import akka.actor.ActorRef;
import akka.actor.Status;
+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.LinkedList;
import java.util.Queue;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.TimeUnit;
-import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
+import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort;
+import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
+import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransactionReply;
+import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
+import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
import org.opendaylight.controller.cluster.datastore.modification.Modification;
-import org.opendaylight.controller.sal.core.spi.data.DOMStoreThreePhaseCommitCohort;
+import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
+import org.opendaylight.controller.md.sal.common.api.data.TransactionCommitFailedException;
import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
/**
* Coordinates commits for a shard ensuring only one concurrent 3-phase commit.
*/
public class ShardCommitCoordinator {
- private static final Logger LOG = LoggerFactory.getLogger(ShardCommitCoordinator.class);
+ // Interface hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
+ public interface CohortDecorator {
+ ShardDataTreeCohort decorate(String transactionID, ShardDataTreeCohort actual);
+ }
private final Cache<String, CohortEntry> cohortCache;
private CohortEntry currentCohortEntry;
+ private final ShardDataTree dataTree;
+
private final Queue<CohortEntry> queuedCohortEntries;
- private final int queueCapacity;
+ private int queueCapacity;
+
+ private final Logger log;
+
+ private final String name;
+
+ 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());
+ }
+ }
+ };
+
+ // This is a hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
+ private CohortDecorator cohortDecorator;
+
+ private ReadyTransactionReply readyTransactionReply;
- public ShardCommitCoordinator(long cacheExpiryTimeoutInSec, int queueCapacity) {
- cohortCache = CacheBuilder.newBuilder().expireAfterAccess(
- cacheExpiryTimeoutInSec, TimeUnit.SECONDS).build();
+ public ShardCommitCoordinator(ShardDataTree dataTree,
+ long cacheExpiryTimeoutInSec, int queueCapacity, ActorRef shardActor, Logger log, String name) {
this.queueCapacity = queueCapacity;
+ this.log = log;
+ this.name = name;
+ this.dataTree = Preconditions.checkNotNull(dataTree);
+
+ cohortCache = CacheBuilder.newBuilder().expireAfterAccess(cacheExpiryTimeoutInSec, TimeUnit.SECONDS).
+ removalListener(cacheRemovalListener).build();
// 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<>();
}
+ public void setQueueCapacity(int queueCapacity) {
+ this.queueCapacity = queueCapacity;
+ }
+
+ private ReadyTransactionReply readyTransactionReply(Shard shard) {
+ if(readyTransactionReply == null) {
+ readyTransactionReply = new ReadyTransactionReply(Serialization.serializedActorPath(shard.self()));
+ }
+
+ return readyTransactionReply;
+ }
+
/**
- * This method caches a 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 modification made by the transaction
+ * 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.
*/
- public void transactionReady(String transactionID, DOMStoreThreePhaseCommitCohort cohort,
- Modification modification) {
+ public void handleForwardedReadyTransaction(ForwardedReadyTransaction ready, ActorRef sender, Shard shard) {
+ log.debug("{}: Readying transaction {}, client version {}", name,
+ ready.getTransactionID(), ready.getTxnClientVersion());
+
+ CohortEntry cohortEntry = new CohortEntry(ready.getTransactionID(), ready.getCohort(),
+ (MutableCompositeModification) ready.getModification());
+ cohortCache.put(ready.getTransactionID(), cohortEntry);
+
+ if(ready.getTxnClientVersion() < DataStoreVersions.LITHIUM_VERSION) {
+ // Return our actor path as we'll handle the three phase commit except if the Tx client
+ // version < Helium-1 version which means the Tx was initiated by a base Helium version node.
+ // In that case, the subsequent 3-phase commit messages won't contain the transactionId so to
+ // maintain backwards compatibility, we create a separate cohort actor to provide the compatible behavior.
+ ActorRef replyActorPath = shard.self();
+ if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
+ log.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", name);
+ replyActorPath = shard.getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
+ ready.getTransactionID()));
+ }
- cohortCache.put(transactionID, new CohortEntry(transactionID, cohort, modification));
+ ReadyTransactionReply readyTransactionReply =
+ new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath),
+ ready.getTxnClientVersion());
+ sender.tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
+ readyTransactionReply, shard.self());
+ } else {
+ 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());
+ }
+ }
}
/**
- * This method handles the canCommit phase for a transaction.
