org.infinispan.statetransfer.StateConsumerImpl Maven / Gradle / Ivy
package org.infinispan.statetransfer;
import static org.infinispan.context.Flag.CACHE_MODE_LOCAL;
import static org.infinispan.context.Flag.IGNORE_RETURN_VALUES;
import static org.infinispan.context.Flag.PUT_FOR_STATE_TRANSFER;
import static org.infinispan.context.Flag.SKIP_LOCKING;
import static org.infinispan.context.Flag.SKIP_OWNERSHIP_CHECK;
import static org.infinispan.context.Flag.SKIP_REMOTE_LOOKUP;
import static org.infinispan.context.Flag.SKIP_SHARED_CACHE_STORE;
import static org.infinispan.context.Flag.SKIP_XSITE_BACKUP;
import static org.infinispan.factories.KnownComponentNames.STATE_TRANSFER_EXECUTOR;
import static org.infinispan.persistence.manager.PersistenceManager.AccessMode.PRIVATE;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.HashMap;
import java.util.HashSet;
import java.util.Iterator;
import java.util.List;
import java.util.Map;
import java.util.Set;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicInteger;
import javax.transaction.SystemException;
import javax.transaction.TransactionManager;
import org.infinispan.Cache;
import org.infinispan.commands.CommandsFactory;
import org.infinispan.commands.write.InvalidateCommand;
import org.infinispan.commands.write.PutKeyValueCommand;
import org.infinispan.commons.CacheException;
import org.infinispan.commons.util.EnumUtil;
import org.infinispan.commons.util.SmallIntSet;
import org.infinispan.commons.util.concurrent.ConcurrentHashSet;
import org.infinispan.configuration.cache.CacheMode;
import org.infinispan.configuration.cache.Configuration;
import org.infinispan.container.DataContainer;
import org.infinispan.container.entries.InternalCacheEntry;
import org.infinispan.context.InvocationContext;
import org.infinispan.context.InvocationContextFactory;
import org.infinispan.context.impl.TxInvocationContext;
import org.infinispan.distexec.DistributedCallable;
import org.infinispan.distribution.DistributionInfo;
import org.infinispan.distribution.TriangleOrderManager;
import org.infinispan.distribution.DistributionManager;
import org.infinispan.distribution.ch.ConsistentHash;
import org.infinispan.distribution.ch.KeyPartitioner;
import org.infinispan.executors.LimitedExecutor;
import org.infinispan.factories.KnownComponentNames;
import org.infinispan.factories.annotations.ComponentName;
import org.infinispan.factories.annotations.Inject;
import org.infinispan.factories.annotations.Start;
import org.infinispan.factories.annotations.Stop;
import org.infinispan.filter.KeyFilter;
import org.infinispan.interceptors.AsyncInterceptorChain;
import org.infinispan.notifications.cachelistener.CacheNotifier;
import org.infinispan.persistence.manager.PersistenceManager;
import org.infinispan.remoting.responses.CacheNotFoundResponse;
import org.infinispan.remoting.responses.Response;
import org.infinispan.remoting.responses.SuccessfulResponse;
import org.infinispan.remoting.rpc.ResponseMode;
import org.infinispan.remoting.rpc.RpcManager;
import org.infinispan.remoting.rpc.RpcOptions;
import org.infinispan.remoting.transport.Address;
import org.infinispan.remoting.transport.jgroups.SuspectException;
import org.infinispan.topology.CacheTopology;
import org.infinispan.transaction.impl.RemoteTransaction;
import org.infinispan.transaction.impl.TransactionTable;
import org.infinispan.transaction.totalorder.TotalOrderLatch;
import org.infinispan.transaction.totalorder.TotalOrderManager;
import org.infinispan.transaction.xa.CacheTransaction;
import org.infinispan.transaction.xa.GlobalTransaction;
import org.infinispan.util.concurrent.BlockingTaskAwareExecutorService;
import org.infinispan.util.concurrent.CommandAckCollector;
import org.infinispan.util.concurrent.TimeoutException;
import org.infinispan.util.logging.Log;
import org.infinispan.util.logging.LogFactory;
import net.jcip.annotations.GuardedBy;
/**
* {@link StateConsumer} implementation.
