build();
 290
 291         listenerInfoMXBean = new ShardDataTreeListenerInfoMXBeanImpl(name, datastoreContext.getDataStoreMXBeanType(),
 292                 self());
 293         listenerInfoMXBean.register();
 294     }
 295
 296     private void setTransactionCommitTimeout() {
 297         transactionCommitTimeout = TimeUnit.MILLISECONDS.convert(
 298                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
 299     }
 300
 301     private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
 302         ActorRef shardRoleChangeNotifier = getContext().actorOf(
 303             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
 304         return Optional.of(shardRoleChangeNotifier);
 305     }
 306
 307     @Override
 308     public final void postStop() throws Exception {
 309         LOG.info("Stopping Shard {}", persistenceId());
 310
 311         super.postStop();
 312
 313         messageRetrySupport.close();
 314
 315         if (txCommitTimeoutCheckSchedule != null) {
 316             txCommitTimeoutCheckSchedule.cancel();
 317         }
 318
 319         commitCoordinator.abortPendingTransactions("Transaction aborted due to shutdown.", this);
 320
 321         shardMBean.unregisterMBean();
 322         listenerInfoMXBean.unregister();
 323     }
 324
 325     @Override
 326     protected final void handleRecover(final Object message) {
 327         LOG.debug("{}: onReceiveRecover: Received message {} from {}", persistenceId(), message.getClass(),
 328             getSender());
 329
 330         super.handleRecover(message);
 331
 332         switch (exportOnRecovery) {
 333             case Json:
 334                 if (message instanceof SnapshotOffer) {
 335                     exportActor.tell(new JsonExportActor.ExportSnapshot(store.readCurrentData().orElseThrow(), name),
 336                             ActorRef.noSender());
 337                 } else if (message instanceof ReplicatedLogEntry replicatedLogEntry) {
 338                     exportActor.tell(new JsonExportActor.ExportJournal(replicatedLogEntry), ActorRef.noSender());
 339                 } else if (message instanceof RecoveryCompleted) {
 340                     exportActor.tell(new JsonExportActor.FinishExport(name), ActorRef.noSender());
 341                     exportActor.tell(PoisonPill.getInstance(), ActorRef.noSender());
 342                 }
 343                 break;
 344             case Off:
 345             default:
 346                 break;
 347         }
 348
 349         if (LOG.isTraceEnabled()) {
 350             appendEntriesReplyTracker.begin();
 351         }
 352     }
 353
 354     @Override
 355     // non-final for TestShard
 356     protected void handleNonRaftCommand(final Object message) {
 357         try (var context = appendEntriesReplyTracker.received(message)) {
 358             final var maybeError = context.error();
 359             if (maybeError.isPresent()) {
 360                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
 361                     maybeError.orElseThrow());
 362             }
 363
 364             store.resetTransactionBatch();
 365
 366             if (message instanceof RequestEnvelope request) {
 367                 handleRequestEnvelope(request);
 368             } else if (MessageAssembler.isHandledMessage(message)) {
 369                 handleRequestAssemblerMessage(message);
 370             } else if (message instanceof ConnectClientRequest request) {
 371                 handleConnectClient(request);
 372             } else if (CreateTransaction.isSerializedType(message)) {
 373                 handleCreateTransaction(message);
 374             } else if (message instanceof BatchedModifications request) {
 375                 handleBatchedModifications(request);
 376             } else if (message instanceof ForwardedReadyTransaction request) {
 377                 handleForwardedReadyTransaction(request);
 378             } else if (message instanceof ReadyLocalTransaction request) {
 379                 handleReadyLocalTransaction(request);
 380             } else if (CanCommitTransaction.isSerializedType(message)) {
 381                 handleCanCommitTransaction(CanCommitTransaction.fromSerializable(message));
 382             } else if (CommitTransaction.isSerializedType(message)) {
 383                 handleCommitTransaction(CommitTransaction.fromSerializable(message));
 384             } else if (AbortTransaction.isSerializedType(message)) {
 385                 handleAbortTransaction(AbortTransaction.fromSerializable(message));
 386             } else if (CloseTransactionChain.isSerializedType(message)) {
 387                 closeTransactionChain(CloseTransactionChain.fromSerializable(message));
 388             } else if (message instanceof DataTreeChangedReply) {
 389                 // Ignore reply
 390             } else if (message instanceof RegisterDataTreeChangeListener request) {
 391                 treeChangeSupport.onMessage(request, isLeader(), hasLeader());
