BUG-5280: refactor CohortEntry
[controller.git] / opendaylight / md-sal / sal-distributed-datastore / src / main / java / org / opendaylight / controller / cluster / datastore / ShardCommitCoordinator.java
index b96e38d76a45aced0e1c326c051cb7fcbd1d82d1..eb0c04dbbd86eaaabde73326baf1b35086073ce1 100644 (file)
@@ -8,26 +8,34 @@
 package org.opendaylight.controller.cluster.datastore;
 
 import akka.actor.ActorRef;
-import akka.actor.Status;
+import akka.actor.Status.Failure;
 import akka.serialization.Serialization;
 import com.google.common.annotations.VisibleForTesting;
 import com.google.common.base.Preconditions;
-import com.google.common.cache.Cache;
-import com.google.common.cache.CacheBuilder;
-import com.google.common.cache.RemovalCause;
-import com.google.common.cache.RemovalListener;
-import com.google.common.cache.RemovalNotification;
+import java.util.ArrayList;
+import java.util.Collection;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.Iterator;
 import java.util.LinkedList;
+import java.util.List;
+import java.util.Map;
 import java.util.Queue;
 import java.util.concurrent.ExecutionException;
-import java.util.concurrent.TimeUnit;
+import org.opendaylight.controller.cluster.datastore.DataTreeCohortActorRegistry.CohortRegistryCommand;
+import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
+import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransactionReply;
-import org.opendaylight.controller.cluster.datastore.modification.Modification;
-import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
-import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
-import org.opendaylight.controller.sal.core.spi.data.DOMStoreWriteTransaction;
+import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
+import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
+import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
+import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
+import org.opendaylight.controller.cluster.datastore.utils.AbstractBatchedModificationsCursor;
+import org.opendaylight.controller.md.sal.common.api.data.TransactionCommitFailedException;
+import org.opendaylight.yangtools.concepts.Identifier;
+import org.opendaylight.yangtools.yang.model.api.SchemaContext;
 import org.slf4j.Logger;
 
 /**
@@ -35,20 +43,24 @@ import org.slf4j.Logger;
  *
  * @author Thomas Pantelis
  */
-public class ShardCommitCoordinator {
+final class ShardCommitCoordinator {
 
     // Interface hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
     public interface CohortDecorator {
-        DOMStoreThreePhaseCommitCohort decorate(String transactionID, DOMStoreThreePhaseCommitCohort actual);
+        ShardDataTreeCohort decorate(Identifier transactionID, ShardDataTreeCohort actual);
     }
 
-    private final Cache<String, CohortEntry> cohortCache;
+    private final Map<Identifier, CohortEntry> cohortCache = new HashMap<>();
 
     private CohortEntry currentCohortEntry;
 
-    private final DOMTransactionFactory transactionFactory;
+    private final ShardDataTree dataTree;
 
-    private final Queue<CohortEntry> queuedCohortEntries;
+    private final DataTreeCohortActorRegistry cohortRegistry = new DataTreeCohortActorRegistry();
+
+    // We use a LinkedList here to avoid synchronization overhead with concurrent queue impls
+    // since this should only be accessed on the shard's dispatcher.
+    private final Queue<CohortEntry> queuedCohortEntries = new LinkedList<>();
 
     private int queueCapacity;
 
@@ -56,55 +68,99 @@ public class ShardCommitCoordinator {
 
     private final String name;
 
-    private final String shardActorPath;
-
-    private final RemovalListener<String, CohortEntry> cacheRemovalListener =
-            new RemovalListener<String, CohortEntry>() {
-                @Override
-                public void onRemoval(RemovalNotification<String, CohortEntry> notification) {
-                    if(notification.getCause() == RemovalCause.EXPIRED) {
-                        log.warn("{}: Transaction {} was timed out of the cache", name, notification.getKey());
-                    }
-                }
-            };
+    private final long cacheExpiryTimeoutInMillis;
 
     // This is a hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
     private CohortDecorator cohortDecorator;
 