+ * This method handles a BatchedModifications message for a transaction being prepared directly on the
+ * Shard actor instead of via a ShardTransaction actor. If there's no currently cached
+ * 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 canCommit the CanCommitTransaction message
- * @param sender the actor that sent the message
- * @param shard the transaction's shard actor
+ * @param batched the BatchedModifications
+ * @param shardActor the transaction's shard actor
+ *
+ * @throws ExecutionException if an error occurs loading the cache
*/
- 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 {}",
- transactionID, shard.path());
+ boolean handleBatchedModifications(BatchedModifications batched, ActorRef sender, Shard shard)
+ throws ExecutionException {
+ CohortEntry cohortEntry = cohortCache.getIfPresent(batched.getTransactionID());
+ if(cohortEntry == null) {
+ cohortEntry = new CohortEntry(batched.getTransactionID(),
+ dataTree.newReadWriteTransaction(batched.getTransactionID(),
+ batched.getTransactionChainID()));
+ cohortCache.put(batched.getTransactionID(), cohortEntry);
}
- // Lookup the cohort entry that was cached previously (or should have been) by
- // transactionReady (via the ForwardedReadyTransaction message).
- final CohortEntry cohortEntry = cohortCache.getIfPresent(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("No cohort entry found for transaction %s", transactionID));
- LOG.error(ex.getMessage());
- sender.tell(new Status.Failure(ex), shard);
- return;
+ if(log.isDebugEnabled()) {
+ log.debug("{}: Applying {} batched modifications for Tx {}", name,
+ batched.getModifications().size(), batched.getTransactionID());
}
- cohortEntry.setCanCommitSender(sender);
- cohortEntry.setShard(shard);
+ cohortEntry.applyModifications(batched.getModifications());
+
+ if(batched.isReady()) {
+ if(log.isDebugEnabled()) {
+ log.debug("{}: Readying Tx {}, client version {}", name,
+ batched.getTransactionID(), batched.getVersion());
+ }
+
+ cohortEntry.ready(cohortDecorator, batched.isDoCommitOnReady());
+
+ 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());
+ }
+
+ return batched.isReady();
+ }
+
+ private void handleCanCommit(CohortEntry cohortEntry) {
+ String transactionID = cohortEntry.getTransactionID();
+
+ if(log.isDebugEnabled()) {
+ log.debug("{}: Processing canCommit for transaction {} for shard {}",
+ name, transactionID, cohortEntry.getShard().self().path());
+ }
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 {}",
- currentCohortEntry.getTransactionID(), transactionID);
+ log.debug("{}: Transaction {} is already in progress - queueing transaction {}",
+ name, currentCohortEntry.getTransactionID(), transactionID);
if(queuedCohortEntries.size() < queueCapacity) {
queuedCohortEntries.offer(cohortEntry);
removeCohortEntry(transactionID);
RuntimeException ex = new RuntimeException(
- String.format("Could not enqueue transaction %s - the maximum commit queue"+
+ String.format("%s: Could not enqueue transaction %s - the maximum commit queue"+
" capacity %d has been reached.",
- transactionID, queueCapacity));
- LOG.error(ex.getMessage());
- sender.tell(new Status.Failure(ex), shard);
+ name, transactionID, queueCapacity));
+ log.error(ex.getMessage());
+ cohortEntry.getReplySender().tell(new Status.Failure(ex), cohortEntry.getShard().self());
}
} else {
// No Tx commit currently in progress - make this the current entry and proceed with
}
}
+ /**
+ * This method handles the canCommit phase for a transaction.
+ *
+ * @param canCommit the CanCommitTransaction message
+ * @param sender the actor that sent the message
+ * @param shard the transaction's shard actor
+ */
+ public void handleCanCommit(String 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);
+ 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.self());
+ return;
+ }
+
+ cohortEntry.setReplySender(sender);
+ cohortEntry.setShard(shard);
+
+ handleCanCommit(cohortEntry);
+ }
+
private void doCanCommit(final CohortEntry cohortEntry) {
+ boolean canCommit = false;
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();
+ canCommit = cohortEntry.getCohort().canCommit().get();
+
+ if(cohortEntry.isDoImmediateCommit()) {
+ if(canCommit) {
+ doCommit(cohortEntry);
+ } else {
+ cohortEntry.getReplySender().tell(new Status.Failure(new TransactionCommitFailedException(
+ "Can Commit failed, no detailed cause available.")), cohortEntry.getShard().self());
+ }
+ } else {
+ cohortEntry.getReplySender().tell(
+ canCommit ? CanCommitTransactionReply.YES.toSerializable() :
+ CanCommitTransactionReply.NO.toSerializable(), cohortEntry.getShard().self());
+ }
+ } catch (Exception e) {
+ log.debug("{}: An exception occurred during canCommit: {}", name, e);
- cohortEntry.getCanCommitSender().tell(
- canCommit ? CanCommitTransactionReply.YES.toSerializable() :
- CanCommitTransactionReply.NO.toSerializable(), cohortEntry.getShard());
+ Throwable failure = e;
+ if(e instanceof ExecutionException) {
+ failure = e.getCause();
+ }
+ cohortEntry.getReplySender().tell(new Status.Failure(failure), cohortEntry.getShard().self());
+ } finally {
if(!canCommit) {
- // Remove the entry from the cache now since the Tx will be aborted.