 392             } else if (message instanceof UpdateSchemaContext request) {
 393                 updateSchemaContext(request);
 394             } else if (message instanceof PeerAddressResolved resolved) {
 395                 setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
 396             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
 397                 commitTimeoutCheck();
 398             } else if (message instanceof DatastoreContext request) {
 399                 onDatastoreContext(request);
 400             } else if (message instanceof RegisterRoleChangeListener) {
 401                 roleChangeNotifier.orElseThrow().forward(message, context());
 402             } else if (message instanceof FollowerInitialSyncUpStatus request) {
 403                 shardMBean.setFollowerInitialSyncStatus(request.isInitialSyncDone());
 404                 context().parent().tell(message, self());
 405             } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
 406                 sender().tell(getShardMBean(), self());
 407             } else if (message instanceof GetShardDataTree) {
 408                 sender().tell(store.getDataTree(), self());
 409             } else if (message instanceof ServerRemoved) {
 410                 context().parent().forward(message, context());
 411             } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
 412                 messageRetrySupport.onTimerMessage(message);
 413             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand request) {
 414                 store.processCohortRegistryCommand(getSender(), request);
 415             } else if (message instanceof MakeLeaderLocal) {
 416                 onMakeLeaderLocal();
 417             } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
 418                 store.resumeNextPendingTransaction();
 419             } else if (GetKnownClients.INSTANCE.equals(message)) {
 420                 handleGetKnownClients();
 421             } else if (!responseMessageSlicer.handleMessage(message)) {
 422                 super.handleNonRaftCommand(message);
 423             }
 424         }
 425     }
 426
 427     private void handleRequestAssemblerMessage(final Object message) {
 428         dispatchers.getDispatcher(DispatcherType.Serialization).execute(() -> {
 429             JavaSerializer.currentSystem().value_$eq((ExtendedActorSystem) context().system());
 430             requestMessageAssembler.handleMessage(message, self());
 431         });
 432     }
 433
 434     @SuppressWarnings("checkstyle:IllegalCatch")
 435     private void handleRequestEnvelope(final RequestEnvelope envelope) {
 436         final long now = ticker().read();
 437         try {
 438             final RequestSuccess<?, ?> success = handleRequest(envelope, now);
 439             if (success != null) {
 440                 final long executionTimeNanos = ticker().read() - now;
 441                 if (success instanceof SliceableMessage) {
 442                     dispatchers.getDispatcher(DispatcherType.Serialization).execute(() ->
 443                         responseMessageSlicer.slice(SliceOptions.builder().identifier(success.getTarget())
 444                             .message(envelope.newSuccessEnvelope(success, executionTimeNanos))
 445                             .sendTo(envelope.getMessage().getReplyTo()).replyTo(self())
 446                             .onFailureCallback(t -> LOG.warn("Error slicing response {}", success, t)).build()));
 447                 } else {
 448                     envelope.sendSuccess(success, executionTimeNanos);
 449                 }
 450             }
 451         } catch (RequestException e) {
 452             LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
 453             envelope.sendFailure(e, ticker().read() - now);
 454         } catch (Exception e) {
 455             LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
 456             envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
 457                 ticker().read() - now);
 458         }
 459     }
 460
 461     private void commitTimeoutCheck() {
 462         store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
 463         commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
 464         requestMessageAssembler.checkExpiredAssembledMessageState();
 465     }
 466
 467     private OptionalLong updateAccess(final SimpleShardDataTreeCohort cohort) {
 468         final FrontendIdentifier frontend = cohort.transactionId().getHistoryId().getClientId().getFrontendId();
 469         final LeaderFrontendState state = knownFrontends.get(frontend);
 470         if (state == null) {
 471             // Not tell-based protocol, do nothing
 472             return OptionalLong.empty();
 473         }
 474
 475         if (isIsolatedLeader()) {
 476             // We are isolated and no new request can come through until we emerge from it. We are still updating
 477             // liveness of frontend when we see it attempting to communicate. Use the last access timer.