-    public ShardCommitCoordinator(DOMTransactionFactory transactionFactory,
-            long cacheExpiryTimeoutInSec, int queueCapacity, ActorRef shardActor, Logger log, String name) {
+    private ReadyTransactionReply readyTransactionReply;
+
+    private Runnable runOnPendingTransactionsComplete;
+
+    ShardCommitCoordinator(ShardDataTree dataTree, long cacheExpiryTimeoutInMillis, int queueCapacity, Logger log,
+            String name) {
 
         this.queueCapacity = queueCapacity;
         this.log = log;
         this.name = name;
-        this.transactionFactory = transactionFactory;
-
-        shardActorPath = Serialization.serializedActorPath(shardActor);
+        this.dataTree = Preconditions.checkNotNull(dataTree);
+        this.cacheExpiryTimeoutInMillis = cacheExpiryTimeoutInMillis;
+    }
 
-        cohortCache = CacheBuilder.newBuilder().expireAfterAccess(cacheExpiryTimeoutInSec, TimeUnit.SECONDS).
-                removalListener(cacheRemovalListener).build();
+    int getQueueSize() {
+        return queuedCohortEntries.size();
+    }
 
-        // We use a LinkedList here to avoid synchronization overhead with concurrent queue impls
-        // since this should only be accessed on the shard's dispatcher.
-        queuedCohortEntries = new LinkedList<>();
+    int getCohortCacheSize() {
+        return cohortCache.size();
     }
 
-    public void setQueueCapacity(int queueCapacity) {
+    void setQueueCapacity(int queueCapacity) {
         this.queueCapacity = queueCapacity;
     }
 
+    private ReadyTransactionReply readyTransactionReply(Shard shard) {
+        if(readyTransactionReply == null) {
+            readyTransactionReply = new ReadyTransactionReply(Serialization.serializedActorPath(shard.self()));
+        }
+
+        return readyTransactionReply;
+    }
+
+    private boolean queueCohortEntry(CohortEntry cohortEntry, ActorRef sender, Shard shard) {
+        if(queuedCohortEntries.size() < queueCapacity) {
+            queuedCohortEntries.offer(cohortEntry);
+
+            log.debug("{}: Enqueued transaction {}, queue size {}", name, cohortEntry.getTransactionID(),
+                    queuedCohortEntries.size());
+
+            return true;
+        } else {
+            cohortCache.remove(cohortEntry.getTransactionID());
+
+            final RuntimeException ex = new RuntimeException(
+                    String.format("%s: Could not enqueue transaction %s - the maximum commit queue"+
+                                  " capacity %d has been reached.",
+                                  name, cohortEntry.getTransactionID(), queueCapacity));
+            log.error(ex.getMessage());
+            sender.tell(new Failure(ex), shard.self());
+            return false;
+        }
+    }
+
     /**
      * This method is called to ready a transaction that was prepared by ShardTransaction actor. It caches
      * the prepared cohort entry for the given transactions ID in preparation for the subsequent 3-phase commit.
      *
-     * @param transactionID the ID of the transaction
-     * @param cohort the cohort to participate in the transaction commit
-     * @param modification the modifications made by the transaction
+     * @param ready the ForwardedReadyTransaction message to process
+     * @param sender the sender of the message
+     * @param shard the transaction's shard actor
+     * @param schema
      */
-    public void transactionReady(String transactionID, DOMStoreThreePhaseCommitCohort cohort,
-            MutableCompositeModification modification) {
+    void handleForwardedReadyTransaction(ForwardedReadyTransaction ready, ActorRef sender, Shard shard,
+            SchemaContext schema) {
+        log.debug("{}: Readying transaction {}, client version {}", name,
+                ready.getTransactionID(), ready.getTxnClientVersion());
+
+        final ShardDataTreeCohort cohort = ready.getTransaction().ready();
+        final CohortEntry cohortEntry = CohortEntry.createReady(ready.getTransactionID(), cohort, cohortRegistry,
+            schema, ready.getTxnClientVersion());
+        cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
+
+        if(!queueCohortEntry(cohortEntry, sender, shard)) {
+            return;
+        }
 
-        cohortCache.put(transactionID, new CohortEntry(transactionID, cohort, modification));
+        if(ready.isDoImmediateCommit()) {
+            cohortEntry.setDoImmediateCommit(true);
+            cohortEntry.setReplySender(sender);
+            cohortEntry.setShard(shard);
+            handleCanCommit(cohortEntry);
+        } else {
+            // The caller does not want immediate commit - the 3-phase commit will be coordinated by the
+            // front-end so send back a ReadyTransactionReply with our actor path.
+            sender.tell(readyTransactionReply(shard), shard.self());
+        }
     }
 
