2 * Copyright (c) 2014 Brocade Communications Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.cluster.datastore;
10 import akka.actor.ActorRef;
11 import akka.actor.Status;
12 import akka.serialization.Serialization;
13 import com.google.common.annotations.VisibleForTesting;
14 import com.google.common.base.Preconditions;
15 import com.google.common.base.Stopwatch;
16 import java.util.HashMap;
17 import java.util.Iterator;
18 import java.util.LinkedList;
20 import java.util.Queue;
21 import java.util.concurrent.ExecutionException;
22 import java.util.concurrent.TimeUnit;
23 import org.opendaylight.controller.cluster.datastore.compat.BackwardsCompatibleThreePhaseCommitCohort;
24 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
25 import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
26 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransactionReply;
27 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
28 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
29 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
30 import org.opendaylight.controller.cluster.datastore.modification.Modification;
31 import org.opendaylight.controller.cluster.datastore.modification.MutableCompositeModification;
32 import org.opendaylight.controller.md.sal.common.api.data.TransactionCommitFailedException;
33 import org.slf4j.Logger;
36 * Coordinates commits for a shard ensuring only one concurrent 3-phase commit.
38 * @author Thomas Pantelis
40 public class ShardCommitCoordinator {
42 // Interface hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
43 public interface CohortDecorator {
44 ShardDataTreeCohort decorate(String transactionID, ShardDataTreeCohort actual);
47 private final Map<String, CohortEntry> cohortCache = new HashMap<>();
49 private CohortEntry currentCohortEntry;
51 private final ShardDataTree dataTree;
53 // We use a LinkedList here to avoid synchronization overhead with concurrent queue impls
54 // since this should only be accessed on the shard's dispatcher.
55 private final Queue<CohortEntry> queuedCohortEntries = new LinkedList<>();
57 private int queueCapacity;
59 private final Logger log;
61 private final String name;
63 private final long cacheExpiryTimeoutInMillis;
65 // This is a hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
66 private CohortDecorator cohortDecorator;
68 private ReadyTransactionReply readyTransactionReply;
70 public ShardCommitCoordinator(ShardDataTree dataTree,
71 long cacheExpiryTimeoutInMillis, int queueCapacity, ActorRef shardActor, Logger log, String name) {
73 this.queueCapacity = queueCapacity;
76 this.dataTree = Preconditions.checkNotNull(dataTree);
77 this.cacheExpiryTimeoutInMillis = cacheExpiryTimeoutInMillis;
80 public void setQueueCapacity(int queueCapacity) {
81 this.queueCapacity = queueCapacity;
84 private ReadyTransactionReply readyTransactionReply(Shard shard) {
85 if(readyTransactionReply == null) {
86 readyTransactionReply = new ReadyTransactionReply(Serialization.serializedActorPath(shard.self()));
89 return readyTransactionReply;
92 private boolean queueCohortEntry(CohortEntry cohortEntry, ActorRef sender, Shard shard) {
93 if(queuedCohortEntries.size() < queueCapacity) {
94 queuedCohortEntries.offer(cohortEntry);
97 cohortCache.remove(cohortEntry.getTransactionID());
99 RuntimeException ex = new RuntimeException(
100 String.format("%s: Could not enqueue transaction %s - the maximum commit queue"+
101 " capacity %d has been reached.",
102 name, cohortEntry.getTransactionID(), queueCapacity));
103 log.error(ex.getMessage());
104 sender.tell(new Status.Failure(ex), shard.self());
110 * This method is called to ready a transaction that was prepared by ShardTransaction actor. It caches
111 * the prepared cohort entry for the given transactions ID in preparation for the subsequent 3-phase commit.
