Slice front-end request messages
[controller.git] / opendaylight / md-sal / sal-distributed-datastore / src / main / java / org / opendaylight / controller / cluster / datastore / Shard.java
index bd44f0b41d3ba30006f750b35199a2dcd58d7f44..ebe5498481d9fe3befdabafbb4976f635a7b15b3 100644 (file)
@@ -11,21 +11,50 @@ package org.opendaylight.controller.cluster.datastore;
 import akka.actor.ActorRef;
 import akka.actor.ActorSelection;
 import akka.actor.Cancellable;
+import akka.actor.ExtendedActorSystem;
 import akka.actor.Props;
+import akka.actor.Status;
+import akka.actor.Status.Failure;
+import akka.serialization.JavaSerializer;
 import akka.serialization.Serialization;
 import com.google.common.annotations.VisibleForTesting;
 import com.google.common.base.Optional;
 import com.google.common.base.Preconditions;
-import com.google.common.base.Throwables;
+import com.google.common.base.Ticker;
+import com.google.common.base.Verify;
+import com.google.common.collect.ImmutableList;
+import com.google.common.collect.ImmutableMap;
+import com.google.common.collect.Range;
 import java.io.IOException;
+import java.util.Arrays;
 import java.util.Collection;
 import java.util.Collections;
 import java.util.Map;
-import java.util.concurrent.ExecutionException;
 import java.util.concurrent.TimeUnit;
 import javax.annotation.Nonnull;
+import javax.annotation.Nullable;
+import org.opendaylight.controller.cluster.access.ABIVersion;
+import org.opendaylight.controller.cluster.access.commands.ConnectClientRequest;
+import org.opendaylight.controller.cluster.access.commands.ConnectClientSuccess;
+import org.opendaylight.controller.cluster.access.commands.LocalHistoryRequest;
+import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
+import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
+import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
+import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
+import org.opendaylight.controller.cluster.access.concepts.FrontendIdentifier;
+import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
+import org.opendaylight.controller.cluster.access.concepts.Request;
+import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
+import org.opendaylight.controller.cluster.access.concepts.RequestException;
+import org.opendaylight.controller.cluster.access.concepts.RequestSuccess;
+import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
+import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
+import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
+import org.opendaylight.controller.cluster.access.concepts.UnsupportedRequestException;
 import org.opendaylight.controller.cluster.common.actor.CommonConfig;
+import org.opendaylight.controller.cluster.common.actor.Dispatchers;
+import org.opendaylight.controller.cluster.common.actor.Dispatchers.DispatcherType;
 import org.opendaylight.controller.cluster.common.actor.MessageTracker;
 import org.opendaylight.controller.cluster.common.actor.MessageTracker.Error;
 import org.opendaylight.controller.cluster.common.actor.MeteringBehavior;
@@ -39,47 +68,53 @@ import org.opendaylight.controller.cluster.datastore.messages.BatchedModificatio
 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
 import org.opendaylight.controller.cluster.datastore.messages.CloseTransactionChain;
 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
-import org.opendaylight.controller.cluster.datastore.messages.CommitTransactionReply;
 import org.opendaylight.controller.cluster.datastore.messages.CreateTransaction;
 import org.opendaylight.controller.cluster.datastore.messages.CreateTransactionReply;
-import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot;
-import org.opendaylight.controller.cluster.datastore.messages.DatastoreSnapshot.ShardSnapshot;
 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
 import org.opendaylight.controller.cluster.datastore.messages.GetShardDataTree;
+import org.opendaylight.controller.cluster.datastore.messages.MakeLeaderLocal;
+import org.opendaylight.controller.cluster.datastore.messages.OnDemandShardState;
 import org.opendaylight.controller.cluster.datastore.messages.PeerAddressResolved;
+import org.opendaylight.controller.cluster.datastore.messages.PersistAbortTransactionPayload;
 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
 import org.opendaylight.controller.cluster.datastore.messages.RegisterChangeListener;
 import org.opendaylight.controller.cluster.datastore.messages.RegisterDataTreeChangeListener;
 import org.opendaylight.controller.cluster.datastore.messages.ShardLeaderStateChanged;
 import org.opendaylight.controller.cluster.datastore.messages.UpdateSchemaContext;
-import org.opendaylight.controller.cluster.datastore.persisted.CommitTransactionPayload;
-import org.opendaylight.controller.cluster.datastore.persisted.DataTreeCandidateSupplier;
-import org.opendaylight.controller.cluster.datastore.utils.Dispatchers;
+import org.opendaylight.controller.cluster.datastore.persisted.AbortTransactionPayload;
+import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot;
+import org.opendaylight.controller.cluster.datastore.persisted.DatastoreSnapshot.ShardSnapshot;
+import org.opendaylight.controller.cluster.messaging.MessageAssembler;
+import org.opendaylight.controller.cluster.messaging.MessageSlicer;
+import org.opendaylight.controller.cluster.messaging.SliceOptions;
 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
 import org.opendaylight.controller.cluster.notifications.RegisterRoleChangeListener;
 import org.opendaylight.controller.cluster.notifications.RoleChangeNotifier;
+import org.opendaylight.controller.cluster.raft.LeadershipTransferFailedException;
 import org.opendaylight.controller.cluster.raft.RaftActor;
 import org.opendaylight.controller.cluster.raft.RaftActorRecoveryCohort;
 import org.opendaylight.controller.cluster.raft.RaftActorSnapshotCohort;
 import org.opendaylight.controller.cluster.raft.RaftState;
 import org.opendaylight.controller.cluster.raft.base.messages.FollowerInitialSyncUpStatus;
+import org.opendaylight.controller.cluster.raft.client.messages.OnDemandRaftState;
 import org.opendaylight.controller.cluster.raft.messages.AppendEntriesReply;
+import org.opendaylight.controller.cluster.raft.messages.RequestLeadership;
 import org.opendaylight.controller.cluster.raft.messages.ServerRemoved;
 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
 import org.opendaylight.yangtools.concepts.Identifier;
-import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidate;
 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
-import org.opendaylight.yangtools.yang.data.api.schema.tree.ModificationType;
+import org.opendaylight.yangtools.yang.data.api.schema.tree.TipProducingDataTree;
 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
+import org.opendaylight.yangtools.yang.model.api.SchemaContextProvider;
 import scala.concurrent.duration.Duration;
 import scala.concurrent.duration.FiniteDuration;
 
 /**
- * A Shard represents a portion of the logical data tree <br/>
+ * A Shard represents a portion of the logical data tree.
+ *
  * <p>
  * Our Shard uses InMemoryDataTree as it's internal representation and delegates all requests it
- * </p>
  */
 public class Shard extends RaftActor {
 
@@ -99,9 +134,27 @@ public class Shard extends RaftActor {
         }
     };
 
+    static final Object RESUME_NEXT_PENDING_TRANSACTION = new Object() {
+        @Override
+        public String toString() {
+            return "resumeNextPendingTransaction";
+        }
+    };
+
     // FIXME: shard names should be encapsulated in their own class and this should be exposed as a constant.
     public static final String DEFAULT_NAME = "default";
 