 478             return OptionalLong.of(state.getLastSeenTicks());
 479         }
 480
 481         // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
 482         return OptionalLong.of(state.getLastConnectTicks());
 483     }
 484
 485     private void disableTracking(final DisableTrackingPayload payload) {
 486         final ClientIdentifier clientId = payload.getIdentifier();
 487         LOG.debug("{}: disabling tracking of {}", persistenceId(), clientId);
 488         frontendMetadata.disableTracking(clientId);
 489
 490         if (isLeader()) {
 491             final FrontendIdentifier frontendId = clientId.getFrontendId();
 492             final LeaderFrontendState frontend = knownFrontends.get(frontendId);
 493             if (frontend != null) {
 494                 if (clientId.equals(frontend.getIdentifier())) {
 495                     if (!(frontend instanceof LeaderFrontendState.Disabled)) {
 496                         verify(knownFrontends.replace(frontendId, frontend,
 497                             new LeaderFrontendState.Disabled(persistenceId(), clientId, store)));
 498                         LOG.debug("{}: leader state for {} disabled", persistenceId(), clientId);
 499                     } else {
 500                         LOG.debug("{}: leader state {} is already disabled", persistenceId(), frontend);
 501                     }
 502                 } else {
 503                     LOG.debug("{}: leader state {} does not match {}", persistenceId(), frontend, clientId);
 504                 }
 505             } else {
 506                 LOG.debug("{}: leader state for {} not found", persistenceId(), clientId);
 507                 knownFrontends.put(frontendId, new LeaderFrontendState.Disabled(persistenceId(), clientId,
 508                     getDataStore()));
 509             }
 510         }
 511     }
 512
 513     private void onMakeLeaderLocal() {
 514         LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
 515         if (isLeader()) {
 516             getSender().tell(new Status.Success(null), getSelf());
 517             return;
 518         }
 519
 520         final ActorSelection leader = getLeader();
 521
 522         if (leader == null) {
 523             // Leader is not present. The cluster is most likely trying to
 524             // elect a leader and we should let that run its normal course
 525
 526             // TODO we can wait for the election to complete and retry the
 527             // request. We can also let the caller retry by sending a flag
 528             // in the response indicating the request is "reTryable".
 529             getSender().tell(new Failure(
 530                     new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
 531                             + "Currently there is no leader for " + persistenceId())),
 532                     getSelf());
 533             return;
 534         }
 535
 536         leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
 537     }
 538
 539     // Acquire our frontend tracking handle and verify generation matches
 540     private @Nullable LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
 541         final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
 542         if (existing != null) {
 543             final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
 544             if (cmp == 0) {
 545                 existing.touch();
 546                 return existing;
 547             }
 548             if (cmp > 0) {
 549                 LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
 550                 throw new RetiredGenerationException(clientId.getGeneration(),
 551                     existing.getIdentifier().getGeneration());
 552             }
 553
 554             LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
 555             existing.retire();
 556             knownFrontends.remove(clientId.getFrontendId());
 557         } else {
 558             LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
 559         }
 560
 561         return null;
 562     }
 563
 564     private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
 565         final LeaderFrontendState ret = findFrontend(clientId);
 566         if (ret != null) {
 567             return ret;
 568         }
 569
 570         // TODO: a dedicated exception would be better, but this is technically true, too
 571         throw new OutOfSequenceEnvelopeException(0);
 572     }
 573
 574     private static @NonNull ABIVersion selectVersion(final ConnectClientRequest message) {
 575         final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
 576         for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
 577             if (clientRange.contains(v)) {
 578                 return v;
 579             }
 580         }
 581
 582         throw new IllegalArgumentException(String.format(
 583             "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
 584             clientRange));
 585     }
 586
 587     @SuppressWarnings("checkstyle:IllegalCatch")
 588     private void handleConnectClient(final ConnectClientRequest message) {
 589         try {
 590             final ClientIdentifier clientId = message.getTarget();
 591             final LeaderFrontendState existing = findFrontend(clientId);
 592             if (existing != null) {
 593                 existing.touch();
 594             }
 595
 596             if (!isLeader() || !isLeaderActive()) {
 597                 LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
 598                                 + "isLeadershipTransferInProgress: {}.",
 599                         persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
 600                 throw new NotLeaderException(getSelf());
 601             }
 602
 603             final ABIVersion selectedVersion = selectVersion(message);
 604             final LeaderFrontendState frontend;
 605             if (existing == null) {
 606                 frontend = new LeaderFrontendState.Enabled(persistenceId(), clientId, store);
 607                 knownFrontends.put(clientId.getFrontendId(), frontend);
 608                 LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
 609             } else {
 610                 frontend = existing;
 611             }
 612
 613             frontend.reconnect();
 614             message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
 615                 ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
 616                 ActorRef.noSender());
 617         } catch (RequestException | RuntimeException e) {
 618             message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
 619         }
 620     }
 621
 622     private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
 623             throws RequestException {
 624         // We are not the leader, hence we want to fail-fast.