     /**
@@ -113,20 +169,16 @@ public class ShardCommitCoordinator {
      * DOMStoreWriteTransaction, one is created. The batched modifications are applied to the write Tx. If
      * the BatchedModifications is ready to commit then a DOMStoreThreePhaseCommitCohort is created.
      *
-     * @param batched the BatchedModifications
-     * @param shardActor the transaction's shard actor
-     *
-     * @throws ExecutionException if an error occurs loading the cache
+     * @param batched the BatchedModifications message to process
+     * @param sender the sender of the message
      */
-    public boolean handleTransactionModifications(BatchedModifications batched)
-            throws ExecutionException {
-        CohortEntry cohortEntry = cohortCache.getIfPresent(batched.getTransactionID());
+    void handleBatchedModifications(BatchedModifications batched, ActorRef sender, Shard shard) {
+        CohortEntry cohortEntry = cohortCache.get(batched.getTransactionID());
         if(cohortEntry == null) {
-            cohortEntry = new CohortEntry(batched.getTransactionID(),
-                    transactionFactory.<DOMStoreWriteTransaction>newTransaction(
-                        TransactionProxy.TransactionType.WRITE_ONLY, batched.getTransactionID(),
-                        batched.getTransactionChainID()));
-            cohortCache.put(batched.getTransactionID(), cohortEntry);
+            cohortEntry = CohortEntry.createOpen(batched.getTransactionID(),
+                    dataTree.newReadWriteTransaction(batched.getTransactionID()),
+                    cohortRegistry, dataTree.getSchemaContext(), batched.getVersion());
+            cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
         }
 
         if(log.isDebugEnabled()) {
@@ -137,99 +189,381 @@ public class ShardCommitCoordinator {
         cohortEntry.applyModifications(batched.getModifications());
 
         if(batched.isReady()) {
+            if(cohortEntry.getLastBatchedModificationsException() != null) {
+                cohortCache.remove(cohortEntry.getTransactionID());
+                throw cohortEntry.getLastBatchedModificationsException();
+            }
+
+            if(cohortEntry.getTotalBatchedModificationsReceived() != batched.getTotalMessagesSent()) {
+                cohortCache.remove(cohortEntry.getTransactionID());
+                throw new IllegalStateException(String.format(
+                        "The total number of batched messages received %d does not match the number sent %d",
+                        cohortEntry.getTotalBatchedModificationsReceived(), batched.getTotalMessagesSent()));
+            }
+
+            if(!queueCohortEntry(cohortEntry, sender, shard)) {
+                return;
+            }
+
             if(log.isDebugEnabled()) {
                 log.debug("{}: Readying Tx {}, client version {}", name,
                         batched.getTransactionID(), batched.getVersion());
             }
 
-            cohortEntry.ready(cohortDecorator);
-        }
+            cohortEntry.ready(cohortDecorator, batched.isDoCommitOnReady());
 
-        return batched.isReady();
+            if(batched.isDoCommitOnReady()) {
+                cohortEntry.setReplySender(sender);
+                cohortEntry.setShard(shard);
+                handleCanCommit(cohortEntry);
+            } else {
+                sender.tell(readyTransactionReply(shard), shard.self());
+            }
+        } else {
+            sender.tell(new BatchedModificationsReply(batched.getModifications().size()), shard.self());
+        }
     }
 
     /**
-     * This method handles the canCommit phase for a transaction.
+     * This method handles {@link ReadyLocalTransaction} message. All transaction modifications have
+     * been prepared beforehand by the sender and we just need to drive them through into the
+     * dataTree.
      *
-     * @param canCommit the CanCommitTransaction message
-     * @param sender the actor that sent the message
+     * @param message the ReadyLocalTransaction message to process
+     * @param sender the sender of the message
      * @param shard the transaction's shard actor
      */
-    public void handleCanCommit(CanCommitTransaction canCommit, final ActorRef sender,
-            final ActorRef shard) {
-        String transactionID = canCommit.getTransactionID();
-        if(log.isDebugEnabled()) {
-            log.debug("{}: Processing canCommit for transaction {} for shard {}",
-                    name, transactionID, shard.path());
+    void handleReadyLocalTransaction(ReadyLocalTransaction message, ActorRef sender, Shard shard) {
+        final ShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(dataTree, message.getModification(),
+                message.getTransactionID());
+        final CohortEntry cohortEntry = CohortEntry.createReady(message.getTransactionID(), cohort, cohortRegistry,
+            dataTree.getSchemaContext(), DataStoreVersions.CURRENT_VERSION);
+        cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
+        cohortEntry.setDoImmediateCommit(message.isDoCommitOnReady());
+
+        if(!queueCohortEntry(cohortEntry, sender, shard)) {
+            return;
+        }
+
+        log.debug("{}: Applying local modifications for Tx {}", name, message.getTransactionID());
+
+        if (message.isDoCommitOnReady()) {
+            cohortEntry.setReplySender(sender);
+            cohortEntry.setShard(shard);
+            handleCanCommit(cohortEntry);
+        } else {
+            sender.tell(readyTransactionReply(shard), shard.self());
+        }
+    }
+
+    Collection<BatchedModifications> createForwardedBatchedModifications(final BatchedModifications from,
+            final int maxModificationsPerBatch) {
+        CohortEntry cohortEntry = getAndRemoveCohortEntry(from.getTransactionID());
+        if(cohortEntry == null || cohortEntry.getTransaction() == null) {
+            return Collections.singletonList(from);
         }
 