+    private static final Collection<ABIVersion> SUPPORTED_ABIVERSIONS;
+
+    static {
+        final ABIVersion[] values = ABIVersion.values();
+        final ABIVersion[] real = Arrays.copyOfRange(values, 1, values.length - 1);
+        SUPPORTED_ABIVERSIONS = ImmutableList.copyOf(real).reverse();
+    }
+
+    // FIXME: make this a dynamic property based on mailbox size and maximum number of clients
+    private static final int CLIENT_MAX_MESSAGES = 1000;
+
     // The state of this Shard
     private final ShardDataTree store;
 
@@ -134,50 +187,76 @@ public class Shard extends RaftActor {
 
     private final ShardTransactionMessageRetrySupport messageRetrySupport;
 
-    protected Shard(AbstractBuilder<?, ?> builder) {
+    private final FrontendMetadata frontendMetadata;
+    private Map<FrontendIdentifier, LeaderFrontendState> knownFrontends = ImmutableMap.of();
+    private boolean paused;
+
+    private final MessageSlicer responseMessageSlicer;
+    private final Dispatchers dispatchers;
+
+    private final MessageAssembler requestMessageAssembler;
+
+    protected Shard(final AbstractBuilder<?, ?> builder) {
         super(builder.getId().toString(), builder.getPeerAddresses(),
                 Optional.of(builder.getDatastoreContext().getShardRaftConfig()), DataStoreVersions.CURRENT_VERSION);
 
         this.name = builder.getId().toString();
         this.datastoreContext = builder.getDatastoreContext();
         this.restoreFromSnapshot = builder.getRestoreFromSnapshot();
+        this.frontendMetadata = new FrontendMetadata(name);
 
         setPersistence(datastoreContext.isPersistent());
 
         LOG.info("Shard created : {}, persistent : {}", name, datastoreContext.isPersistent());
 
-        store = new ShardDataTree(builder.getSchemaContext(), builder.getTreeType(),
-                new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher"),
-                new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher"), name);
+        ShardDataTreeChangeListenerPublisherActorProxy treeChangeListenerPublisher =
+                new ShardDataTreeChangeListenerPublisherActorProxy(getContext(), name + "-DTCL-publisher", name);
+        ShardDataChangeListenerPublisherActorProxy dataChangeListenerPublisher =
+                new ShardDataChangeListenerPublisherActorProxy(getContext(), name + "-DCL-publisher", name);
+        if (builder.getDataTree() != null) {
+            store = new ShardDataTree(this, builder.getSchemaContext(), builder.getDataTree(),
+                    treeChangeListenerPublisher, dataChangeListenerPublisher, name, frontendMetadata);
+        } else {
+            store = new ShardDataTree(this, builder.getSchemaContext(), builder.getTreeType(),
+                    builder.getDatastoreContext().getStoreRoot(), treeChangeListenerPublisher,
+                    dataChangeListenerPublisher, name, frontendMetadata);
+        }
 
-        shardMBean = ShardMBeanFactory.getShardStatsMBean(name.toString(),
-                datastoreContext.getDataStoreMXBeanType());
-        shardMBean.setShard(this);
+        shardMBean = ShardMBeanFactory.getShardStatsMBean(name, datastoreContext.getDataStoreMXBeanType(), this);
 
         if (isMetricsCaptureEnabled()) {
             getContext().become(new MeteringBehavior(this));
         }
 
-        commitCoordinator = new ShardCommitCoordinator(store,
-                datastoreContext.getShardCommitQueueExpiryTimeoutInMillis(),
-                datastoreContext.getShardTransactionCommitQueueCapacity(), LOG, this.name);
+        commitCoordinator = new ShardCommitCoordinator(store, LOG, this.name);
 
         setTransactionCommitTimeout();
 
         // create a notifier actor for each cluster member
-        roleChangeNotifier = createRoleChangeNotifier(name.toString());
+        roleChangeNotifier = createRoleChangeNotifier(name);
 
         appendEntriesReplyTracker = new MessageTracker(AppendEntriesReply.class,
                 getRaftActorContext().getConfigParams().getIsolatedCheckIntervalInMillis());
 
+        dispatchers = new Dispatchers(context().system().dispatchers());
         transactionActorFactory = new ShardTransactionActorFactory(store, datastoreContext,
-                new Dispatchers(context().system().dispatchers()).getDispatcherPath(
-                        Dispatchers.DispatcherType.Transaction), self(), getContext(), shardMBean);
+            dispatchers.getDispatcherPath(Dispatchers.DispatcherType.Transaction),
+                self(), getContext(), shardMBean, builder.getId().getShardName());
 
         snapshotCohort = ShardSnapshotCohort.create(getContext(), builder.getId().getMemberName(), store, LOG,
             this.name);
 
         messageRetrySupport = new ShardTransactionMessageRetrySupport(this);
+
+        responseMessageSlicer = MessageSlicer.builder().logContext(this.name)
+                .messageSliceSize(datastoreContext.getMaximumMessageSliceSize())
+                .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
+                .expireStateAfterInactivity(2, TimeUnit.MINUTES).build();
+
+        requestMessageAssembler = MessageAssembler.builder().logContext(this.name)
+                .fileBackedStreamFactory(getRaftActorContext().getFileBackedOutputStreamFactory())
+                .assembledMessageCallback((message, sender) -> self().tell(message, sender))
+                .expireStateAfterInactivity(datastoreContext.getRequestTimeout(), TimeUnit.NANOSECONDS).build();
     }
 
     private void setTransactionCommitTimeout() {
@@ -185,7 +264,7 @@ public class Shard extends RaftActor {
                 datastoreContext.getShardTransactionCommitTimeoutInSeconds(), TimeUnit.SECONDS) / 2;
     }
 
-    private Optional<ActorRef> createRoleChangeNotifier(String shardId) {
+    private Optional<ActorRef> createRoleChangeNotifier(final String shardId) {
         ActorRef shardRoleChangeNotifier = this.getContext().actorOf(
             RoleChangeNotifier.getProps(shardId), shardId + "-notifier");
         return Optional.of(shardRoleChangeNotifier);
@@ -199,7 +278,7 @@ public class Shard extends RaftActor {
 
         messageRetrySupport.close();
 
-        if(txCommitTimeoutCheckSchedule != null) {
+        if (txCommitTimeoutCheckSchedule != null) {
             txCommitTimeoutCheckSchedule.cancel();
         }
 
@@ -221,14 +300,22 @@ public class Shard extends RaftActor {
 
     @Override
     protected void handleNonRaftCommand(final Object message) {
-        try (final MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
+        try (MessageTracker.Context context = appendEntriesReplyTracker.received(message)) {
             final Optional<Error> maybeError = context.error();
             if (maybeError.isPresent()) {
                 LOG.trace("{} : AppendEntriesReply failed to arrive at the expected interval {}", persistenceId(),
                     maybeError.get());
             }
 