 625         if (!isLeader() || paused || !isLeaderActive()) {
 626             LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
 627                             + "isLeadershipTransferInProgress: {}, paused: {}",
 628                     persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
 629             throw new NotLeaderException(getSelf());
 630         }
 631
 632         final var request = envelope.getMessage();
 633         if (request instanceof TransactionRequest<?> txReq) {
 634             final var clientId = txReq.getTarget().getHistoryId().getClientId();
 635             return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
 636         } else if (request instanceof LocalHistoryRequest<?> lhReq) {
 637             final var clientId = lhReq.getTarget().getClientId();
 638             return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
 639         } else {
 640             LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
 641             throw new UnsupportedRequestException(request);
 642         }
 643     }
 644
 645     private void handleGetKnownClients() {
 646         final ImmutableSet<ClientIdentifier> clients;
 647         if (isLeader()) {
 648             clients = knownFrontends.values().stream()
 649                     .map(LeaderFrontendState::getIdentifier)
 650                     .collect(ImmutableSet.toImmutableSet());
 651         } else {
 652             clients = frontendMetadata.getClients();
 653         }
 654         sender().tell(new GetKnownClientsReply(clients), self());
 655     }
 656
 657     private boolean hasLeader() {
 658         return getLeaderId() != null;
 659     }
 660
 661     final int getPendingTxCommitQueueSize() {
 662         return store.getQueueSize();
 663     }
 664
 665     final int getCohortCacheSize() {
 666         return commitCoordinator.getCohortCacheSize();
 667     }
 668
 669     @Override
 670     protected final Optional<ActorRef> getRoleChangeNotifier() {
 671         return roleChangeNotifier;
 672     }
 673
 674     final String getShardName() {
 675         return shardName;
 676     }
 677
 678     @Override
 679     protected final LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
 680             final short leaderPayloadVersion) {
 681         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
 682                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
 683     }
 684
 685     private void onDatastoreContext(final DatastoreContext context) {
 686         datastoreContext = verifyNotNull(context);
 687
 688         setTransactionCommitTimeout();
 689
 690         setPersistence(datastoreContext.isPersistent());
 691
 692         updateConfigParams(datastoreContext.getShardRaftConfig());
 693     }
 694
 695     // applyState() will be invoked once consensus is reached on the payload
 696     // non-final for mocking
 697     void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
 698         final boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
 699         if (canSkipPayload) {
 700             applyState(self(), id, payload);
 701         } else {
 702             // We are faking the sender
 703             persistData(self(), id, payload, batchHint);
 704         }
 705     }
 706
 707     @Deprecated(since = "9.0.0", forRemoval = true)
 708     private void handleCommitTransaction(final CommitTransaction commit) {
 709         final var txId = commit.getTransactionId();
 710         if (isLeader()) {
 711             askProtocolEncountered(txId);
 712             commitCoordinator.handleCommit(txId, getSender(), this);
 713         } else {
 714             final var leader = getLeader();
 715             if (leader == null) {
 716                 messageRetrySupport.addMessageToRetry(commit, getSender(), "Could not commit transaction " + txId);
 717             } else {
 718                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
 719                 leader.forward(commit, getContext());
 720             }
 721         }
 722     }
 723
 724     @Deprecated(since = "9.0.0", forRemoval = true)
 725     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
 726         final var txId = canCommit.getTransactionId();
 727         LOG.debug("{}: Can committing transaction {}", persistenceId(), txId);
 728
 729         if (isLeader()) {
 730             askProtocolEncountered(txId);
 731             commitCoordinator.handleCanCommit(txId, getSender(), this);
 732         } else {
 733             final var leader = getLeader();
 734             if (leader == null) {
 735                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
 736                         "Could not canCommit transaction " + txId);
 737             } else {
 738                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
 739                 leader.forward(canCommit, getContext());
 740             }
 741         }
 742     }
 743
 744     @SuppressWarnings("checkstyle:IllegalCatch")
 745     @Deprecated(since = "9.0.0", forRemoval = true)
 746     private void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
 747         askProtocolEncountered(batched.getTransactionId());
 748
 749         try {
 750             commitCoordinator.handleBatchedModifications(batched, sender, this);
 751         } catch (Exception e) {
 752             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
 753                     batched.getTransactionId(), e);
 754             sender.tell(new Failure(e), getSelf());
 755         }
 756     }
 757
 758     @Deprecated(since = "9.0.0", forRemoval = true)
 759     private void handleBatchedModifications(final BatchedModifications batched) {
 760         // This message is sent to prepare the modifications transaction directly on the Shard as an
 761         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
 762         // BatchedModifications message, the caller sets the ready flag in the message indicating
 763         // modifications are complete. The reply contains the cohort actor path (this actor) for the caller
 764         // to initiate the 3-phase commit. This also avoids the overhead of sending an additional
 765         // ReadyTransaction message.