113 public void handleForwardedReadyTransaction(ForwardedReadyTransaction ready, ActorRef sender, Shard shard) {
114 log.debug("{}: Readying transaction {}, client version {}", name,
115 ready.getTransactionID(), ready.getTxnClientVersion());
117 CohortEntry cohortEntry = new CohortEntry(ready.getTransactionID(), ready.getCohort(),
118 (MutableCompositeModification) ready.getModification());
119 cohortCache.put(ready.getTransactionID(), cohortEntry);
121 if(!queueCohortEntry(cohortEntry, sender, shard)) {
125 if(ready.getTxnClientVersion() < DataStoreVersions.LITHIUM_VERSION) {
126 // Return our actor path as we'll handle the three phase commit except if the Tx client
127 // version < Helium-1 version which means the Tx was initiated by a base Helium version node.
128 // In that case, the subsequent 3-phase commit messages won't contain the transactionId so to
129 // maintain backwards compatibility, we create a separate cohort actor to provide the compatible behavior.
130 ActorRef replyActorPath = shard.self();
131 if(ready.getTxnClientVersion() < DataStoreVersions.HELIUM_1_VERSION) {
132 log.debug("{}: Creating BackwardsCompatibleThreePhaseCommitCohort", name);
133 replyActorPath = shard.getContext().actorOf(BackwardsCompatibleThreePhaseCommitCohort.props(
134 ready.getTransactionID()));
137 ReadyTransactionReply readyTransactionReply =
138 new ReadyTransactionReply(Serialization.serializedActorPath(replyActorPath),
139 ready.getTxnClientVersion());
140 sender.tell(ready.isReturnSerialized() ? readyTransactionReply.toSerializable() :
141 readyTransactionReply, shard.self());
143 if(ready.isDoImmediateCommit()) {
144 cohortEntry.setDoImmediateCommit(true);
145 cohortEntry.setReplySender(sender);
146 cohortEntry.setShard(shard);
147 handleCanCommit(cohortEntry);
149 // The caller does not want immediate commit - the 3-phase commit will be coordinated by the
150 // front-end so send back a ReadyTransactionReply with our actor path.
151 sender.tell(readyTransactionReply(shard), shard.self());
157 * This method handles a BatchedModifications message for a transaction being prepared directly on the
158 * Shard actor instead of via a ShardTransaction actor. If there's no currently cached
159 * DOMStoreWriteTransaction, one is created. The batched modifications are applied to the write Tx. If
160 * the BatchedModifications is ready to commit then a DOMStoreThreePhaseCommitCohort is created.
162 * @param batched the BatchedModifications
163 * @param shardActor the transaction's shard actor
165 * @throws ExecutionException if an error occurs loading the cache
167 void handleBatchedModifications(BatchedModifications batched, ActorRef sender, Shard shard)
168 throws ExecutionException {
169 CohortEntry cohortEntry = cohortCache.get(batched.getTransactionID());
170 if(cohortEntry == null) {
171 cohortEntry = new CohortEntry(batched.getTransactionID(),
172 dataTree.newReadWriteTransaction(batched.getTransactionID(),
173 batched.getTransactionChainID()));
174 cohortCache.put(batched.getTransactionID(), cohortEntry);
177 if(log.isDebugEnabled()) {
178 log.debug("{}: Applying {} batched modifications for Tx {}", name,
179 batched.getModifications().size(), batched.getTransactionID());
182 cohortEntry.applyModifications(batched.getModifications());
184 if(batched.isReady()) {
185 if(!queueCohortEntry(cohortEntry, sender, shard)) {
189 if(log.isDebugEnabled()) {
190 log.debug("{}: Readying Tx {}, client version {}", name,
191 batched.getTransactionID(), batched.getVersion());
194 cohortEntry.ready(cohortDecorator, batched.isDoCommitOnReady());
196 if(batched.isDoCommitOnReady()) {
197 cohortEntry.setReplySender(sender);
198 cohortEntry.setShard(shard);
199 handleCanCommit(cohortEntry);
201 sender.tell(readyTransactionReply(shard), shard.self());
204 sender.tell(new BatchedModificationsReply(batched.getModifications().size()), shard.self());
209 * This method handles {@link ReadyLocalTransaction} message. All transaction modifications have
210 * been prepared beforehand by the sender and we just need to drive them through into the dataTree.