+        cohortEntry.applyModifications(from.getModifications());
+
+        final LinkedList<BatchedModifications> newModifications = new LinkedList<>();
+        cohortEntry.getTransaction().getSnapshot().applyToCursor(new AbstractBatchedModificationsCursor() {
+            @Override
+            protected BatchedModifications getModifications() {
+                if(newModifications.isEmpty() ||
+                        newModifications.getLast().getModifications().size() >= maxModificationsPerBatch) {
+                    newModifications.add(new BatchedModifications(from.getTransactionID(), from.getVersion()));
+                }
+
+                return newModifications.getLast();
+            }
+        });
+
+        BatchedModifications last = newModifications.getLast();
+        last.setDoCommitOnReady(from.isDoCommitOnReady());
+        last.setReady(from.isReady());
+        last.setTotalMessagesSent(newModifications.size());
+        return newModifications;
+    }
+
+    private void handleCanCommit(CohortEntry cohortEntry) {
+        cohortEntry.updateLastAccessTime();
+
+        if(currentCohortEntry != null) {
+            // There's already a Tx commit in progress so we can't process this entry yet - but it's in the
+            // queue and will get processed after all prior entries complete.
+
+            if(log.isDebugEnabled()) {
+                log.debug("{}: Commit for Tx {} already in progress - skipping canCommit for {} for now",
+                        name, currentCohortEntry.getTransactionID(), cohortEntry.getTransactionID());
+            }
+
+            return;
+        }
+
+        // No Tx commit currently in progress - check if this entry is the next one in the queue, If so make
+        // it the current entry and proceed with canCommit.
+        // Purposely checking reference equality here.
+        if(queuedCohortEntries.peek() == cohortEntry) {
+            currentCohortEntry = queuedCohortEntries.poll();
+            doCanCommit(currentCohortEntry);
+        } else {
+            if(log.isDebugEnabled()) {
+                log.debug("{}: Tx {} is the next pending canCommit - skipping {} for now", name,
+                        queuedCohortEntries.peek() != null ? queuedCohortEntries.peek().getTransactionID() : "???",
+                                cohortEntry.getTransactionID());
+            }
+        }
+    }
+
+    /**
+     * This method handles the canCommit phase for a transaction.
+     *
+     * @param transactionID the ID of the transaction to canCommit
+     * @param sender the actor to which to send the response
+     * @param shard the transaction's shard actor
+     */
+    void handleCanCommit(Identifier transactionID, final ActorRef sender, final Shard shard) {
         // Lookup the cohort entry that was cached previously (or should have been) by
         // transactionReady (via the ForwardedReadyTransaction message).
-        final CohortEntry cohortEntry = cohortCache.getIfPresent(transactionID);
+        final CohortEntry cohortEntry = cohortCache.get(transactionID);
         if(cohortEntry == null) {
             // Either canCommit was invoked before ready(shouldn't happen)  or a long time passed
             // between canCommit and ready and the entry was expired from the cache.
             IllegalStateException ex = new IllegalStateException(
                     String.format("%s: No cohort entry found for transaction %s", name, transactionID));
             log.error(ex.getMessage());
-            sender.tell(new Status.Failure(ex), shard);
+            sender.tell(new Failure(ex), shard.self());
             return;
         }
 
-        cohortEntry.setCanCommitSender(sender);
+        cohortEntry.setReplySender(sender);
         cohortEntry.setShard(shard);
 