 766
 767         // If we're not the leader then forward to the leader. This is a safety measure - we shouldn't
 768         // normally get here if we're not the leader as the front-end (TransactionProxy) should determine
 769         // the primary/leader shard. However with timing and caching on the front-end, there's a small
 770         // window where it could have a stale leader during leadership transitions.
 771         //
 772         boolean isLeaderActive = isLeaderActive();
 773         if (isLeader() && isLeaderActive) {
 774             handleBatchedModificationsLocal(batched, getSender());
 775         } else {
 776             final var leader = getLeader();
 777             if (!isLeaderActive || leader == null) {
 778                 messageRetrySupport.addMessageToRetry(batched, getSender(),
 779                         "Could not process BatchedModifications " + batched.getTransactionId());
 780             } else {
 781                 // If this is not the first batch and leadership changed in between batched messages,
 782                 // we need to reconstruct previous BatchedModifications from the transaction
 783                 // DataTreeModification, honoring the max batched modification count, and forward all the
 784                 // previous BatchedModifications to the new leader.
 785                 final var newModifications = commitCoordinator.createForwardedBatchedModifications(batched,
 786                     datastoreContext.getShardBatchedModificationCount());
 787
 788                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
 789                         newModifications.size(), leader);
 790
 791                 for (BatchedModifications bm : newModifications) {
 792                     leader.forward(bm, getContext());
 793                 }
 794             }
 795         }
 796     }
 797
 798     private boolean failIfIsolatedLeader(final ActorRef sender) {
 799         if (isIsolatedLeader()) {
 800             sender.tell(new Failure(new NoShardLeaderException(String.format(
 801                     "Shard %s was the leader but has lost contact with all of its followers. Either all"
 802                     + " other follower nodes are down or this node is isolated by a network partition.",
 803                     persistenceId()))), getSelf());
 804             return true;
 805         }
 806
 807         return false;
 808     }
 809
 810     protected boolean isIsolatedLeader() {
 811         return getRaftState() == RaftState.IsolatedLeader;
 812     }
 813
 814     @SuppressWarnings("checkstyle:IllegalCatch")
 815     @Deprecated(since = "9.0.0", forRemoval = true)
 816    private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
 817         final var txId = message.getTransactionId();
 818         LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), txId);
 819
 820         final var isLeaderActive = isLeaderActive();
 821         if (isLeader() && isLeaderActive) {
 822             askProtocolEncountered(txId);
 823             try {
 824                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
 825             } catch (Exception e) {
 826                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(), txId, e);
 827                 getSender().tell(new Failure(e), getSelf());
 828             }
 829         } else {
 830             final var leader = getLeader();
 831             if (!isLeaderActive || leader == null) {
 832                 messageRetrySupport.addMessageToRetry(message, getSender(),
 833                         "Could not process ready local transaction " + txId);
 834             } else {
 835                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
 836                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 837                 leader.forward(message, getContext());
 838             }
 839         }
 840     }
 841
 842     @Deprecated(since = "9.0.0", forRemoval = true)
 843     private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
 844         LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
 845
 846         final var isLeaderActive = isLeaderActive();
 847         if (isLeader() && isLeaderActive) {
 848             askProtocolEncountered(forwardedReady.getTransactionId());
 849             commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
 850         } else {
 851             final var leader = getLeader();
 852             if (!isLeaderActive || leader == null) {
 853                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
 854                         "Could not process forwarded ready transaction " + forwardedReady.getTransactionId());
 855             } else {
 856                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 857
 858                 final var readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
 859                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit(),
 860                         forwardedReady.getParticipatingShardNames());
 861                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
 862                 leader.forward(readyLocal, getContext());
 863             }
 864         }
 865     }
 866
 867     @Deprecated(since = "9.0.0", forRemoval = true)
 868     private void handleAbortTransaction(final AbortTransaction abort) {
 869         final var transactionId = abort.getTransactionId();
 870         askProtocolEncountered(transactionId);
 871         doAbortTransaction(transactionId, getSender());