-            if (CreateTransaction.isSerializedType(message)) {
+            store.resetTransactionBatch();
+
+            if (message instanceof RequestEnvelope) {
+                handleRequestEnvelope((RequestEnvelope)message);
+            } else if (requestMessageAssembler.isHandledMessage(message)) {
+                handleRequestAssemblerMessage(message);
+            } else if (message instanceof ConnectClientRequest) {
+                handleConnectClient((ConnectClientRequest)message);
+            } else if (CreateTransaction.isSerializedType(message)) {
                 handleCreateTransaction(message);
             } else if (message instanceof BatchedModifications) {
                 handleBatchedModifications((BatchedModifications)message);
@@ -252,40 +339,240 @@ public class Shard extends RaftActor {
                 updateSchemaContext((UpdateSchemaContext) message);
             } else if (message instanceof PeerAddressResolved) {
                 PeerAddressResolved resolved = (PeerAddressResolved) message;
-                setPeerAddress(resolved.getPeerId().toString(),
-                        resolved.getPeerAddress());
+                setPeerAddress(resolved.getPeerId(), resolved.getPeerAddress());
             } else if (TX_COMMIT_TIMEOUT_CHECK_MESSAGE.equals(message)) {
-                commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
-            } else if(message instanceof DatastoreContext) {
+                commitTimeoutCheck();
+            } else if (message instanceof DatastoreContext) {
                 onDatastoreContext((DatastoreContext)message);
-            } else if(message instanceof RegisterRoleChangeListener){
+            } else if (message instanceof RegisterRoleChangeListener) {
                 roleChangeNotifier.get().forward(message, context());
             } else if (message instanceof FollowerInitialSyncUpStatus) {
                 shardMBean.setFollowerInitialSyncStatus(((FollowerInitialSyncUpStatus) message).isInitialSyncDone());
                 context().parent().tell(message, self());
-            } else if(GET_SHARD_MBEAN_MESSAGE.equals(message)){
+            } else if (GET_SHARD_MBEAN_MESSAGE.equals(message)) {
                 sender().tell(getShardMBean(), self());
-            } else if(message instanceof GetShardDataTree) {
+            } else if (message instanceof GetShardDataTree) {
                 sender().tell(store.getDataTree(), self());
-            } else if(message instanceof ServerRemoved){
+            } else if (message instanceof ServerRemoved) {
                 context().parent().forward(message, context());
-            } else if(ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
+            } else if (ShardTransactionMessageRetrySupport.TIMER_MESSAGE_CLASS.isInstance(message)) {
                 messageRetrySupport.onTimerMessage(message);
             } else if (message instanceof DataTreeCohortActorRegistry.CohortRegistryCommand) {
-                commitCoordinator.processCohortRegistryCommand(getSender(),
+                store.processCohortRegistryCommand(getSender(),
                         (DataTreeCohortActorRegistry.CohortRegistryCommand) message);
-            } else {
+            } else if (message instanceof PersistAbortTransactionPayload) {
+                final TransactionIdentifier txId = ((PersistAbortTransactionPayload) message).getTransactionId();
+                persistPayload(txId, AbortTransactionPayload.create(txId), true);
+            } else if (message instanceof MakeLeaderLocal) {
+                onMakeLeaderLocal();
+            } else if (RESUME_NEXT_PENDING_TRANSACTION.equals(message)) {
+                store.resumeNextPendingTransaction();
+            } else if (!responseMessageSlicer.handleMessage(message)) {
                 super.handleNonRaftCommand(message);
             }
         }
     }
 