216 void handleReadyLocalTransaction(ReadyLocalTransaction message, ActorRef sender, Shard shard) {
217 final ShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(dataTree, message.getModification());
218 final CohortEntry cohortEntry = new CohortEntry(message.getTransactionID(), cohort);
219 cohortCache.put(message.getTransactionID(), cohortEntry);
220 cohortEntry.setDoImmediateCommit(message.isDoCommitOnReady());
222 if(!queueCohortEntry(cohortEntry, sender, shard)) {
226 log.debug("{}: Applying local modifications for Tx {}", name, message.getTransactionID());
228 if (message.isDoCommitOnReady()) {
229 cohortEntry.setReplySender(sender);
230 cohortEntry.setShard(shard);
231 handleCanCommit(cohortEntry);
233 sender.tell(readyTransactionReply(shard), shard.self());
237 private void handleCanCommit(CohortEntry cohortEntry) {
238 String transactionID = cohortEntry.getTransactionID();
240 cohortEntry.updateLastAccessTime();
242 if(currentCohortEntry != null) {
243 // There's already a Tx commit in progress so we can't process this entry yet - but it's in the
244 // queue and will get processed after all prior entries complete.
246 if(log.isDebugEnabled()) {
247 log.debug("{}: Commit for Tx {} already in progress - skipping canCommit for {} for now",
248 name, currentCohortEntry.getTransactionID(), transactionID);
254 // No Tx commit currently in progress - check if this entry is the next one in the queue, If so make
255 // it the current entry and proceed with canCommit.
256 // Purposely checking reference equality here.
257 if(queuedCohortEntries.peek() == cohortEntry) {
258 currentCohortEntry = queuedCohortEntries.poll();
259 doCanCommit(currentCohortEntry);
261 if(log.isDebugEnabled()) {
262 log.debug("{}: Tx {} is the next pending canCommit - skipping {} for now",
263 name, queuedCohortEntries.peek().getTransactionID(), transactionID);
269 * This method handles the canCommit phase for a transaction.
271 * @param canCommit the CanCommitTransaction message
272 * @param sender the actor that sent the message
273 * @param shard the transaction's shard actor
275 public void handleCanCommit(String transactionID, final ActorRef sender, final Shard shard) {
276 // Lookup the cohort entry that was cached previously (or should have been) by
277 // transactionReady (via the ForwardedReadyTransaction message).
278 final CohortEntry cohortEntry = cohortCache.get(transactionID);
279 if(cohortEntry == null) {
280 // Either canCommit was invoked before ready(shouldn't happen) or a long time passed
281 // between canCommit and ready and the entry was expired from the cache.
282 IllegalStateException ex = new IllegalStateException(
283 String.format("%s: No cohort entry found for transaction %s", name, transactionID));
284 log.error(ex.getMessage());
285 sender.tell(new Status.Failure(ex), shard.self());
289 cohortEntry.setReplySender(sender);
290 cohortEntry.setShard(shard);
292 handleCanCommit(cohortEntry);
295 private void doCanCommit(final CohortEntry cohortEntry) {
296 boolean canCommit = false;
298 // We block on the future here so we don't have to worry about possibly accessing our
299 // state on a different thread outside of our dispatcher. Also, the data store
300 // currently uses a same thread executor anyway.