 872     }
 873
 874     final void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
 875         commitCoordinator.handleAbort(transactionID, sender, this);
 876     }
 877
 878     @Deprecated(since = "9.0.0", forRemoval = true)
 879     private void handleCreateTransaction(final Object message) {
 880         if (isLeader()) {
 881             createTransaction(CreateTransaction.fromSerializable(message));
 882         } else if (getLeader() != null) {
 883             getLeader().forward(message, getContext());
 884         } else {
 885             getSender().tell(new Failure(new NoShardLeaderException(
 886                     "Could not create a shard transaction", persistenceId())), getSelf());
 887         }
 888     }
 889
 890     @Deprecated(since = "9.0.0", forRemoval = true)
 891     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
 892         if (isLeader()) {
 893             final var id = closeTransactionChain.getIdentifier();
 894             askProtocolEncountered(id.getClientId());
 895             store.closeTransactionChain(id);
 896         } else if (getLeader() != null) {
 897             getLeader().forward(closeTransactionChain, getContext());
 898         } else {
 899             LOG.warn("{}: Could not close transaction {}", persistenceId(), closeTransactionChain.getIdentifier());
 900         }
 901     }
 902
 903     @Deprecated(since = "9.0.0", forRemoval = true)
 904     @SuppressWarnings("checkstyle:IllegalCatch")
 905     private void createTransaction(final CreateTransaction createTransaction) {
 906         askProtocolEncountered(createTransaction.getTransactionId());
 907
 908         try {
 909             if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
 910                     && failIfIsolatedLeader(getSender())) {
 911                 return;
 912             }
 913
 914             final var transactionActor = createTransaction(createTransaction.getTransactionType(),
 915                 createTransaction.getTransactionId());
 916
 917             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
 918                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
 919         } catch (Exception e) {
 920             getSender().tell(new Failure(e), getSelf());
 921         }
 922     }
 923
 924     @Deprecated(since = "9.0.0", forRemoval = true)
 925     private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
 926         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
 927         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
 928             transactionId);
 929     }
 930
 931     // Called on leader only
 932     @Deprecated(since = "9.0.0", forRemoval = true)
 933     private void askProtocolEncountered(final TransactionIdentifier transactionId) {
 934         askProtocolEncountered(transactionId.getHistoryId().getClientId());
 935     }
 936
 937     // Called on leader only
 938     @Deprecated(since = "9.0.0", forRemoval = true)
 939     private void askProtocolEncountered(final ClientIdentifier clientId) {
 940         final var frontend = clientId.getFrontendId();
 941         final var state = knownFrontends.get(frontend);
 942         if (!(state instanceof LeaderFrontendState.Disabled)) {
 943             LOG.debug("{}: encountered ask-based client {}, disabling transaction tracking", persistenceId(), clientId);
 944             if (knownFrontends.isEmpty()) {
 945                 knownFrontends = new HashMap<>();
 946             }
 947             knownFrontends.put(frontend, new LeaderFrontendState.Disabled(persistenceId(), clientId, getDataStore()));
 948
 949             persistPayload(clientId, DisableTrackingPayload.create(clientId,
 950                 datastoreContext.getInitialPayloadSerializedBufferCapacity()), false);
 951         }
 952     }
 953
 954     private void updateSchemaContext(final UpdateSchemaContext message) {
 955         updateSchemaContext(message.getEffectiveModelContext());
 956     }
 957
 958     @VisibleForTesting
 959     void updateSchemaContext(final @NonNull EffectiveModelContext schemaContext) {
 960         store.updateSchemaContext(schemaContext);
 961     }
 962
 963     private boolean isMetricsCaptureEnabled() {
 964         CommonConfig config = new CommonConfig(getContext().system().settings().config());
 965         return config.isMetricCaptureEnabled();
 966     }
 967
 968     @Override
 969     protected final RaftActorSnapshotCohort getRaftActorSnapshotCohort() {
 970         return snapshotCohort;
 971     }
 972
 973     @Override
 974     protected final RaftActorRecoveryCohort getRaftActorRecoveryCohort() {
 975         if (restoreFromSnapshot == null) {
 976             return ShardRecoveryCoordinator.create(store, persistenceId(), LOG);
 977         }
 978
 979         return ShardRecoveryCoordinator.forSnapshot(store, persistenceId(), LOG, restoreFromSnapshot.getSnapshot());
 980     }
 981
 982     @Override
 983     // non-final for testing
 984     protected void onRecoveryComplete() {
 985         restoreFromSnapshot = null;
 986
 987         //notify shard manager
 988         getContext().parent().tell(new ActorInitialized(getSelf()), ActorRef.noSender());
 989
 990         // Being paranoid here - this method should only be called once but just in case...
 991         if (txCommitTimeoutCheckSchedule == null) {
 992             // Schedule a message to be periodically sent to check if the current in-progress
 993             // transaction should be expired and aborted.