-        if(currentCohortEntry != null) {
-            // There's already a Tx commit in progress - attempt to queue this entry to be
-            // committed after the current Tx completes.
-            log.debug("{}: Transaction {} is already in progress - queueing transaction {}",
-                    name, currentCohortEntry.getTransactionID(), transactionID);
+        handleCanCommit(cohortEntry);
+    }
+
+    private void doCanCommit(final CohortEntry cohortEntry) {
+        boolean canCommit = false;
+        try {
+            canCommit = cohortEntry.canCommit();
 
-            if(queuedCohortEntries.size() < queueCapacity) {
-                queuedCohortEntries.offer(cohortEntry);
+            log.debug("{}: canCommit for {}: {}", name, cohortEntry.getTransactionID(), canCommit);
+
+            if(cohortEntry.isDoImmediateCommit()) {
+                if(canCommit) {
+                    doCommit(cohortEntry);
+                } else {
+                    cohortEntry.getReplySender().tell(new Failure(new TransactionCommitFailedException(
+                                "Can Commit failed, no detailed cause available.")), cohortEntry.getShard().self());
+                }
             } else {
-                removeCohortEntry(transactionID);
-
-                RuntimeException ex = new RuntimeException(
-                        String.format("%s: Could not enqueue transaction %s - the maximum commit queue"+
-                                      " capacity %d has been reached.",
-                                      name, transactionID, queueCapacity));
-                log.error(ex.getMessage());
-                sender.tell(new Status.Failure(ex), shard);
+                cohortEntry.getReplySender().tell(
+                        canCommit ? CanCommitTransactionReply.yes(cohortEntry.getClientVersion()).toSerializable() :
+                            CanCommitTransactionReply.no(cohortEntry.getClientVersion()).toSerializable(),
+                        cohortEntry.getShard().self());
             }
-        } else {
-            // No Tx commit currently in progress - make this the current entry and proceed with
-            // canCommit.
+        } catch (Exception e) {
+            log.debug("{}: An exception occurred during canCommit", name, e);
+
+            Throwable failure = e;
+            if(e instanceof ExecutionException) {
+                failure = e.getCause();
+            }
+
+            cohortEntry.getReplySender().tell(new Failure(failure), cohortEntry.getShard().self());
+        } finally {
+            if(!canCommit) {
+                // Remove the entry from the cache now.
+                currentTransactionComplete(cohortEntry.getTransactionID(), true);
+            }
+        }
+    }
+
+    private boolean doCommit(CohortEntry cohortEntry) {
+        log.debug("{}: Committing transaction {}", name, cohortEntry.getTransactionID());
+
+        boolean success = false;
+
+        // We perform the preCommit phase here atomically with the commit phase. This is an
+        // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
+        // coordination of preCommit across shards in case of failure but preCommit should not
+        // normally fail since we ensure only one concurrent 3-phase commit.
+
+        try {
+            cohortEntry.preCommit();
+
+            cohortEntry.getShard().continueCommit(cohortEntry);
+
             cohortEntry.updateLastAccessTime();
-            currentCohortEntry = cohortEntry;
 
-            doCanCommit(cohortEntry);
+            success = true;
+        } catch (Exception e) {
+            log.error("{} An exception occurred while preCommitting transaction {}",
+                    name, cohortEntry.getTransactionID(), e);
+            cohortEntry.getReplySender().tell(new Failure(e), cohortEntry.getShard().self());
+
+            currentTransactionComplete(cohortEntry.getTransactionID(), true);
         }
+
+        return success;
     }
 
-    private void doCanCommit(final CohortEntry cohortEntry) {
+    /**
+     * This method handles the preCommit and commit phases for a transaction.
+     *
+     * @param transactionID the ID of the transaction to commit
+     * @param sender the actor to which to send the response
+     * @param shard the transaction's shard actor
+     * @return true if the transaction was successfully prepared, false otherwise.
+     */
+    boolean handleCommit(final Identifier transactionID, final ActorRef sender, final Shard shard) {
+        // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
+        // this transaction.
+        final CohortEntry cohortEntry = getCohortEntryIfCurrent(transactionID);
+        if(cohortEntry == null) {
+            // We're not the current Tx - the Tx was likely expired b/c it took too long in
+            // between the canCommit and commit messages.
+            IllegalStateException ex = new IllegalStateException(
+                    String.format("%s: Cannot commit transaction %s - it is not the current transaction",
+                            name, transactionID));
+            log.error(ex.getMessage());
+            sender.tell(new Failure(ex), shard.self());
+            return false;
+        }
+
+        cohortEntry.setReplySender(sender);
+        return doCommit(cohortEntry);
+    }
+
+    void handleAbort(final Identifier transactionID, final ActorRef sender, final Shard shard) {
+        CohortEntry cohortEntry = getCohortEntryIfCurrent(transactionID);
+        if(cohortEntry != null) {
+            // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
+            // aborted during replication in which case we may still commit locally if replication
+            // succeeds.
+            currentTransactionComplete(transactionID, false);
+        } else {
+            cohortEntry = getAndRemoveCohortEntry(transactionID);
+        }
+
+        if(cohortEntry == null) {
+            return;
+        }
+
+        log.debug("{}: Aborting transaction {}", name, transactionID);
 