301 canCommit = cohortEntry.getCohort().canCommit().get();
303 log.debug("{}: canCommit for {}: {}", name, cohortEntry.getTransactionID(), canCommit);
305 if(cohortEntry.isDoImmediateCommit()) {
307 doCommit(cohortEntry);
309 cohortEntry.getReplySender().tell(new Status.Failure(new TransactionCommitFailedException(
310 "Can Commit failed, no detailed cause available.")), cohortEntry.getShard().self());
313 cohortEntry.getReplySender().tell(
314 canCommit ? CanCommitTransactionReply.YES.toSerializable() :
315 CanCommitTransactionReply.NO.toSerializable(), cohortEntry.getShard().self());
317 } catch (Exception e) {
318 log.debug("{}: An exception occurred during canCommit", name, e);
320 Throwable failure = e;
321 if(e instanceof ExecutionException) {
322 failure = e.getCause();
325 cohortEntry.getReplySender().tell(new Status.Failure(failure), cohortEntry.getShard().self());
328 // Remove the entry from the cache now.
329 currentTransactionComplete(cohortEntry.getTransactionID(), true);
334 private boolean doCommit(CohortEntry cohortEntry) {
335 log.debug("{}: Committing transaction {}", name, cohortEntry.getTransactionID());
337 boolean success = false;
339 // We perform the preCommit phase here atomically with the commit phase. This is an
340 // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
341 // coordination of preCommit across shards in case of failure but preCommit should not
342 // normally fail since we ensure only one concurrent 3-phase commit.
345 // We block on the future here so we don't have to worry about possibly accessing our
346 // state on a different thread outside of our dispatcher. Also, the data store
347 // currently uses a same thread executor anyway.
348 cohortEntry.getCohort().preCommit().get();
350 cohortEntry.getShard().continueCommit(cohortEntry);
352 cohortEntry.updateLastAccessTime();
355 } catch (Exception e) {
356 log.error("{} An exception occurred while preCommitting transaction {}",
357 name, cohortEntry.getTransactionID(), e);
358 cohortEntry.getReplySender().tell(new akka.actor.Status.Failure(e), cohortEntry.getShard().self());
360 currentTransactionComplete(cohortEntry.getTransactionID(), true);
366 boolean handleCommit(final String transactionID, final ActorRef sender, final Shard shard) {
367 // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
369 final CohortEntry cohortEntry = getCohortEntryIfCurrent(transactionID);
370 if(cohortEntry == null) {
371 // We're not the current Tx - the Tx was likely expired b/c it took too long in
372 // between the canCommit and commit messages.
373 IllegalStateException ex = new IllegalStateException(
374 String.format("%s: Cannot commit transaction %s - it is not the current transaction",
375 name, transactionID));
376 log.error(ex.getMessage());
377 sender.tell(new akka.actor.Status.Failure(ex), shard.self());
381 cohortEntry.setReplySender(sender);
382 return doCommit(cohortEntry);
386 * Returns the cohort entry for the Tx commit currently in progress if the given transaction ID
387 * matches the current entry.
389 * @param transactionID the ID of the transaction
390 * @return the current CohortEntry or null if the given transaction ID does not match the
393 public CohortEntry getCohortEntryIfCurrent(String transactionID) {
394 if(isCurrentTransaction(transactionID)) {
395 return currentCohortEntry;
401 public CohortEntry getCurrentCohortEntry() {
402 return currentCohortEntry;
405 public CohortEntry getAndRemoveCohortEntry(String transactionID) {
406 return cohortCache.remove(transactionID);
409 public boolean isCurrentTransaction(String transactionID) {
410 return currentCohortEntry != null &&
411 currentCohortEntry.getTransactionID().equals(transactionID);
415 * This method is called when a transaction is complete, successful or not. If the given
416 * given transaction ID matches the current in-progress transaction, the next cohort entry,
417 * if any, is dequeued and processed.
419 * @param transactionID the ID of the completed transaction
420 * @param removeCohortEntry if true the CohortEntry for the transaction is also removed from
423 public void currentTransactionComplete(String transactionID, boolean removeCohortEntry) {
424 if(removeCohortEntry) {
425 cohortCache.remove(transactionID);
428 if(isCurrentTransaction(transactionID)) {
429 currentCohortEntry = null;
431 log.debug("{}: currentTransactionComplete: {}", name, transactionID);
433 maybeProcessNextCohortEntry();
437 private void maybeProcessNextCohortEntry() {
438 // Check if there's a next cohort entry waiting in the queue and if it is ready to commit. Also
439 // clean out expired entries.
