private void updateTopologyId(TopologyAffectedCommand command) throws InterruptedException {
// set the topology id if it was not set before (ie. this is local command)
// TODO Make tx commands extend FlagAffectedCommand so we can use CACHE_MODE_LOCAL in
// TransactionTable.cleanupStaleTransactions
if (command.getTopologyId() == -1) {
CacheTopology cacheTopology = stateTransferManager.getCacheTopology();
if (cacheTopology != null) {
command.setTopologyId(cacheTopology.getTopologyId());
}
}
}
private void waitForStateTransfer(int expectedTopologyId, Cache... caches) {
waitForRehashToComplete(caches);
for (Cache c : caches) {
CacheTopology cacheTopology =
extractComponent(c, StateTransferManager.class).getCacheTopology();
assertEquals(
String.format(
"Wrong topology on cache %s, expected %d and got %s",
c, expectedTopologyId, cacheTopology),
expectedTopologyId,
cacheTopology.getTopologyId());
}
}
@Override
public boolean isAcceptable(Response response, Address sender) {
if (response.isSuccessful() && response.isValid()) {
ManagerStatusResponse value =
(ManagerStatusResponse) ((SuccessfulResponse) response).getResponseValue();
for (Entry<String, CacheStatusResponse> entry : value.getCaches().entrySet()) {
CacheStatusResponse csr = entry.getValue();
CacheTopology cacheTopology = csr.getCacheTopology();
CacheTopology stableTopology = csr.getStableTopology();
CacheTopology replaceCacheTopology = seenTopologies.get(cacheTopology);
if (replaceCacheTopology == null) {
seenTopologies.put(cacheTopology, cacheTopology);
replaceCacheTopology = cacheTopology;
}
CacheTopology replaceStableTopology = null;
// If the don't equal check if we replace - note stableTopology can be null
if (!cacheTopology.equals(stableTopology)) {
replaceStableTopology = seenTopologies.get(stableTopology);
if (replaceStableTopology == null) {
seenTopologies.put(stableTopology, stableTopology);
}
} else {
// Since they were equal replace it with the cache topology we are going to use
replaceStableTopology =
replaceCacheTopology != null ? replaceCacheTopology : cacheTopology;
}
CacheJoinInfo info = csr.getCacheJoinInfo();
CacheJoinInfo replaceInfo = seenInfos.get(info);
if (replaceInfo == null) {
seenInfos.put(info, info);
}
if (replaceCacheTopology != null
|| replaceStableTopology != null
|| replaceInfo != null) {
entry.setValue(
new CacheStatusResponse(
replaceInfo != null ? replaceInfo : info,
replaceCacheTopology != null ? replaceCacheTopology : cacheTopology,
replaceStableTopology != null ? replaceStableTopology : stableTopology,
csr.getAvailabilityMode()));
}
}
}
return true;
}
@Override
public void notifyTopologyChanged(
CacheTopology oldTopology, CacheTopology newTopology, int newTopologyId, boolean pre) {
if (!topologyChangedListeners.isEmpty()) {
EventImpl<K, V> e = EventImpl.createEvent(cache, TOPOLOGY_CHANGED);
e.setPre(pre);
if (oldTopology != null) {
e.setConsistentHashAtStart(oldTopology.getReadConsistentHash());
}
e.setConsistentHashAtEnd(newTopology.getWriteConsistentHash());
e.setNewTopologyId(newTopologyId);
for (CacheEntryListenerInvocation<K, V> listener : topologyChangedListeners)
listener.invoke(e);
}
}
private InboundTransferTask addTransfer(Address source, Set<Integer> segmentsFromSource) {
synchronized (this) {
segmentsFromSource.removeAll(
transfersBySegment.keySet()); // already in progress segments are excluded
if (!segmentsFromSource.isEmpty()) {
InboundTransferTask inboundTransfer =
new InboundTransferTask(
segmentsFromSource,
source,
cacheTopology.getTopologyId(),
this,
rpcManager,
commandsFactory,
timeout,
cacheName);
for (int segmentId : segmentsFromSource) {
transfersBySegment.put(segmentId, inboundTransfer);
}
List<InboundTransferTask> inboundTransfers =
transfersBySource.get(inboundTransfer.getSource());
if (inboundTransfers == null) {
inboundTransfers = new ArrayList<InboundTransferTask>();
transfersBySource.put(inboundTransfer.getSource(), inboundTransfers);
}
inboundTransfers.add(inboundTransfer);
taskQueue.add(inboundTransfer);
return inboundTransfer;
} else {
return null;
}
}
}
void onTaskCompletion(InboundTransferTask inboundTransfer) {
log.tracef("Completion of inbound transfer task: %s ", inboundTransfer);
removeTransfer(inboundTransfer);
if (activeTopologyUpdates.get() == 0) {
notifyEndOfTopologyUpdate(cacheTopology.getTopologyId());
}
}
/**
* First moves the prepared transactions originated on the leavers into the recovery cache and
* then cleans up the transactions that are not yet prepared.