*
* @author [email protected]
* @since 5.2
*/
public class StateConsumerImpl implements StateConsumer {
private static final Log log = LogFactory.getLog(StateConsumerImpl.class);
private static final boolean trace = log.isTraceEnabled();
private static final int NO_REBALANCE_IN_PROGRESS = -1;
private static final long STATE_TRANSFER_FLAGS = EnumUtil.bitSetOf(PUT_FOR_STATE_TRANSFER, CACHE_MODE_LOCAL,
IGNORE_RETURN_VALUES, SKIP_REMOTE_LOOKUP,
SKIP_SHARED_CACHE_STORE, SKIP_OWNERSHIP_CHECK,
SKIP_XSITE_BACKUP);
private Cache cache;
private StateTransferManager stateTransferManager;
private String cacheName;
private Configuration configuration;
private RpcManager rpcManager;
private TransactionManager transactionManager; // optional
private CommandsFactory commandsFactory;
private TransactionTable transactionTable; // optional
private DataContainer dataContainer;
private PersistenceManager persistenceManager;
private AsyncInterceptorChain interceptorChain;
private InvocationContextFactory icf;
private StateTransferLock stateTransferLock;
private CacheNotifier cacheNotifier;
private TotalOrderManager totalOrderManager;
private BlockingTaskAwareExecutorService remoteCommandsExecutor;
private long timeout;
private boolean isFetchEnabled;
private boolean isTransactional;
private boolean isInvalidationMode;
private boolean isTotalOrder;
private volatile KeyInvalidationListener keyInvalidationListener; //for test purpose only!
private CommitManager commitManager;
private ExecutorService stateTransferExecutor;
private CommandAckCollector commandAckCollector;
private TriangleOrderManager triangleOrderManager;
private DistributionManager distributionManager;
private KeyPartitioner keyPartitioner;
private volatile CacheTopology cacheTopology;
/**
* Indicates if there is a state transfer in progress. It is set to the new topology id when onTopologyUpdate with
* isRebalance==true is called.
* It is changed back to NO_REBALANCE_IN_PROGRESS when a topology update with a null pending CH is received.
*/
private final AtomicInteger stateTransferTopologyId = new AtomicInteger(NO_REBALANCE_IN_PROGRESS);
/**
* Indicates if there is a rebalance in progress and there the local node has not yet received
* all the new segments yet. It is set to true when rebalance starts and becomes when all inbound transfers have completed
* (before stateTransferTopologyId is set back to NO_REBALANCE_IN_PROGRESS).
*/
private final AtomicBoolean waitingForState = new AtomicBoolean(false);
private final Object transferMapsLock = new Object();
/**
* A map that keeps track of current inbound state transfers by source address. There could be multiple transfers
* flowing in from the same source (but for different segments) so the values are lists. This works in tandem with
* transfersBySegment so they always need to be kept in sync and updates to both of them need to be atomic.
*/
@GuardedBy("transferMapsLock")
private final Map> transfersBySource = new HashMap<>();
/**
* A map that keeps track of current inbound state transfers by segment id. There is at most one transfers per segment.
* This works in tandem with transfersBySource so they always need to be kept in sync and updates to both of them
* need to be atomic.
*/
@GuardedBy("transferMapsLock")
private final Map transfersBySegment = new HashMap<>();
/**
* Push RPCs on a background thread
*/
private LimitedExecutor stateRequestExecutor;
private volatile boolean ownsData = false;
private RpcOptions synchronousRpcOptions;
private RpcOptions synchronousIgnoreLeaversRpcOptions;
public StateConsumerImpl() {
}
/**
* Stops applying incoming state. Also stops tracking updated keys. Should be called at the end of state transfer or
* when a ClearCommand is committed during state transfer.