 994             final var period = FiniteDuration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
 995             txCommitTimeoutCheckSchedule = getContext().system().scheduler().schedule(
 996                     period, period, getSelf(),
 997                     TX_COMMIT_TIMEOUT_CHECK_MESSAGE, getContext().dispatcher(), ActorRef.noSender());
 998         }
 999     }
1000
1001     @Override
1002     protected final void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
1003         if (data instanceof Payload payload) {
1004             if (payload instanceof DisableTrackingPayload disableTracking) {
1005                 disableTracking(disableTracking);
1006                 return;
1007             }
1008
1009             try {
1010                 store.applyReplicatedPayload(identifier, payload);
1011             } catch (DataValidationFailedException | IOException e) {
1012                 LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
1013             }
1014         } else {
1015             LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
1016         }
1017     }
1018
1019     @Override
1020     protected final void onStateChanged() {
1021         boolean isLeader = isLeader();
1022         boolean hasLeader = hasLeader();
1023         treeChangeSupport.onLeadershipChange(isLeader, hasLeader);
1024
1025         // If this actor is no longer the leader close all the transaction chains
1026         if (!isLeader) {
1027             if (LOG.isDebugEnabled()) {
1028                 LOG.debug(
1029                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
1030                     persistenceId(), getId());
1031             }
1032
1033             paused = false;
1034             store.purgeLeaderState();
1035         }
1036
1037         if (hasLeader && !isIsolatedLeader()) {
1038             messageRetrySupport.retryMessages();
1039         }
1040     }
1041
1042     @Override
1043     protected final void onLeaderChanged(final String oldLeader, final String newLeader) {
1044         shardMBean.incrementLeadershipChangeCount();
1045         paused = false;
1046
1047         if (!isLeader()) {
1048             if (!knownFrontends.isEmpty()) {
1049                 LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
1050                 knownFrontends = ImmutableMap.of();
1051             }
1052
1053             requestMessageAssembler.close();
1054
1055             if (!hasLeader()) {
1056                 // No leader anywhere, nothing else to do
1057                 return;
1058             }
1059
1060             // Another leader was elected. If we were the previous leader and had pending transactions, convert
1061             // them to transaction messages and send to the new leader.
1062             ActorSelection leader = getLeader();
1063             if (leader != null) {
1064                 // Clears all pending transactions and converts them to messages to be forwarded to a new leader.
1065                 Collection<?> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
1066                     datastoreContext.getShardBatchedModificationCount());
1067
1068                 if (!messagesToForward.isEmpty()) {
1069                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
1070                             messagesToForward.size(), leader);
1071
1072                     for (Object message : messagesToForward) {
1073                         LOG.debug("{}: Forwarding pending transaction message {}", persistenceId(), message);
1074
1075                         leader.tell(message, self());
1076                     }
1077                 }
1078             } else {
1079                 commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
1080                         + "change and the leader address isn't available.", this);
1081             }
1082         } else {
1083             // We have become the leader, we need to reconstruct frontend state
1084             knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1085             LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
1086         }
1087
1088         if (!isIsolatedLeader()) {
1089             messageRetrySupport.retryMessages();
1090         }
1091     }
1092
1093     @Override
1094     protected final void pauseLeader(final Runnable operation) {
1095         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
1096         paused = true;
1097
1098         // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
1099         knownFrontends.values().forEach(LeaderFrontendState::retire);
1100         knownFrontends = ImmutableMap.of();
1101
1102         store.setRunOnPendingTransactionsComplete(operation);
1103     }
1104
1105     @Override
1106     protected final void unpauseLeader() {
1107         LOG.debug("{}: In unpauseLeader", persistenceId());
1108         paused = false;
1109
1110         store.setRunOnPendingTransactionsComplete(null);
1111
1112         // Restore tell-based protocol state as if we were becoming the leader
1113         knownFrontends = verifyNotNull(frontendMetadata.toLeaderState(this));
1114     }
1115
1116     @Override
1117     protected final OnDemandRaftState.AbstractBuilder<?, ?> newOnDemandRaftStateBuilder() {
1118         return OnDemandShardState.newBuilder()
1119             .treeChangeListenerActors(treeChangeSupport.getListenerActors())
1120             .commitCohortActors(store.getCohortActors());
1121     }
1122
1123     @Override
1124     public final String persistenceId() {
1125         return name;
1126     }
1127
1128     @Override
1129     public final String journalPluginId() {
1130         // This method may be invoked from super constructor (wonderful), hence we also need to handle the case of
1131         // the field being uninitialized because our constructor is not finished.