+    private void handleRequestAssemblerMessage(Object message) {
+        dispatchers.getDispatcher(DispatcherType.Serialization).execute(() -> {
+            JavaSerializer.currentSystem().value_$eq((ExtendedActorSystem) context().system());
+            requestMessageAssembler.handleMessage(message, self());
+        });
+    }
+
+    @SuppressWarnings("checkstyle:IllegalCatch")
+    private void handleRequestEnvelope(final RequestEnvelope envelope) {
+        final long now = ticker().read();
+        try {
+            final RequestSuccess<?, ?> success = handleRequest(envelope, now);
+            if (success != null) {
+                final long executionTimeNanos = ticker().read() - now;
+                if (success instanceof SliceableMessage) {
+                    dispatchers.getDispatcher(DispatcherType.Serialization).execute(() ->
+                        responseMessageSlicer.slice(SliceOptions.builder().identifier(success.getTarget())
+                            .message(envelope.newSuccessEnvelope(success, executionTimeNanos))
+                            .sendTo(envelope.getMessage().getReplyTo()).replyTo(self())
+                            .onFailureCallback(t -> {
+                                LOG.warn("Error slicing response {}", success, t);
+                            }).build()));
+                } else {
+                    envelope.sendSuccess(success, executionTimeNanos);
+                }
+            }
+        } catch (RequestException e) {
+            LOG.debug("{}: request {} failed", persistenceId(), envelope, e);
+            envelope.sendFailure(e, ticker().read() - now);
+        } catch (Exception e) {
+            LOG.debug("{}: request {} caused failure", persistenceId(), envelope, e);
+            envelope.sendFailure(new RuntimeRequestException("Request failed to process", e),
+                ticker().read() - now);
+        }
+    }
+
+    private void commitTimeoutCheck() {
+        store.checkForExpiredTransactions(transactionCommitTimeout, this::updateAccess);
+        commitCoordinator.checkForExpiredTransactions(transactionCommitTimeout, this);
+        requestMessageAssembler.checkExpiredAssembledMessageState();
+    }
+
+    private Optional<Long> updateAccess(final SimpleShardDataTreeCohort cohort) {
+        final FrontendIdentifier frontend = cohort.getIdentifier().getHistoryId().getClientId().getFrontendId();
+        final LeaderFrontendState state = knownFrontends.get(frontend);
+        if (state == null) {
+            // Not tell-based protocol, do nothing
+            return Optional.absent();
+        }
+
+        if (isIsolatedLeader()) {
+            // We are isolated and no new request can come through until we emerge from it. We are still updating
+            // liveness of frontend when we see it attempting to communicate. Use the last access timer.
+            return Optional.of(state.getLastSeenTicks());
+        }
+
+        // If this frontend has freshly connected, give it some time to catch up before killing its transactions.
+        return Optional.of(state.getLastConnectTicks());
+    }
+
+    private void onMakeLeaderLocal() {
+        LOG.debug("{}: onMakeLeaderLocal received", persistenceId());
+        if (isLeader()) {
+            getSender().tell(new Status.Success(null), getSelf());
+            return;
+        }
+
+        final ActorSelection leader = getLeader();
+
+        if (leader == null) {
+            // Leader is not present. The cluster is most likely trying to
+            // elect a leader and we should let that run its normal course
+
+            // TODO we can wait for the election to complete and retry the
+            // request. We can also let the caller retry by sending a flag
+            // in the response indicating the request is "reTryable".
+            getSender().tell(new Failure(
+                    new LeadershipTransferFailedException("We cannot initiate leadership transfer to local node. "
+                            + "Currently there is no leader for " + persistenceId())),
+                    getSelf());
+            return;
+        }
+
+        leader.tell(new RequestLeadership(getId(), getSender()), getSelf());
+    }
+
+    // Acquire our frontend tracking handle and verify generation matches
+    @Nullable
+    private LeaderFrontendState findFrontend(final ClientIdentifier clientId) throws RequestException {
+        final LeaderFrontendState existing = knownFrontends.get(clientId.getFrontendId());
+        if (existing != null) {
+            final int cmp = Long.compareUnsigned(existing.getIdentifier().getGeneration(), clientId.getGeneration());
+            if (cmp == 0) {
+                existing.touch();
+                return existing;
+            }
+            if (cmp > 0) {
+                LOG.debug("{}: rejecting request from outdated client {}", persistenceId(), clientId);
+                throw new RetiredGenerationException(existing.getIdentifier().getGeneration());
+            }
+
+            LOG.info("{}: retiring state {}, outdated by request from client {}", persistenceId(), existing, clientId);
+            existing.retire();
+            knownFrontends.remove(clientId.getFrontendId());
+        } else {
+            LOG.debug("{}: client {} is not yet known", persistenceId(), clientId);
+        }
+
+        return null;
+    }
+
+    private LeaderFrontendState getFrontend(final ClientIdentifier clientId) throws RequestException {
+        final LeaderFrontendState ret = findFrontend(clientId);
+        if (ret != null) {
+            return ret;
+        }
+
+        // TODO: a dedicated exception would be better, but this is technically true, too
+        throw new OutOfSequenceEnvelopeException(0);
+    }
+
+    private static @Nonnull ABIVersion selectVersion(final ConnectClientRequest message) {
+        final Range<ABIVersion> clientRange = Range.closed(message.getMinVersion(), message.getMaxVersion());
+        for (ABIVersion v : SUPPORTED_ABIVERSIONS) {
+            if (clientRange.contains(v)) {
+                return v;
+            }
+        }
+
+        throw new IllegalArgumentException(String.format(
+            "No common version between backend versions %s and client versions %s", SUPPORTED_ABIVERSIONS,
+            clientRange));
+    }
+
+    @SuppressWarnings("checkstyle:IllegalCatch")
+    private void handleConnectClient(final ConnectClientRequest message) {
+        try {
+            final ClientIdentifier clientId = message.getTarget();
+            final LeaderFrontendState existing = findFrontend(clientId);
+            if (existing != null) {
+                existing.touch();
+            }
+
+            if (!isLeader() || !isLeaderActive()) {
+                LOG.info("{}: not currently leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
+                                + "isLeadershipTransferInProgress: {}.",
+                        persistenceId(), message, isLeader(), isLeaderActive(), isLeadershipTransferInProgress());
+                throw new NotLeaderException(getSelf());
+            }
+
+            final ABIVersion selectedVersion = selectVersion(message);
+            final LeaderFrontendState frontend;
+            if (existing == null) {
+                frontend = new LeaderFrontendState(persistenceId(), clientId, store);
+                knownFrontends.put(clientId.getFrontendId(), frontend);
+                LOG.debug("{}: created state {} for client {}", persistenceId(), frontend, clientId);
+            } else {
+                frontend = existing;
+            }
+
+            frontend.reconnect();
+            message.getReplyTo().tell(new ConnectClientSuccess(message.getTarget(), message.getSequence(), getSelf(),
+                ImmutableList.of(), store.getDataTree(), CLIENT_MAX_MESSAGES).toVersion(selectedVersion),
+                ActorRef.noSender());
+        } catch (RequestException | RuntimeException e) {
+            message.getReplyTo().tell(new Failure(e), ActorRef.noSender());
+        }
+    }
+
+    private @Nullable RequestSuccess<?, ?> handleRequest(final RequestEnvelope envelope, final long now)
+            throws RequestException {
+        // We are not the leader, hence we want to fail-fast.
+        if (!isLeader() || paused || !isLeaderActive()) {
+            LOG.debug("{}: not currently active leader, rejecting request {}. isLeader: {}, isLeaderActive: {},"
+                            + "isLeadershipTransferInProgress: {}, paused: {}",
+                    persistenceId(), envelope, isLeader(), isLeaderActive(), isLeadershipTransferInProgress(), paused);
+            throw new NotLeaderException(getSelf());
+        }
+
+        final Request<?, ?> request = envelope.getMessage();
+        if (request instanceof TransactionRequest) {
+            final TransactionRequest<?> txReq = (TransactionRequest<?>)request;
+            final ClientIdentifier clientId = txReq.getTarget().getHistoryId().getClientId();
+            return getFrontend(clientId).handleTransactionRequest(txReq, envelope, now);
+        } else if (request instanceof LocalHistoryRequest) {
+            final LocalHistoryRequest<?> lhReq = (LocalHistoryRequest<?>)request;
+            final ClientIdentifier clientId = lhReq.getTarget().getClientId();
+            return getFrontend(clientId).handleLocalHistoryRequest(lhReq, envelope, now);
+        } else {
+            LOG.warn("{}: rejecting unsupported request {}", persistenceId(), request);
+            throw new UnsupportedRequestException(request);
+        }
+    }
+
     private boolean hasLeader() {
         return getLeaderId() != null;
     }
 
     public int getPendingTxCommitQueueSize() {
-        return commitCoordinator.getQueueSize();
+        return store.getQueueSize();
     }
 
     public int getCohortCacheSize() {
@@ -298,66 +585,41 @@ public class Shard extends RaftActor {
     }
 
     @Override
-    protected LeaderStateChanged newLeaderStateChanged(String memberId, String leaderId, short leaderPayloadVersion) {
+    protected LeaderStateChanged newLeaderStateChanged(final String memberId, final String leaderId,
+            final short leaderPayloadVersion) {
         return isLeader() ? new ShardLeaderStateChanged(memberId, leaderId, store.getDataTree(), leaderPayloadVersion)
                 : new ShardLeaderStateChanged(memberId, leaderId, leaderPayloadVersion);
     }
 
-    protected void onDatastoreContext(DatastoreContext context) {
+    protected void onDatastoreContext(final DatastoreContext context) {
         datastoreContext = context;
 
-        commitCoordinator.setQueueCapacity(datastoreContext.getShardTransactionCommitQueueCapacity());
-
         setTransactionCommitTimeout();
 