*/
@Override
public void stopApplyingState() {
if (trace) log.tracef("Stop keeping track of changed keys for state transfer");
commitManager.stopTrack(PUT_FOR_STATE_TRANSFER);
}
@Inject
public void init(Cache cache,
@ComponentName(STATE_TRANSFER_EXECUTOR) ExecutorService stateTransferExecutor,
StateTransferManager stateTransferManager,
AsyncInterceptorChain interceptorChain,
InvocationContextFactory icf,
Configuration configuration,
RpcManager rpcManager,
TransactionManager transactionManager,
CommandsFactory commandsFactory,
PersistenceManager persistenceManager,
DataContainer dataContainer,
TransactionTable transactionTable,
StateTransferLock stateTransferLock,
CacheNotifier cacheNotifier,
TotalOrderManager totalOrderManager,
@ComponentName(
KnownComponentNames.REMOTE_COMMAND_EXECUTOR) BlockingTaskAwareExecutorService remoteCommandsExecutor,
CommitManager commitManager,
CommandAckCollector commandAckCollector,
TriangleOrderManager triangleOrderManager,
DistributionManager distributionManager, KeyPartitioner keyPartitioner) {
this.cache = cache;
this.cacheName = cache.getName();
this.stateTransferExecutor = stateTransferExecutor;
this.stateTransferManager = stateTransferManager;
this.interceptorChain = interceptorChain;
this.icf = icf;
this.configuration = configuration;
this.rpcManager = rpcManager;
this.transactionManager = transactionManager;
this.commandsFactory = commandsFactory;
this.persistenceManager = persistenceManager;
this.dataContainer = dataContainer;
this.transactionTable = transactionTable;
this.stateTransferLock = stateTransferLock;
this.cacheNotifier = cacheNotifier;
this.totalOrderManager = totalOrderManager;
this.remoteCommandsExecutor = remoteCommandsExecutor;
this.commitManager = commitManager;
this.commandAckCollector = commandAckCollector;
this.triangleOrderManager = triangleOrderManager;
this.distributionManager = distributionManager;
this.keyPartitioner = keyPartitioner;
isInvalidationMode = configuration.clustering().cacheMode().isInvalidation();
isTransactional = configuration.transaction().transactionMode().isTransactional();
isTotalOrder = configuration.transaction().transactionProtocol().isTotalOrder();
timeout = configuration.clustering().stateTransfer().timeout();
}
public boolean hasActiveTransfers() {
synchronized (transferMapsLock) {
return !transfersBySource.isEmpty();
}
}
@Override
public boolean isStateTransferInProgress() {
return stateTransferTopologyId.get() != NO_REBALANCE_IN_PROGRESS;
}
@Override
public boolean isStateTransferInProgressForKey(Object key) {
if (isInvalidationMode) {
// In invalidation mode it is of not much relevance if the key is actually being transferred right now.
// A false response to this will just mean the usual remote lookup before a write operation is not
// performed and a null is assumed. But in invalidation mode the user must expect the data can disappear
// from cache at any time so this null previous value should not cause any trouble.
return false;
}
DistributionInfo distributionInfo = distributionManager.getCacheTopology().getDistribution(key);
return distributionInfo.isWriteOwner() && !distributionInfo.isReadOwner();
}
@Override
public boolean ownsData() {
return ownsData;
}
@Override
public void onTopologyUpdate(final CacheTopology cacheTopology, final boolean isRebalance) {
final boolean isMember = cacheTopology.getMembers().contains(rpcManager.getAddress());
if (trace) log.tracef("Received new topology for cache %s, isRebalance = %b, isMember = %b, topology = %s", cacheName, isRebalance, isMember, cacheTopology);
if (!ownsData && isMember) {
ownsData = true;
} else if (ownsData && !isMember) {
// This can happen after a merge, if the local node was in a minority partition.
ownsData = false;
}
// If a member leaves/crashes immediately after a rebalance was started, the new CH_UPDATE
// command may be executed before the REBALANCE_START command, so it has to start the rebalance.