+        final ActorRef self = shard.getSelf();
         try {
-            // We block on the future here so we don't have to worry about possibly accessing our
-            // state on a different thread outside of our dispatcher. Also, the data store
-            // currently uses a same thread executor anyway.
-            Boolean canCommit = cohortEntry.getCohort().canCommit().get();
+            cohortEntry.abort();
 
-            cohortEntry.getCanCommitSender().tell(
-                    canCommit ? CanCommitTransactionReply.YES.toSerializable() :
-                        CanCommitTransactionReply.NO.toSerializable(), cohortEntry.getShard());
+            shard.getShardMBean().incrementAbortTransactionsCount();
 
-            if(!canCommit) {
-                // Remove the entry from the cache now since the Tx will be aborted.
-                removeCohortEntry(cohortEntry.getTransactionID());
+            if(sender != null) {
+                sender.tell(AbortTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), self);
+            }
+        } catch (Exception e) {
+            log.error("{}: An exception happened during abort", name, e);
+
+            if(sender != null) {
+                sender.tell(new Failure(e), self);
+            }
+        }
+    }
+
+    void checkForExpiredTransactions(final long timeout, final Shard shard) {
+        CohortEntry cohortEntry = getCurrentCohortEntry();
+        if(cohortEntry != null) {
+            if(cohortEntry.isExpired(timeout)) {
+                log.warn("{}: Current transaction {} has timed out after {} ms - aborting",
+                        name, cohortEntry.getTransactionID(), timeout);
+
+                handleAbort(cohortEntry.getTransactionID(), null, shard);
+            }
+        }
+
+        cleanupExpiredCohortEntries();
+    }
+
+    void abortPendingTransactions(final String reason, final Shard shard) {
+        if(currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
+            return;
+        }
+
+        List<CohortEntry> cohortEntries = getAndClearPendingCohortEntries();
+
+        log.debug("{}: Aborting {} pending queued transactions", name, cohortEntries.size());
+
+        for(CohortEntry cohortEntry: cohortEntries) {
+            if(cohortEntry.getReplySender() != null) {
+                cohortEntry.getReplySender().tell(new Failure(new RuntimeException(reason)), shard.self());
+            }
+        }
+    }
+
+    private List<CohortEntry> getAndClearPendingCohortEntries() {
+        List<CohortEntry> cohortEntries = new ArrayList<>();
+
+        if(currentCohortEntry != null) {
+            cohortEntries.add(currentCohortEntry);
+            cohortCache.remove(currentCohortEntry.getTransactionID());
+            currentCohortEntry = null;
+        }
+
+        for(CohortEntry cohortEntry: queuedCohortEntries) {
+            cohortEntries.add(cohortEntry);
+            cohortCache.remove(cohortEntry.getTransactionID());
+        }
+
+        queuedCohortEntries.clear();
+        return cohortEntries;
+    }
+
+    Collection<Object> convertPendingTransactionsToMessages(final int maxModificationsPerBatch) {
+        if(currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
+            return Collections.emptyList();
+        }
+
+        Collection<Object> messages = new ArrayList<>();
+        List<CohortEntry> cohortEntries = getAndClearPendingCohortEntries();
+        for(CohortEntry cohortEntry: cohortEntries) {
+            if(cohortEntry.isExpired(cacheExpiryTimeoutInMillis) || cohortEntry.isAborted()) {
+                continue;
             }
-        } catch (InterruptedException | ExecutionException e) {
-            log.debug("{}: An exception occurred during canCommit: {}", name, e);
 