440 Iterator<CohortEntry> iter = queuedCohortEntries.iterator();
441 while(iter.hasNext()) {
442 CohortEntry next = iter.next();
443 if(next.isReadyToCommit()) {
444 if(currentCohortEntry == null) {
445 if(log.isDebugEnabled()) {
446 log.debug("{}: Next entry to canCommit {}", name, next);
450 currentCohortEntry = next;
451 currentCohortEntry.updateLastAccessTime();
452 doCanCommit(currentCohortEntry);
456 } else if(next.isExpired(cacheExpiryTimeoutInMillis)) {
457 log.warn("{}: canCommit for transaction {} was not received within {} ms - entry removed from cache",
458 name, next.getTransactionID(), cacheExpiryTimeoutInMillis);
461 cohortCache.remove(next.getTransactionID());
468 void cleanupExpiredCohortEntries() {
469 maybeProcessNextCohortEntry();
473 void setCohortDecorator(CohortDecorator cohortDecorator) {
474 this.cohortDecorator = cohortDecorator;
477 static class CohortEntry {
478 private final String transactionID;
479 private ShardDataTreeCohort cohort;
480 private final ReadWriteShardDataTreeTransaction transaction;
481 private ActorRef replySender;
483 private boolean doImmediateCommit;
484 private final Stopwatch lastAccessTimer = Stopwatch.createStarted();
486 CohortEntry(String transactionID, ReadWriteShardDataTreeTransaction transaction) {
487 this.transaction = Preconditions.checkNotNull(transaction);
488 this.transactionID = transactionID;
491 CohortEntry(String transactionID, ShardDataTreeCohort cohort,
492 MutableCompositeModification compositeModification) {
493 this.transactionID = transactionID;
494 this.cohort = cohort;
495 this.transaction = null;
498 CohortEntry(String transactionID, ShardDataTreeCohort cohort) {
499 this.transactionID = transactionID;
500 this.cohort = cohort;
501 this.transaction = null;
504 void updateLastAccessTime() {
505 lastAccessTimer.reset();
506 lastAccessTimer.start();
509 String getTransactionID() {
510 return transactionID;
513 ShardDataTreeCohort getCohort() {
517 void applyModifications(Iterable<Modification> modifications) {
518 for (Modification modification : modifications) {
519 modification.apply(transaction.getSnapshot());
523 void ready(CohortDecorator cohortDecorator, boolean doImmediateCommit) {
524 Preconditions.checkState(cohort == null, "cohort was already set");
526 setDoImmediateCommit(doImmediateCommit);
528 cohort = transaction.ready();
530 if(cohortDecorator != null) {
531 // Call the hook for unit tests.
532 cohort = cohortDecorator.decorate(transactionID, cohort);
536 boolean isReadyToCommit() {
537 return replySender != null;
540 boolean isExpired(long expireTimeInMillis) {
541 return lastAccessTimer.elapsed(TimeUnit.MILLISECONDS) >= expireTimeInMillis;
544 boolean isDoImmediateCommit() {
545 return doImmediateCommit;
548 void setDoImmediateCommit(boolean doImmediateCommit) {
549 this.doImmediateCommit = doImmediateCommit;
552 ActorRef getReplySender() {
556 void setReplySender(ActorRef replySender) {
557 this.replySender = replySender;
564 void setShard(Shard shard) {
569 public String toString() {
570 StringBuilder builder = new StringBuilder();
571 builder.append("CohortEntry [transactionID=").append(transactionID).append(", doImmediateCommit=")
572 .append(doImmediateCommit).append("]");
573 return builder.toString();