-        if(datastoreContext.isPersistent() && !persistence().isRecoveryApplicable()) {
-            setPersistence(true);
-        } else if(!datastoreContext.isPersistent() && persistence().isRecoveryApplicable()) {
-            setPersistence(false);
-        }
+        setPersistence(datastoreContext.isPersistent());
 
         updateConfigParams(datastoreContext.getShardRaftConfig());
     }
 
-    private static boolean isEmptyCommit(final DataTreeCandidate candidate) {
-        return ModificationType.UNMODIFIED.equals(candidate.getRootNode().getModificationType());
-    }
-
-    void continueCommit(final CohortEntry cohortEntry) {
-        final DataTreeCandidate candidate = cohortEntry.getCandidate();
-        final TransactionIdentifier transactionId = cohortEntry.getTransactionID();
-
-        // If we do not have any followers and we are not using persistence
-        // or if cohortEntry has no modifications
-        // we can apply modification to the state immediately
-        if ((!hasFollowers() && !persistence().isRecoveryApplicable()) || isEmptyCommit(candidate)) {
-            applyModificationToState(cohortEntry.getReplySender(), transactionId, candidate);
-            return;
-        }
-
-        final Payload payload;
-        try {
-            payload = CommitTransactionPayload.create(transactionId, candidate);
-        } catch (IOException e) {
-            LOG.error("{}: failed to encode transaction {} candidate {}", persistenceId(), transactionId, candidate,
-                e);
-            // TODO: do we need to do something smarter here?
-            throw Throwables.propagate(e);
+    // applyState() will be invoked once consensus is reached on the payload
+    void persistPayload(final Identifier id, final Payload payload, final boolean batchHint) {
+        boolean canSkipPayload = !hasFollowers() && !persistence().isRecoveryApplicable();
+        if (canSkipPayload) {
+            applyState(self(), id, payload);
+        } else {
+            // We are faking the sender
+            persistData(self(), id, payload, batchHint);
         }
-
-        persistData(cohortEntry.getReplySender(), cohortEntry.getTransactionID(), payload);
     }
 
     private void handleCommitTransaction(final CommitTransaction commit) {
         if (isLeader()) {
-            if(!commitCoordinator.handleCommit(commit.getTransactionID(), getSender(), this)) {
-                shardMBean.incrementFailedTransactionsCount();
-            }
+            commitCoordinator.handleCommit(commit.getTransactionId(), getSender(), this);
         } else {
             ActorSelection leader = getLeader();
             if (leader == null) {
                 messageRetrySupport.addMessageToRetry(commit, getSender(),
-                        "Could not commit transaction " + commit.getTransactionID());
+                        "Could not commit transaction " + commit.getTransactionId());
             } else {
                 LOG.debug("{}: Forwarding CommitTransaction to leader {}", persistenceId(), leader);
                 leader.forward(commit, getContext());
@@ -365,96 +627,16 @@ public class Shard extends RaftActor {
         }
     }
 
-    private void finishCommit(@Nonnull final ActorRef sender, @Nonnull final Identifier transactionID,
-            @Nonnull final CohortEntry cohortEntry) {
-        LOG.debug("{}: Finishing commit for transaction {}", persistenceId(), cohortEntry.getTransactionID());
-
-        try {
-            try {
-            cohortEntry.commit();
-            } catch(ExecutionException e) {
-                // We may get a "store tree and candidate base differ" IllegalStateException from commit under
-                // certain edge case scenarios so we'll try to re-apply the candidate from scratch as a last
-                // resort. Eg, we're a follower and a tx payload is replicated but the leader goes down before
-                // applying it to the state. We then become the leader and a second tx is pre-committed and
-                // replicated. When consensus occurs, this will cause the first tx to be applied as a foreign
-                // candidate via applyState prior to the second tx. Since the second tx has already been
-                // pre-committed, when it gets here to commit it will get an IllegalStateException.
-
-                // FIXME - this is not an ideal way to handle this scenario. This is temporary - a cleaner
-                // solution will be forthcoming.
-                if(e.getCause() instanceof IllegalStateException) {
-                    LOG.debug("{}: commit failed for transaction {} - retrying as foreign candidate", persistenceId(),
-                            transactionID, e);
-                    store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
-                } else {
-                    throw e;
-                }
-            }
-
-            sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), getSelf());
-
-            shardMBean.incrementCommittedTransactionCount();
-            shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
-
-        } catch (Exception e) {
-            sender.tell(new akka.actor.Status.Failure(e), getSelf());
-
-            LOG.error("{}, An exception occurred while committing transaction {}", persistenceId(),
-                    transactionID, e);
-            shardMBean.incrementFailedTransactionsCount();
-        } finally {
-            commitCoordinator.currentTransactionComplete(transactionID, true);
-        }
-    }
-
-    private void finishCommit(@Nonnull final ActorRef sender, final @Nonnull Identifier transactionID) {
-        // With persistence enabled, this method is called via applyState by the leader strategy
-        // after the commit has been replicated to a majority of the followers.
-
-        CohortEntry cohortEntry = commitCoordinator.getCohortEntryIfCurrent(transactionID);
-        if (cohortEntry == null) {
-            // The transaction is no longer the current commit. This can happen if the transaction
-            // was aborted prior, most likely due to timeout in the front-end. We need to finish
-            // committing the transaction though since it was successfully persisted and replicated
-            // however we can't use the original cohort b/c it was already preCommitted and may
-            // conflict with the current commit or may have been aborted so we commit with a new
-            // transaction.
-            cohortEntry = commitCoordinator.getAndRemoveCohortEntry(transactionID);
-            if(cohortEntry != null) {
-                try {
-                    store.applyForeignCandidate(transactionID, cohortEntry.getCandidate());
-                } catch (DataValidationFailedException e) {
-                    shardMBean.incrementFailedTransactionsCount();
-                    LOG.error("{}: Failed to re-apply transaction {}", persistenceId(), transactionID, e);
-                }
-
-                sender.tell(CommitTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(),
-                        getSelf());
-            } else {
-                // This really shouldn't happen - it likely means that persistence or replication
-                // took so long to complete such that the cohort entry was expired from the cache.
-                IllegalStateException ex = new IllegalStateException(
-                        String.format("%s: Could not finish committing transaction %s - no CohortEntry found",
-                                persistenceId(), transactionID));
-                LOG.error(ex.getMessage());
-                sender.tell(new akka.actor.Status.Failure(ex), getSelf());
-            }
-        } else {
-            finishCommit(sender, transactionID, cohortEntry);
-        }
-    }
-
     private void handleCanCommitTransaction(final CanCommitTransaction canCommit) {
-        LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionID());
+        LOG.debug("{}: Can committing transaction {}", persistenceId(), canCommit.getTransactionId());
 
         if (isLeader()) {
-        commitCoordinator.handleCanCommit(canCommit.getTransactionID(), getSender(), this);
+            commitCoordinator.handleCanCommit(canCommit.getTransactionId(), getSender(), this);
         } else {
             ActorSelection leader = getLeader();
             if (leader == null) {
                 messageRetrySupport.addMessageToRetry(canCommit, getSender(),
-                        "Could not canCommit transaction " + canCommit.getTransactionID());
+                        "Could not canCommit transaction " + canCommit.getTransactionId());
             } else {
                 LOG.debug("{}: Forwarding CanCommitTransaction to leader {}", persistenceId(), leader);
                 leader.forward(canCommit, getContext());
@@ -462,17 +644,18 @@ public class Shard extends RaftActor {
         }
     }
 