-            // Remove the entry from the cache now since the Tx will be aborted.
-            removeCohortEntry(cohortEntry.getTransactionID());
-            cohortEntry.getCanCommitSender().tell(new Status.Failure(e), cohortEntry.getShard());
+            final LinkedList<BatchedModifications> newModifications = new LinkedList<>();
+            cohortEntry.getDataTreeModification().applyToCursor(new AbstractBatchedModificationsCursor() {
+                @Override
+                protected BatchedModifications getModifications() {
+                    if(newModifications.isEmpty() ||
+                            newModifications.getLast().getModifications().size() >= maxModificationsPerBatch) {
+                        newModifications.add(new BatchedModifications(cohortEntry.getTransactionID(),
+                                cohortEntry.getClientVersion()));
         }
+
+                    return newModifications.getLast();
+                }
+            });
+
+            if(!newModifications.isEmpty()) {
+                BatchedModifications last = newModifications.getLast();
+                last.setDoCommitOnReady(cohortEntry.isDoImmediateCommit());
+                last.setReady(true);
+                last.setTotalMessagesSent(newModifications.size());
+                messages.addAll(newModifications);
+
+                if(!cohortEntry.isDoImmediateCommit() && cohortEntry.getState() == CohortEntry.State.CAN_COMMITTED) {
+                    messages.add(new CanCommitTransaction(cohortEntry.getTransactionID(),
+                            cohortEntry.getClientVersion()));
+                }
+
+                if(!cohortEntry.isDoImmediateCommit() && cohortEntry.getState() == CohortEntry.State.PRE_COMMITTED) {
+                    messages.add(new CommitTransaction(cohortEntry.getTransactionID(),
+                            cohortEntry.getClientVersion()));
+                }
+            }
+        }
+
+        return messages;
     }
 
     /**
@@ -240,7 +574,7 @@ public class ShardCommitCoordinator {
      * @return the current CohortEntry or null if the given transaction ID does not match the
      *         current entry.
      */
-    public CohortEntry getCohortEntryIfCurrent(String transactionID) {
+    CohortEntry getCohortEntryIfCurrent(Identifier transactionID) {
         if(isCurrentTransaction(transactionID)) {
             return currentCohortEntry;
         }
@@ -248,21 +582,15 @@ public class ShardCommitCoordinator {
         return null;
     }
 
-    public CohortEntry getCurrentCohortEntry() {
+    CohortEntry getCurrentCohortEntry() {
         return currentCohortEntry;
     }
 
-    public CohortEntry getAndRemoveCohortEntry(String transactionID) {
-        CohortEntry cohortEntry = cohortCache.getIfPresent(transactionID);
-        cohortCache.invalidate(transactionID);
-        return cohortEntry;
+    CohortEntry getAndRemoveCohortEntry(Identifier transactionID) {
+        return cohortCache.remove(transactionID);
     }
 
-    public void removeCohortEntry(String transactionID) {
-        cohortCache.invalidate(transactionID);
-    }
-
-    public boolean isCurrentTransaction(String transactionID) {
+    boolean isCurrentTransaction(Identifier transactionID) {
         return currentCohortEntry != null &&
                 currentCohortEntry.getTransactionID().equals(transactionID);
     }
@@ -276,106 +604,77 @@ public class ShardCommitCoordinator {
      * @param removeCohortEntry if true the CohortEntry for the transaction is also removed from
      *        the cache.
      */
-    public void currentTransactionComplete(String transactionID, boolean removeCohortEntry) {
+    void currentTransactionComplete(Identifier transactionID, boolean removeCohortEntry) {
         if(removeCohortEntry) {
-            removeCohortEntry(transactionID);
+            cohortCache.remove(transactionID);
         }
 
         if(isCurrentTransaction(transactionID)) {
-            // Dequeue the next cohort entry waiting in the queue.
-            currentCohortEntry = queuedCohortEntries.poll();
-            if(currentCohortEntry != null) {
-                currentCohortEntry.updateLastAccessTime();
-                doCanCommit(currentCohortEntry);
-            }
-        }
-    }
+            currentCohortEntry = null;
 