boolean startRebalance = isRebalance;
if (!isRebalance) {
if (cacheTopology.getPendingCH() != null && this.cacheTopology.getPendingCH() == null) {
if (trace) log.tracef("Forcing startRebalance = true");
startRebalance = true;
}
}
if (startRebalance) {
// Only update the rebalance topology id when starting the rebalance, as we're going to ignore any state
// response with a smaller topology id
stateTransferTopologyId.compareAndSet(NO_REBALANCE_IN_PROGRESS, cacheTopology.getTopologyId());
cacheNotifier.notifyDataRehashed(cacheTopology.getCurrentCH(), cacheTopology.getPendingCH(),
cacheTopology.getUnionCH(), cacheTopology.getTopologyId(), true);
}
awaitTotalOrderTransactions(cacheTopology, startRebalance);
// Make sure we don't send a REBALANCE_CONFIRM command before we've added all the transfer tasks
// even if some of the tasks are removed and re-added
waitingForState.set(false);
final ConsistentHash newWriteCh = cacheTopology.getWriteConsistentHash();
final CacheTopology previousCacheTopology = this.cacheTopology;
final ConsistentHash previousReadCh =
previousCacheTopology != null ? previousCacheTopology.getReadConsistentHash() : null;
final ConsistentHash previousWriteCh =
previousCacheTopology != null ? previousCacheTopology.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;
triangleOrderManager.updateCacheTopology(cacheTopology);
if (distributionManager != null) {
distributionManager.setCacheTopology(cacheTopology);
}
if (startRebalance) {
if (trace) log.tracef("Start keeping track of keys for rebalance");
commitManager.stopTrack(PUT_FOR_STATE_TRANSFER);
commitManager.startTrack(PUT_FOR_STATE_TRANSFER);
}
stateTransferLock.releaseExclusiveTopologyLock();
stateTransferLock.notifyTopologyInstalled(cacheTopology.getTopologyId());
remoteCommandsExecutor.checkForReadyTasks();
try {
// fetch transactions and data segments from other owners if this is enabled
if (isTransactional || isFetchEnabled) {
Set addedSegments;
if (previousWriteCh == null) {
// If we have any segments assigned in the initial CH, it means we are the first member.
// If we are not the first member, we can only add segments via rebalance.
addedSegments = Collections.emptySet();
// TODO Perhaps we should only do this once we are a member, as listener installation should happen only on cache members?
if (configuration.clustering().cacheMode().isDistributed()) {
Collection callables = getClusterListeners(cacheTopology);
for (DistributedCallable callable : callables) {
callable.setEnvironment(cache, null);
try {
callable.call();
} catch (Exception e) {
log.clusterListenerInstallationFailure(e);
}
}
}
if (trace) {
log.tracef("On cache %s we have: added segments: %s", cacheName, addedSegments);
}
} else {
Set previousSegments = getOwnedSegments(previousWriteCh);
Set newSegments = getOwnedSegments(newWriteCh);
SmallIntSet removedSegments;
if (newSegments.size() == newWriteCh.getNumSegments()) {
// Optimization for replicated caches
removedSegments = new SmallIntSet();
} else {
removedSegments = new SmallIntSet(previousSegments);
removedSegments.removeAll(newSegments);
}
// This is a rebalance, we need to request the segments we own in the new CH.
addedSegments = new SmallIntSet(newSegments);
addedSegments.removeAll(previousSegments);
if (trace) {
log.tracef("On cache %s we have: new segments: %s; old segments: %s", cacheName, newSegments, previousSegments);
log.tracef("On cache %s we have: added segments: %s; removed segments: %s", cacheName, addedSegments, removedSegments);
}
// remove inbound transfers for segments we no longer own
cancelTransfers(removedSegments);
if (!startRebalance && !addedSegments.isEmpty()) {
// If the last owner of a segment leaves the cluster, a new set of owners is assigned,
// but the new owners should not try to retrieve the segment from each other.