*
* @param cacheTopology
*/
@Override
public void cleanupLeaverTransactions(CacheTopology cacheTopology) {
// We only care about transactions originated before this topology update
if (getMinTopologyId() >= cacheTopology.getTopologyId()) return;
Iterator<RemoteTransaction> it = getRemoteTransactions().iterator();
while (it.hasNext()) {
RecoveryAwareRemoteTransaction recTx = (RecoveryAwareRemoteTransaction) it.next();
if (recTx.getTopologyId() < cacheTopology.getTopologyId()) {
recTx.computeOrphan(cacheTopology.getMembers());
if (recTx.isInDoubt()) {
recoveryManager.registerInDoubtTransaction(recTx);
it.remove();
}
}
}
// this cleans up the transactions that are not yet prepared
super.cleanupLeaverTransactions(cacheTopology);
}
public void cleanupStaleTransactions(CacheTopology cacheTopology) {
int topologyId = cacheTopology.getTopologyId();
List<Address> members = cacheTopology.getMembers();
// We only care about transactions originated before this topology update
if (getMinTopologyId() >= topologyId) return;
log.tracef(
"Checking for transactions originated on leavers. Current members are %s, remote transactions: %d",
members, remoteTransactions.size());
Set<GlobalTransaction> toKill = new HashSet<GlobalTransaction>();
for (Map.Entry<GlobalTransaction, RemoteTransaction> e : remoteTransactions.entrySet()) {
GlobalTransaction gt = e.getKey();
RemoteTransaction remoteTx = e.getValue();
log.tracef("Checking transaction %s", gt);
// The topology id check is needed for joiners
if (remoteTx.getTopologyId() < topologyId && !members.contains(gt.getAddress())) {
toKill.add(gt);
}
}
if (toKill.isEmpty()) {
log.tracef("No global transactions pertain to originator(s) who have left the cluster.");
} else {
log.tracef("%s global transactions pertain to leavers and need to be killed", toKill.size());
}
for (GlobalTransaction gtx : toKill) {
log.tracef("Killing remote transaction originating on leaver %s", gtx);
RollbackCommand rc = new RollbackCommand(cacheName, gtx);
rc.init(invoker, icc, TransactionTable.this);
try {
rc.perform(null);
log.tracef("Rollback of transaction %s complete.", gtx);
} catch (Throwable e) {
log.unableToRollbackGlobalTx(gtx, e);
}
}
log.tracef(
"Completed cleaning transactions originating on leavers. Remote transactions remaining: %d",
remoteTransactions.size());
}
/**
* Receive notification of updated keys right before they are committed in DataContainer.
*
* @param key the key that is being modified
*/
@Override
public void addUpdatedKey(Object key) {
// grab a copy of the reference to prevent issues if another thread calls stopApplyingState()
// between null check and actual usage
final Set<Object> localUpdatedKeys = updatedKeys;
if (localUpdatedKeys != null) {
if (cacheTopology.getWriteConsistentHash().isKeyLocalToNode(rpcManager.getAddress(), key)) {
localUpdatedKeys.add(key);
}
}
}
public void applyState(Address sender, int topologyId, Collection<StateChunk> stateChunks) {
if (trace) {
log.tracef(
"Before applying the received state the data container of cache %s has %d keys",
cacheName, dataContainer.size());
}
for (StateChunk stateChunk : stateChunks) {
// it's possible to receive a late message so we must be prepared to ignore segments we no
// longer own
// todo [anistor] this check should be based on topologyId
if (!cacheTopology
.getWriteConsistentHash()
.getSegmentsForOwner(rpcManager.getAddress())
.contains(stateChunk.getSegmentId())) {
log.warnf(
"Discarding received cache entries for segment %d of cache %s because they do not belong to this node.",
stateChunk.getSegmentId(), cacheName);
continue;
}
// notify the inbound task that a chunk of cache entries was received
InboundTransferTask inboundTransfer;
synchronized (this) {
inboundTransfer = transfersBySegment.get(stateChunk.getSegmentId());
}
if (inboundTransfer != null) {
if (stateChunk.getCacheEntries() != null) {
doApplyState(sender, stateChunk.getSegmentId(), stateChunk.getCacheEntries());
}
inboundTransfer.onStateReceived(stateChunk.getSegmentId(), stateChunk.isLastChunk());
} else {
log.warnf(
"Received unsolicited state from node %s for segment %d of cache %s",
sender, stateChunk.getSegmentId(), cacheName);
}
}
if (trace) {
log.tracef(
"After applying the received state the data container of cache %s has %d keys",