-    protected void handleBatchedModificationsLocal(BatchedModifications batched, ActorRef sender) {
+    @SuppressWarnings("checkstyle:IllegalCatch")
+    protected void handleBatchedModificationsLocal(final BatchedModifications batched, final ActorRef sender) {
         try {
             commitCoordinator.handleBatchedModifications(batched, sender, this);
         } catch (Exception e) {
             LOG.error("{}: Error handling BatchedModifications for Tx {}", persistenceId(),
-                    batched.getTransactionID(), e);
-            sender.tell(new akka.actor.Status.Failure(e), getSelf());
+                    batched.getTransactionId(), e);
+            sender.tell(new Failure(e), getSelf());
         }
     }
 
-    private void handleBatchedModifications(BatchedModifications batched) {
+    private void handleBatchedModifications(final BatchedModifications batched) {
         // This message is sent to prepare the modifications transaction directly on the Shard as an
         // optimization to avoid the extra overhead of a separate ShardTransaction actor. On the last
         // BatchedModifications message, the caller sets the ready flag in the message indicating
@@ -492,30 +675,31 @@ public class Shard extends RaftActor {
             ActorSelection leader = getLeader();
             if (!isLeaderActive || leader == null) {
                 messageRetrySupport.addMessageToRetry(batched, getSender(),
-                        "Could not commit transaction " + batched.getTransactionID());
+                        "Could not commit transaction " + batched.getTransactionId());
             } else {
                 // If this is not the first batch and leadership changed in between batched messages,
                 // we need to reconstruct previous BatchedModifications from the transaction
                 // DataTreeModification, honoring the max batched modification count, and forward all the
                 // previous BatchedModifications to the new leader.
-                Collection<BatchedModifications> newModifications = commitCoordinator.createForwardedBatchedModifications(
-                        batched, datastoreContext.getShardBatchedModificationCount());
+                Collection<BatchedModifications> newModifications = commitCoordinator
+                        .createForwardedBatchedModifications(batched,
+                                datastoreContext.getShardBatchedModificationCount());
 
                 LOG.debug("{}: Forwarding {} BatchedModifications to leader {}", persistenceId(),
                         newModifications.size(), leader);
 
-                for(BatchedModifications bm: newModifications) {
+                for (BatchedModifications bm : newModifications) {
                     leader.forward(bm, getContext());
                 }
             }
         }
     }
 
-    private boolean failIfIsolatedLeader(ActorRef sender) {
-        if(isIsolatedLeader()) {
-            sender.tell(new akka.actor.Status.Failure(new NoShardLeaderException(String.format(
-                    "Shard %s was the leader but has lost contact with all of its followers. Either all" +
-                    " other follower nodes are down or this node is isolated by a network partition.",
+    private boolean failIfIsolatedLeader(final ActorRef sender) {
+        if (isIsolatedLeader()) {
+            sender.tell(new Failure(new NoShardLeaderException(String.format(
+                    "Shard %s was the leader but has lost contact with all of its followers. Either all"
+                    " other follower nodes are down or this node is isolated by a network partition.",
                     persistenceId()))), getSelf());
             return true;
         }
@@ -527,8 +711,9 @@ public class Shard extends RaftActor {
         return getRaftState() == RaftState.IsolatedLeader;
     }
 
+    @SuppressWarnings("checkstyle:IllegalCatch")
     private void handleReadyLocalTransaction(final ReadyLocalTransaction message) {
-        LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionID());
+        LOG.debug("{}: handleReadyLocalTransaction for {}", persistenceId(), message.getTransactionId());
 
         boolean isLeaderActive = isLeaderActive();
         if (isLeader() && isLeaderActive) {
@@ -536,14 +721,14 @@ public class Shard extends RaftActor {
                 commitCoordinator.handleReadyLocalTransaction(message, getSender(), this);
             } catch (Exception e) {
                 LOG.error("{}: Error handling ReadyLocalTransaction for Tx {}", persistenceId(),
-                        message.getTransactionID(), e);
-                getSender().tell(new akka.actor.Status.Failure(e), getSelf());
+                        message.getTransactionId(), e);
+                getSender().tell(new Failure(e), getSelf());
             }
         } else {
             ActorSelection leader = getLeader();
             if (!isLeaderActive || leader == null) {
                 messageRetrySupport.addMessageToRetry(message, getSender(),
-                        "Could not commit transaction " + message.getTransactionID());
+                        "Could not commit transaction " + message.getTransactionId());
             } else {
                 LOG.debug("{}: Forwarding ReadyLocalTransaction to leader {}", persistenceId(), leader);
                 message.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
@@ -552,22 +737,21 @@ public class Shard extends RaftActor {
         }
     }
 
-    private void handleForwardedReadyTransaction(ForwardedReadyTransaction forwardedReady) {
-        LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionID());
+    private void handleForwardedReadyTransaction(final ForwardedReadyTransaction forwardedReady) {
+        LOG.debug("{}: handleForwardedReadyTransaction for {}", persistenceId(), forwardedReady.getTransactionId());
 
         boolean isLeaderActive = isLeaderActive();
         if (isLeader() && isLeaderActive) {
-            commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this,
-                    store.getSchemaContext());
+            commitCoordinator.handleForwardedReadyTransaction(forwardedReady, getSender(), this);
         } else {
             ActorSelection leader = getLeader();
             if (!isLeaderActive || leader == null) {
                 messageRetrySupport.addMessageToRetry(forwardedReady, getSender(),
-                        "Could not commit transaction " + forwardedReady.getTransactionID());
+                        "Could not commit transaction " + forwardedReady.getTransactionId());
             } else {
                 LOG.debug("{}: Forwarding ForwardedReadyTransaction to leader {}", persistenceId(), leader);
 
-                ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionID(),
+                ReadyLocalTransaction readyLocal = new ReadyLocalTransaction(forwardedReady.getTransactionId(),
                         forwardedReady.getTransaction().getSnapshot(), forwardedReady.isDoImmediateCommit());
                 readyLocal.setRemoteVersion(getCurrentBehavior().getLeaderPayloadVersion());
                 leader.forward(readyLocal, getContext());
@@ -576,10 +760,10 @@ public class Shard extends RaftActor {
     }
 
     private void handleAbortTransaction(final AbortTransaction abort) {
-        doAbortTransaction(abort.getTransactionID(), getSender());
+        doAbortTransaction(abort.getTransactionId(), getSender());
     }
 
-    void doAbortTransaction(final TransactionIdentifier transactionID, final ActorRef sender) {
+    void doAbortTransaction(final Identifier transactionID, final ActorRef sender) {
         commitCoordinator.handleAbort(transactionID, sender, this);
     }
 
@@ -589,19 +773,22 @@ public class Shard extends RaftActor {
         } else if (getLeader() != null) {
             getLeader().forward(message, getContext());
         } else {
-            getSender().tell(new akka.actor.Status.Failure(new NoShardLeaderException(
+            getSender().tell(new Failure(new NoShardLeaderException(
                     "Could not create a shard transaction", persistenceId())), getSelf());
         }
     }
 
     private void closeTransactionChain(final CloseTransactionChain closeTransactionChain) {
-        store.closeTransactionChain(closeTransactionChain.getIdentifier());
+        final LocalHistoryIdentifier id = closeTransactionChain.getIdentifier();
+        store.closeTransactionChain(id, null);
+        store.purgeTransactionChain(id, null);
     }
 