-    @VisibleForTesting
-    void setCohortDecorator(CohortDecorator cohortDecorator) {
-        this.cohortDecorator = cohortDecorator;
-    }
-
-
-    static class CohortEntry {
-        private final String transactionID;
-        private DOMStoreThreePhaseCommitCohort cohort;
-        private final MutableCompositeModification compositeModification;
-        private final DOMStoreWriteTransaction transaction;
-        private ActorRef canCommitSender;
-        private ActorRef shard;
-        private long lastAccessTime;
-
-        CohortEntry(String transactionID, DOMStoreWriteTransaction transaction) {
-            this.compositeModification = new MutableCompositeModification();
-            this.transaction = transaction;
-            this.transactionID = transactionID;
-        }
-
-        CohortEntry(String transactionID, DOMStoreThreePhaseCommitCohort cohort,
-                MutableCompositeModification compositeModification) {
-            this.transactionID = transactionID;
-            this.cohort = cohort;
-            this.compositeModification = compositeModification;
-            this.transaction = null;
-        }
+            log.debug("{}: currentTransactionComplete: {}", name, transactionID);
 
-        void updateLastAccessTime() {
-            lastAccessTime = System.currentTimeMillis();
-        }
-
-        long getLastAccessTime() {
-            return lastAccessTime;
-        }
-
-        String getTransactionID() {
-            return transactionID;
+            maybeProcessNextCohortEntry();
         }
+    }
 
-        DOMStoreThreePhaseCommitCohort getCohort() {
-            return cohort;
-        }
+    private void maybeProcessNextCohortEntry() {
+        // Check if there's a next cohort entry waiting in the queue and if it is ready to commit. Also
+        // clean out expired entries.
+        final Iterator<CohortEntry> iter = queuedCohortEntries.iterator();
+        while(iter.hasNext()) {
+            final CohortEntry next = iter.next();
+            if(next.isReadyToCommit()) {
+                if(currentCohortEntry == null) {
+                    if(log.isDebugEnabled()) {
+                        log.debug("{}: Next entry to canCommit {}", name, next);
+                    }
 
-        MutableCompositeModification getModification() {
-            return compositeModification;
-        }
+                    iter.remove();
+                    currentCohortEntry = next;
+                    currentCohortEntry.updateLastAccessTime();
+                    doCanCommit(currentCohortEntry);
+                }
 
-        void applyModifications(Iterable<Modification> modifications) {
-            for(Modification modification: modifications) {
-                compositeModification.addModification(modification);
-                modification.apply(transaction);
+                break;
+            } else if(next.isExpired(cacheExpiryTimeoutInMillis)) {
+                log.warn("{}: canCommit for transaction {} was not received within {} ms - entry removed from cache",
+                        name, next.getTransactionID(), cacheExpiryTimeoutInMillis);
+            } else if(!next.isAborted()) {
+                break;
             }
-        }
 
-        void ready(CohortDecorator cohortDecorator) {
-            Preconditions.checkState(cohort == null, "cohort was already set");
+            iter.remove();
+            cohortCache.remove(next.getTransactionID());
+        }
 
-            cohort = transaction.ready();
+        maybeRunOperationOnPendingTransactionsComplete();
+    }
 
-            if(cohortDecorator != null) {
-                // Call the hook for unit tests.
-                cohort = cohortDecorator.decorate(transactionID, cohort);
-            }
-        }
+    void cleanupExpiredCohortEntries() {
+        maybeProcessNextCohortEntry();
+    }
 
-        ActorRef getCanCommitSender() {
-            return canCommitSender;
-        }
+    void setRunOnPendingTransactionsComplete(Runnable operation) {
+        runOnPendingTransactionsComplete = operation;
+        maybeRunOperationOnPendingTransactionsComplete();
+    }
 
-        void setCanCommitSender(ActorRef canCommitSender) {
-            this.canCommitSender = canCommitSender;
-        }
+    private void maybeRunOperationOnPendingTransactionsComplete() {
+        if(runOnPendingTransactionsComplete != null && currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
+            log.debug("{}: Pending transactions complete - running operation {}", name, runOnPendingTransactionsComplete);
 
-        ActorRef getShard() {
-            return shard;
+            runOnPendingTransactionsComplete.run();
+            runOnPendingTransactionsComplete = null;
         }
+    }
 
-        void setShard(ActorRef shard) {
-            this.shard = shard;
-        }
+    @VisibleForTesting
+    void setCohortDecorator(CohortDecorator cohortDecorator) {
+        this.cohortDecorator = cohortDecorator;
+    }
 
-        boolean hasModifications(){
-            return compositeModification.getModifications().size() > 0;
-        }
+   void processCohortRegistryCommand(ActorRef sender, CohortRegistryCommand message) {
+        cohortRegistry.process(sender, message);
     }
 }