cacheName, dataContainer.size());
synchronized (this) {
log.tracef("Segments not received yet for cache %s: %s", cacheName, transfersBySource);
}
}
}
private Address findSource(int segmentId, Set<Address> excludedSources) {
List<Address> owners = cacheTopology.getReadConsistentHash().locateOwnersForSegment(segmentId);
if (owners.size() == 1 && owners.get(0).equals(rpcManager.getAddress())) {
return null;
}
for (int i = owners.size() - 1;
i >= 0;
i--) { // iterate backwards because we prefer to fetch from newer nodes
Address o = owners.get(i);
if (!o.equals(rpcManager.getAddress()) && !excludedSources.contains(o)) {
return o;
}
}
log.noLiveOwnersFoundForSegment(segmentId, cacheName, owners, excludedSources);
return null;
}
private void requestTransactions(
Set<Integer> segments, Map<Address, Set<Integer>> sources, Set<Address> excludedSources) {
findSources(segments, sources, excludedSources);
boolean seenFailures = false;
while (true) {
Set<Integer> failedSegments = new HashSet<Integer>();
for (Map.Entry<Address, Set<Integer>> e : sources.entrySet()) {
Address source = e.getKey();
Set<Integer> segmentsFromSource = e.getValue();
List<TransactionInfo> transactions =
getTransactions(source, segmentsFromSource, cacheTopology.getTopologyId());
if (transactions != null) {
applyTransactions(source, transactions);
} else {
// if requesting the transactions failed we need to retry from another source
failedSegments.addAll(segmentsFromSource);
excludedSources.add(source);
}
}
if (failedSegments.isEmpty()) {
break;
}
// look for other sources for all failed segments
seenFailures = true;
sources.clear();
findSources(failedSegments, sources, excludedSources);
}
if (seenFailures) {
// start fresh when next step starts (fetching segments)
sources.clear();
}
}
private int getSegment(Object key) {
// there we can use any CH version because the routing table is not involved
return cacheTopology.getReadConsistentHash().getSegment(key);
}
@Override
public void onTopologyUpdate(CacheTopology cacheTopology, boolean isRebalance) {
if (trace)
log.tracef(
"Received new CH %s for cache %s", cacheTopology.getWriteConsistentHash(), cacheName);
int numStartedTopologyUpdates = activeTopologyUpdates.incrementAndGet();
if (isRebalance) {
rebalanceInProgress.set(true);
}
final ConsistentHash previousCh =
this.cacheTopology != null ? this.cacheTopology.getWriteConsistentHash() : null;
// Ensures writes to the data container use the right consistent hash
// No need for a try/finally block, since it's just an assignment
stateTransferLock.acquireExclusiveTopologyLock();
this.cacheTopology = cacheTopology;
if (numStartedTopologyUpdates == 1) {
updatedKeys = new ConcurrentHashSet<Object>();
}
stateTransferLock.releaseExclusiveTopologyLock();
stateTransferLock.notifyTopologyInstalled(cacheTopology.getTopologyId());
try {
// fetch transactions and data segments from other owners if this is enabled
if (isTransactional || isFetchEnabled) {
Set<Integer> addedSegments;
if (previousCh == null) {
// we start fresh, without any data, so we need to pull everything we own according to
// writeCh
addedSegments = getOwnedSegments(cacheTopology.getWriteConsistentHash());
if (trace) {
log.tracef("On cache %s we have: added segments: %s", cacheName, addedSegments);
}
} else {
Set<Integer> previousSegments = getOwnedSegments(previousCh);
Set<Integer> newSegments = getOwnedSegments(cacheTopology.getWriteConsistentHash());
Set<Integer> removedSegments = new HashSet<Integer>(previousSegments);
removedSegments.removeAll(newSegments);
// This is a rebalance, we need to request the segments we own in the new CH.
addedSegments = new HashSet<Integer>(newSegments);
addedSegments.removeAll(previousSegments);
if (trace) {
log.tracef(
"On cache %s we have: removed segments: %s; new segments: %s; old segments: %s; added segments: %s",
cacheName, removedSegments, newSegments, previousSegments, addedSegments);
}
// remove inbound transfers and any data for segments we no longer own
cancelTransfers(removedSegments);
// If L1.onRehash is enabled, "removed" segments are actually moved to L1. The new (and
// old) owners
// will automatically add the nodes that no longer own a key to that key's requestors
// list.