1132         if (datastoreContext != null && !datastoreContext.isPersistent()) {
1133             return NON_PERSISTENT_JOURNAL_ID;
1134         }
1135         return super.journalPluginId();
1136     }
1137
1138     @VisibleForTesting
1139     final ShardCommitCoordinator getCommitCoordinator() {
1140         return commitCoordinator;
1141     }
1142
1143     // non-final for mocking
1144     DatastoreContext getDatastoreContext() {
1145         return datastoreContext;
1146     }
1147
1148     @VisibleForTesting
1149     final ShardDataTree getDataStore() {
1150         return store;
1151     }
1152
1153     @VisibleForTesting
1154     // non-final for mocking
1155     ShardStats getShardMBean() {
1156         return shardMBean;
1157     }
1158
1159     public static Builder builder() {
1160         return new Builder();
1161     }
1162
1163     public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
1164         private final Class<? extends S> shardClass;
1165         private ShardIdentifier id;
1166         private Map<String, String> peerAddresses = Collections.emptyMap();
1167         private DatastoreContext datastoreContext;
1168         private EffectiveModelContextProvider schemaContextProvider;
1169         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
1170         private DataTree dataTree;
1171
1172         private volatile boolean sealed;
1173
1174         AbstractBuilder(final Class<? extends S> shardClass) {
1175             this.shardClass = shardClass;
1176         }
1177
1178         final void checkSealed() {
1179             checkState(!sealed, "Builder is already sealed - further modifications are not allowed");
1180         }
1181
1182         @SuppressWarnings("unchecked")
1183         private T self() {
1184             return (T) this;
1185         }
1186
1187         public T id(final ShardIdentifier newId) {
1188             checkSealed();
1189             id = newId;
1190             return self();
1191         }
1192
1193         public T peerAddresses(final Map<String, String> newPeerAddresses) {
1194             checkSealed();
1195             peerAddresses = newPeerAddresses;
1196             return self();
1197         }
1198
1199         public T datastoreContext(final DatastoreContext newDatastoreContext) {
1200             checkSealed();
1201             datastoreContext = newDatastoreContext;
1202             return self();
1203         }
1204
1205         public T schemaContextProvider(final EffectiveModelContextProvider newSchemaContextProvider) {
1206             checkSealed();
1207             schemaContextProvider = requireNonNull(newSchemaContextProvider);
1208             return self();
1209         }
1210
1211         public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
1212             checkSealed();
1213             restoreFromSnapshot = newRestoreFromSnapshot;
1214             return self();
1215         }
1216
1217         public T dataTree(final DataTree newDataTree) {
1218             checkSealed();
1219             dataTree = newDataTree;
1220             return self();
1221         }
1222
1223         public ShardIdentifier getId() {
1224             return id;
1225         }
1226
1227         public Map<String, String> getPeerAddresses() {
1228             return peerAddresses;
1229         }
1230
1231         public DatastoreContext getDatastoreContext() {
1232             return datastoreContext;
1233         }
1234
1235         public EffectiveModelContext getSchemaContext() {
1236             return verifyNotNull(schemaContextProvider.getEffectiveModelContext());
1237         }
1238
1239         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
1240             return restoreFromSnapshot;
1241         }
1242
1243         public DataTree getDataTree() {
1244             return dataTree;
1245         }
1246
1247         public TreeType getTreeType() {
1248             return switch (datastoreContext.getLogicalStoreType()) {
1249                 case CONFIGURATION -> TreeType.CONFIGURATION;
1250                 case OPERATIONAL -> TreeType.OPERATIONAL;
1251             };
1252         }
1253
1254         protected void verify() {
1255             requireNonNull(id, "id should not be null");
1256             requireNonNull(peerAddresses, "peerAddresses should not be null");
1257             requireNonNull(datastoreContext, "dataStoreContext should not be null");
1258             requireNonNull(schemaContextProvider, "schemaContextProvider should not be null");
1259         }
1260
1261         public Props props() {
1262             sealed = true;
1263             verify();
1264             return Props.create(shardClass, this);
1265         }
1266     }
1267
1268     public static class Builder extends AbstractBuilder<Builder, Shard> {
1269         Builder() {
1270             this(Shard.class);
1271         }
1272
1273         Builder(final Class<? extends Shard> shardClass) {
1274             super(shardClass);
1275         }
1276     }
1277
1278     Ticker ticker() {
1279         return Ticker.systemTicker();
1280     }
1281
1282     void scheduleNextPendingTransaction() {
1283         self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
1284     }
1285 }