- removeCohortEntry(cohortEntry.getTransactionID());
+ // Remove the entry from the cache now.
+ currentTransactionComplete(cohortEntry.getTransactionID(), true);
}
- } catch (InterruptedException | ExecutionException e) {
- LOG.debug("An exception occurred during canCommit", e);
+ }
+ }
+
+ 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 {
+ // 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.
+ cohortEntry.getCohort().preCommit().get();
- // 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());
+ cohortEntry.getShard().continueCommit(cohortEntry);
+
+ cohortEntry.updateLastAccessTime();
+
+ success = true;
+ } catch (Exception e) {
+ log.error("{} An exception occurred while preCommitting transaction {}",
+ name, cohortEntry.getTransactionID(), e);
+ cohortEntry.getReplySender().tell(new akka.actor.Status.Failure(e), cohortEntry.getShard().self());
+
+ currentTransactionComplete(cohortEntry.getTransactionID(), true);
+ }
+
+ return success;
+ }
+
+ boolean handleCommit(final String 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 akka.actor.Status.Failure(ex), shard.self());
+ return false;
}
+
+ return doCommit(cohortEntry);
}
/**
// Dequeue the next cohort entry waiting in the queue.
currentCohortEntry = queuedCohortEntries.poll();
if(currentCohortEntry != null) {
+ currentCohortEntry.updateLastAccessTime();
doCanCommit(currentCohortEntry);
}
}
}
+ @VisibleForTesting
+ void setCohortDecorator(CohortDecorator cohortDecorator) {
+ this.cohortDecorator = cohortDecorator;
+ }
+
+
static class CohortEntry {
private final String transactionID;
- private final DOMStoreThreePhaseCommitCohort cohort;
- private final Modification modification;
- private ActorRef canCommitSender;
- private ActorRef shard;
+ private ShardDataTreeCohort cohort;
+ private final ReadWriteShardDataTreeTransaction transaction;
+ private ActorRef replySender;
+ private Shard shard;
private long lastAccessTime;
+ private boolean doImmediateCommit;
- CohortEntry(String transactionID, DOMStoreThreePhaseCommitCohort cohort,
- Modification modification) {
+ CohortEntry(String transactionID, ReadWriteShardDataTreeTransaction transaction) {
+ this.transaction = Preconditions.checkNotNull(transaction);
+ this.transactionID = transactionID;
+ }
+
+ CohortEntry(String transactionID, ShardDataTreeCohort cohort,
+ MutableCompositeModification compositeModification) {
this.transactionID = transactionID;
this.cohort = cohort;
- this.modification = modification;
+ this.transaction = null;
}
void updateLastAccessTime() {
return transactionID;
}
- DOMStoreThreePhaseCommitCohort getCohort() {
+ ShardDataTreeCohort getCohort() {
return cohort;
}
- Modification getModification() {
- return modification;
+ void applyModifications(Iterable<Modification> modifications) {
+ for (Modification modification : modifications) {
+ modification.apply(transaction.getSnapshot());
+ }
+ }
+
+ void ready(CohortDecorator cohortDecorator, boolean doImmediateCommit) {
+ Preconditions.checkState(cohort == null, "cohort was already set");
+
+ setDoImmediateCommit(doImmediateCommit);
+
+ cohort = transaction.ready();
+
+ if(cohortDecorator != null) {
+ // Call the hook for unit tests.
+ cohort = cohortDecorator.decorate(transactionID, cohort);
+ }
+ }
+
+ boolean isDoImmediateCommit() {
+ return doImmediateCommit;
+ }
+
+ void setDoImmediateCommit(boolean doImmediateCommit) {
+ this.doImmediateCommit = doImmediateCommit;
}
- ActorRef getCanCommitSender() {
- return canCommitSender;
+ ActorRef getReplySender() {
+ return replySender;
}
- void setCanCommitSender(ActorRef canCommitSender) {
- this.canCommitSender = canCommitSender;
+ void setReplySender(ActorRef replySender) {
+ this.replySender = replySender;
}
- ActorRef getShard() {
+ Shard getShard() {
return shard;
}
- void setShard(ActorRef shard) {
+ void setShard(Shard shard) {
this.shard = shard;
}
}