invalidateSegments(newSegments, removedSegments);
// check if any of the existing transfers should be restarted from a different source
// because the initial source is no longer a member
Set<Address> members =
new HashSet<Address>(cacheTopology.getReadConsistentHash().getMembers());
synchronized (this) {
for (Iterator<Address> it = transfersBySource.keySet().iterator(); it.hasNext(); ) {
Address source = it.next();
if (!members.contains(source)) {
if (trace) {
log.tracef(
"Removing inbound transfers from source %s for cache %s", source, cacheName);
}
List<InboundTransferTask> inboundTransfers = transfersBySource.get(source);
it.remove();
for (InboundTransferTask inboundTransfer : inboundTransfers) {
// these segments will be restarted if they are still in new write CH
if (trace) {
log.tracef(
"Removing inbound transfers for segments %s from source %s for cache %s",
inboundTransfer.getSegments(), source, cacheName);
}
transfersBySegment.keySet().removeAll(inboundTransfer.getSegments());
addedSegments.addAll(inboundTransfer.getUnfinishedSegments());
}
}
}
// exclude those that are already in progress from a valid source
addedSegments.removeAll(transfersBySegment.keySet());
}
}
if (!addedSegments.isEmpty()) {
addTransfers(addedSegments); // add transfers for new or restarted segments
}
}
} finally {
stateTransferLock.notifyTransactionDataReceived(cacheTopology.getTopologyId());
if (activeTopologyUpdates.decrementAndGet() == 0) {
notifyEndOfTopologyUpdate(cacheTopology.getTopologyId());
}
}
}
public void testReplace() throws Exception {
cache(0).put("myKey", "myValue");
// add an interceptor on second node that will block REPLACE commands right after
// EntryWrappingInterceptor until we are ready
final CountDownLatch replaceStartedLatch = new CountDownLatch(1);
final CountDownLatch replaceProceedLatch = new CountDownLatch(1);
boolean isVersioningEnabled = cache(0).getCacheConfiguration().versioning().enabled();
cacheConfigBuilder
.customInterceptors()
.addInterceptor()
.after(
isVersioningEnabled
? VersionedEntryWrappingInterceptor.class
: EntryWrappingInterceptor.class)
.interceptor(
new CommandInterceptor() {
@Override
protected Object handleDefault(InvocationContext ctx, VisitableCommand cmd)
throws Throwable {
if (cmd instanceof ReplaceCommand) {
// signal we encounter a REPLACE
replaceStartedLatch.countDown();
// wait until it is ok to continue with REPLACE
if (!replaceProceedLatch.await(15, TimeUnit.SECONDS)) {
throw new TimeoutException();
}
}
return super.handleDefault(ctx, cmd);
}
});
// do not allow coordinator to send topology updates to node B
final ClusterTopologyManager ctm0 =
TestingUtil.extractGlobalComponent(manager(0), ClusterTopologyManager.class);
ctm0.setRebalancingEnabled(false);
log.info("Adding a new node ..");
addClusterEnabledCacheManager(cacheConfigBuilder);
log.info("Added a new node");
// node B is not a member yet and rebalance has not started yet
CacheTopology cacheTopology =
advancedCache(1).getComponentRegistry().getStateTransferManager().getCacheTopology();
assertNull(cacheTopology.getPendingCH());
assertTrue(cacheTopology.getMembers().contains(address(0)));
assertFalse(cacheTopology.getMembers().contains(address(1)));
assertFalse(cacheTopology.getCurrentCH().getMembers().contains(address(1)));
// no keys should be present on node B yet because state transfer is blocked
assertTrue(cache(1).keySet().isEmpty());
// initiate a REPLACE
Future<Object> getFuture =
fork(
new Callable<Object>() {
@Override
public Object call() throws Exception {
try {
return cache(1).replace("myKey", "newValue");
} catch (Exception e) {
log.errorf(e, "REPLACE failed: %s", e.getMessage());
throw e;
}
}
});
// wait for REPLACE command on node B to reach beyond *EntryWrappingInterceptor, where it will
// block.
// the value seen so far is null
if (!replaceStartedLatch.await(15, TimeUnit.SECONDS)) {
throw new TimeoutException();
}
// paranoia, yes the value is still missing from data container
assertTrue(cache(1).keySet().isEmpty());
// allow rebalance to start
ctm0.setRebalancingEnabled(true);
// wait for state transfer to end
TestingUtil.waitForRehashToComplete(cache(0), cache(1));
// the state should be already transferred now
assertEquals(1, cache(1).keySet().size());
// allow REPLACE to continue
replaceProceedLatch.countDown();
Object oldVal = getFuture.get(15, TimeUnit.SECONDS);
assertNotNull(oldVal);
assertEquals("myValue", oldVal);
assertEquals("newValue", cache(0).get("myKey"));
assertEquals("newValue", cache(1).get("myKey"));
}