-    private void createTransaction(CreateTransaction createTransaction) {
+    @SuppressWarnings("checkstyle:IllegalCatch")
+    private void createTransaction(final CreateTransaction createTransaction) {
         try {
-            if(TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY &&
-                    failIfIsolatedLeader(getSender())) {
+            if (TransactionType.fromInt(createTransaction.getTransactionType()) != TransactionType.READ_ONLY
+                    && failIfIsolatedLeader(getSender())) {
                 return;
             }
 
@@ -611,29 +798,16 @@ public class Shard extends RaftActor {
             getSender().tell(new CreateTransactionReply(Serialization.serializedActorPath(transactionActor),
                     createTransaction.getTransactionId(), createTransaction.getVersion()).toSerializable(), getSelf());
         } catch (Exception e) {
-            getSender().tell(new akka.actor.Status.Failure(e), getSelf());
+            getSender().tell(new Failure(e), getSelf());
         }
     }
 
-    private ActorRef createTransaction(int transactionType, TransactionIdentifier transactionId) {
+    private ActorRef createTransaction(final int transactionType, final TransactionIdentifier transactionId) {
         LOG.debug("{}: Creating transaction : {} ", persistenceId(), transactionId);
         return transactionActorFactory.newShardTransaction(TransactionType.fromInt(transactionType),
             transactionId);
     }
 
-    private void commitWithNewTransaction(final BatchedModifications modification) {
-        ReadWriteShardDataTreeTransaction tx = store.newReadWriteTransaction(modification.getTransactionID());
-        modification.apply(tx.getSnapshot());
-        try {
-            snapshotCohort.syncCommitTransaction(tx);
-            shardMBean.incrementCommittedTransactionCount();
-            shardMBean.setLastCommittedTransactionTime(System.currentTimeMillis());
-        } catch (Exception e) {
-            shardMBean.incrementFailedTransactionsCount();
-            LOG.error("{}: Failed to commit", persistenceId(), e);
-        }
-    }
-
     private void updateSchemaContext(final UpdateSchemaContext message) {
         updateSchemaContext(message.getSchemaContext());
     }
@@ -669,7 +843,7 @@ public class Shard extends RaftActor {
         getContext().parent().tell(new ActorInitialized(), getSelf());
 
         // Being paranoid here - this method should only be called once but just in case...
-        if(txCommitTimeoutCheckSchedule == null) {
+        if (txCommitTimeoutCheckSchedule == null) {
             // Schedule a message to be periodically sent to check if the current in-progress
             // transaction should be expired and aborted.
             FiniteDuration period = Duration.create(transactionCommitTimeout / 3, TimeUnit.MILLISECONDS);
@@ -681,40 +855,14 @@ public class Shard extends RaftActor {
 
     @Override
     protected void applyState(final ActorRef clientActor, final Identifier identifier, final Object data) {
-        if (data instanceof DataTreeCandidateSupplier) {
-            if (clientActor == null) {
-                // No clientActor indicates a replica coming from the leader
-                try {
-                    store.applyForeignCandidate(identifier, ((DataTreeCandidateSupplier)data).getCandidate().getValue());
-                } catch (DataValidationFailedException | IOException e) {
-                    LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
-                }
-            } else {
-                // Replication consensus reached, proceed to commit
-                finishCommit(clientActor, identifier);
+        if (data instanceof Payload) {
+            try {
+                store.applyReplicatedPayload(identifier, (Payload)data);
+            } catch (DataValidationFailedException | IOException e) {
+                LOG.error("{}: Error applying replica {}", persistenceId(), identifier, e);
             }
         } else {
-            LOG.error("{}: Unknown state received {} ClassLoader {}", persistenceId(), data,
-                data.getClass().getClassLoader());
-        }
-    }
-
-    private void applyModificationToState(ActorRef clientActor, Identifier identifier, Object modification) {
-        if(modification == null) {
-            LOG.error(
-                    "{}: modification is null - this is very unexpected, clientActor = {}, identifier = {}",
-                    persistenceId(), identifier, clientActor != null ? clientActor.path().toString() : null);
-        } else if(clientActor == null) {
-            // There's no clientActor to which to send a commit reply so we must be applying
-            // replicated state from the leader.
-
-            // The only implementation we know of is BatchedModifications, which also carries a transaction
-            // identifier -- which we really need that.
-            Preconditions.checkArgument(modification instanceof BatchedModifications);
-            commitWithNewTransaction((BatchedModifications)modification);
-        } else {
-            // This must be the OK to commit after replication consensus.
-            finishCommit(clientActor, identifier);
+            LOG.error("{}: Unknown state for {} received {}", persistenceId(), identifier, data);
         }
     }
 
@@ -727,57 +875,106 @@ public class Shard extends RaftActor {
 
         // If this actor is no longer the leader close all the transaction chains
         if (!isLeader) {
-            if(LOG.isDebugEnabled()) {
+            if (LOG.isDebugEnabled()) {
                 LOG.debug(
                     "{}: onStateChanged: Closing all transaction chains because shard {} is no longer the leader",
                     persistenceId(), getId());
             }
 
-            store.closeAllTransactionChains();
+            paused = false;
+            store.purgeLeaderState();
         }
 
-        if(hasLeader && !isIsolatedLeader()) {
+        if (hasLeader && !isIsolatedLeader()) {
             messageRetrySupport.retryMessages();
         }
     }
 
     @Override
-    protected void onLeaderChanged(String oldLeader, String newLeader) {
+    protected void onLeaderChanged(final String oldLeader, final String newLeader) {
         shardMBean.incrementLeadershipChangeCount();
+        paused = false;
+
+        if (!isLeader()) {
+            if (!knownFrontends.isEmpty()) {
+                LOG.debug("{}: removing frontend state for {}", persistenceId(), knownFrontends.keySet());
+                knownFrontends = ImmutableMap.of();
+            }
+
+            requestMessageAssembler.close();
+
+            if (!hasLeader()) {
+                // No leader anywhere, nothing else to do
+                return;
+            }
 
-        boolean hasLeader = hasLeader();
-        if(hasLeader && !isLeader()) {
             // Another leader was elected. If we were the previous leader and had pending transactions, convert
             // them to transaction messages and send to the new leader.
             ActorSelection leader = getLeader();
-            if(leader != null) {
-                Collection<Object> messagesToForward = commitCoordinator.convertPendingTransactionsToMessages(
-                        datastoreContext.getShardBatchedModificationCount());
+            if (leader != null) {
+                Collection<?> messagesToForward = convertPendingTransactionsToMessages();
 
-                if(!messagesToForward.isEmpty()) {
+                if (!messagesToForward.isEmpty()) {
                     LOG.debug("{}: Forwarding {} pending transaction messages to leader {}", persistenceId(),
                             messagesToForward.size(), leader);
 
-                    for(Object message: messagesToForward) {
+                    for (Object message : messagesToForward) {
                         leader.tell(message, self());
                     }
                 }
             } else {
-                commitCoordinator.abortPendingTransactions(
-                        "The transacton was aborted due to inflight leadership change and the leader address isn't available.",
-                        this);
+                commitCoordinator.abortPendingTransactions("The transacton was aborted due to inflight leadership "
+                        + "change and the leader address isn't available.", this);
             }
+        } else {
+            // We have become the leader, we need to reconstruct frontend state
+            knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
+            LOG.debug("{}: became leader with frontend state for {}", persistenceId(), knownFrontends.keySet());
         }
 
-        if(hasLeader && !isIsolatedLeader()) {
+        if (!isIsolatedLeader()) {
             messageRetrySupport.retryMessages();
         }
     }
 
+    /**
+     * Clears all pending transactions and converts them to messages to be forwarded to a new leader.
+     *
+     * @return the converted messages
+     */
+    public Collection<?> convertPendingTransactionsToMessages() {
+        return commitCoordinator.convertPendingTransactionsToMessages(
+                datastoreContext.getShardBatchedModificationCount());
+    }
+
     @Override
-    protected void pauseLeader(Runnable operation) {
+    protected void pauseLeader(final Runnable operation) {
         LOG.debug("{}: In pauseLeader, operation: {}", persistenceId(), operation);
-        commitCoordinator.setRunOnPendingTransactionsComplete(operation);
+        paused = true;
+
+        // Tell-based protocol can replay transaction state, so it is safe to blow it up when we are paused.
+        knownFrontends.values().forEach(LeaderFrontendState::retire);
+        knownFrontends = ImmutableMap.of();
+
+        store.setRunOnPendingTransactionsComplete(operation);
+    }
+
+    @Override
+    protected void unpauseLeader() {
+        LOG.debug("{}: In unpauseLeader", persistenceId());
+        paused = false;
+
+        store.setRunOnPendingTransactionsComplete(null);
+
+        // Restore tell-based protocol state as if we were becoming the leader
+        knownFrontends = Verify.verifyNotNull(frontendMetadata.toLeaderState(this));
+    }
+
+    @Override
+    protected OnDemandRaftState.AbstractBuilder<?> newOnDemandRaftStateBuilder() {
+        return OnDemandShardState.newBuilder().treeChangeListenerActors(treeChangeSupport.getListenerActors())
+                .dataChangeListenerActors(changeSupport.getListenerActors())
+                .commitCohortActors(store.getCohortActors());
     }
 
     @Override
@@ -808,16 +1005,17 @@ public class Shard extends RaftActor {
         return new Builder();
     }
 
-    public static abstract class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
+    public abstract static class AbstractBuilder<T extends AbstractBuilder<T, S>, S extends Shard> {
         private final Class<S> shardClass;
         private ShardIdentifier id;
         private Map<String, String> peerAddresses = Collections.emptyMap();
         private DatastoreContext datastoreContext;
-        private SchemaContext schemaContext;
+        private SchemaContextProvider schemaContextProvider;
         private DatastoreSnapshot.ShardSnapshot restoreFromSnapshot;
+        private TipProducingDataTree dataTree;
         private volatile boolean sealed;
 
-        protected AbstractBuilder(Class<S> shardClass) {
+        protected AbstractBuilder(final Class<S> shardClass) {
             this.shardClass = shardClass;
         }
 
@@ -830,33 +1028,39 @@ public class Shard extends RaftActor {
             return (T) this;
         }
 
-        public T id(ShardIdentifier id) {
+        public T id(final ShardIdentifier newId) {
             checkSealed();
-            this.id = id;
+            this.id = newId;
             return self();
         }
 
-        public T peerAddresses(Map<String, String> peerAddresses) {
+        public T peerAddresses(final Map<String, String> newPeerAddresses) {
             checkSealed();
-            this.peerAddresses = peerAddresses;
+            this.peerAddresses = newPeerAddresses;
             return self();
         }
 
-        public T datastoreContext(DatastoreContext datastoreContext) {
+        public T datastoreContext(final DatastoreContext newDatastoreContext) {
             checkSealed();
-            this.datastoreContext = datastoreContext;
+            this.datastoreContext = newDatastoreContext;
             return self();
         }
 
-        public T schemaContext(SchemaContext schemaContext) {
+        public T schemaContextProvider(final SchemaContextProvider schemaContextProvider) {
             checkSealed();
-            this.schemaContext = schemaContext;
+            this.schemaContextProvider = Preconditions.checkNotNull(schemaContextProvider);
             return self();
         }
 
-        public T restoreFromSnapshot(DatastoreSnapshot.ShardSnapshot restoreFromSnapshot) {
+        public T restoreFromSnapshot(final DatastoreSnapshot.ShardSnapshot newRestoreFromSnapshot) {
             checkSealed();
-            this.restoreFromSnapshot = restoreFromSnapshot;
+            this.restoreFromSnapshot = newRestoreFromSnapshot;
+            return self();
+        }
+
+        public T dataTree(final TipProducingDataTree newDataTree) {
+            checkSealed();
+            this.dataTree = newDataTree;
             return self();
         }
 
@@ -873,29 +1077,34 @@ public class Shard extends RaftActor {
         }
 
         public SchemaContext getSchemaContext() {
-            return schemaContext;
+            return Verify.verifyNotNull(schemaContextProvider.getSchemaContext());
         }
 
         public DatastoreSnapshot.ShardSnapshot getRestoreFromSnapshot() {
             return restoreFromSnapshot;
         }
 
+        public TipProducingDataTree getDataTree() {
+            return dataTree;
+        }
+
         public TreeType getTreeType() {
             switch (datastoreContext.getLogicalStoreType()) {
-            case CONFIGURATION:
-                return TreeType.CONFIGURATION;
-            case OPERATIONAL:
-                return TreeType.OPERATIONAL;
+                case CONFIGURATION:
+                    return TreeType.CONFIGURATION;
+                case OPERATIONAL:
+                    return TreeType.OPERATIONAL;
+                default:
+                    throw new IllegalStateException("Unhandled logical store type "
+                            + datastoreContext.getLogicalStoreType());
             }
-
-            throw new IllegalStateException("Unhandled logical store type " + datastoreContext.getLogicalStoreType());
         }
 
         protected void verify() {
             Preconditions.checkNotNull(id, "id should not be null");
             Preconditions.checkNotNull(peerAddresses, "peerAddresses should not be null");
             Preconditions.checkNotNull(datastoreContext, "dataStoreContext should not be null");
-            Preconditions.checkNotNull(schemaContext, "schemaContext should not be null");
+            Preconditions.checkNotNull(schemaContextProvider, "schemaContextProvider should not be null");
         }
 
         public Props props() {
@@ -910,4 +1119,12 @@ public class Shard extends RaftActor {
             super(Shard.class);
         }
     }
+
+    Ticker ticker() {
+        return Ticker.systemTicker();
+    }
+
+    void scheduleNextPendingTransaction() {
+        self().tell(RESUME_NEXT_PENDING_TRANSACTION, ActorRef.noSender());
+    }
 }