// If this happens during a rebalance, we might have already sent our rebalance
// confirmation, so the coordinator won't wait for us to retrieve those segments anyway.
log.debugf("Not requesting segments %s because the last owner left the cluster",
addedSegments);
addedSegments.clear();
}
// check if any of the existing transfers should be restarted from a different source because
// the initial source is no longer a member
restartBrokenTransfers(cacheTopology, addedSegments);
}
if (!addedSegments.isEmpty()) {
// add transfers for new or restarted segments
addTransfers(addedSegments);
}
}
int rebalanceTopologyId = stateTransferTopologyId.get();
if (trace) log.tracef("Topology update processed, stateTransferTopologyId = %d, startRebalance = %s, pending CH = %s",
(Object)rebalanceTopologyId, startRebalance, cacheTopology.getPendingCH());
if (rebalanceTopologyId != NO_REBALANCE_IN_PROGRESS) {
// there was a rebalance in progress
if (!startRebalance && cacheTopology.getPendingCH() == null) {
// we have received a topology update without a pending CH, signalling the end of the rebalance
boolean changed = stateTransferTopologyId.compareAndSet(rebalanceTopologyId, NO_REBALANCE_IN_PROGRESS);
if (changed) {
stopApplyingState();
// if the coordinator changed, we might get two concurrent topology updates,
// but we only want to notify the @DataRehashed listeners once
cacheNotifier.notifyDataRehashed(previousReadCh, cacheTopology.getCurrentCH(), previousWriteCh,
cacheTopology.getTopologyId(), false);
if (trace) {
log.tracef("Unlock State Transfer in Progress for topology ID %s", cacheTopology.getTopologyId());
}
if (isTotalOrder) {
totalOrderManager.notifyStateTransferEnd();
}
}
}
}
} finally {
stateTransferLock.notifyTransactionDataReceived(cacheTopology.getTopologyId());
remoteCommandsExecutor.checkForReadyTasks();
// Only set the flag here, after all the transfers have been added to the transfersBySource map
if (stateTransferTopologyId.get() != NO_REBALANCE_IN_PROGRESS && isMember) {
waitingForState.set(true);
}
notifyEndOfRebalanceIfNeeded(cacheTopology.getTopologyId(), cacheTopology.getRebalanceId());
// Remove the transactions whose originators have left the cache.
// Need to do it now, after we have applied any transactions from other nodes,
// and after notifyTransactionDataReceived - otherwise the RollbackCommands would block.
if (transactionTable != null) {
transactionTable.cleanupLeaverTransactions(rpcManager.getTransport().getMembers());
}
commandAckCollector.onMembersChange(newWriteCh.getMembers());
// Any data for segments we do not own should be removed from data container and cache store
// We need to discard data from all segments we don't own, not just those we previously owned,
// when we lose membership (e.g. because there was a merge, the local partition was in degraded mode
// and the other partition was available) or when L1 is enabled.
Set removedSegments;
boolean wasMember =
previousWriteCh != null && previousWriteCh.getMembers().contains(rpcManager.getAddress());
if (isMember || wasMember) {
removedSegments = new HashSet<>(newWriteCh.getNumSegments());
for (int i = 0; i < newWriteCh.getNumSegments(); i++) {
removedSegments.add(i);
}
Set newSegments = getOwnedSegments(newWriteCh);
removedSegments.removeAll(newSegments);
try {
removeStaleData(removedSegments);
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
throw new CacheException(e);
}
}
}
}
private void awaitTotalOrderTransactions(CacheTopology cacheTopology, boolean isRebalance) {
//in total order, we should wait for remote transactions before proceeding
if (isTotalOrder) {
if (trace) {
log.trace("State Transfer in Total Order cache. Waiting for remote transactions to finish");
}
try {
for (TotalOrderLatch block : totalOrderManager.notifyStateTransferStart(cacheTopology.getTopologyId(), isRebalance)) {
block.awaitUntilUnBlock();
}
} catch (InterruptedException e) {
//interrupted...
Thread.currentThread().interrupt();
throw new CacheException(e);
}
if (trace) {
log.trace(
"State Transfer in Total Order cache. All remote transactions are finished. Moving on...");
}
}
}
private void notifyEndOfRebalanceIfNeeded(int topologyId, int rebalanceId) {
if (waitingForState.get() && !hasActiveTransfers()) {
if (waitingForState.compareAndSet(true, false)) {
log.debugf("Finished receiving of segments for cache %s for topology %d.", cacheName, topologyId);
stopApplyingState();
stateTransferManager.notifyEndOfRebalance(topologyId, rebalanceId);
}
}
}
private Set getOwnedSegments(ConsistentHash consistentHash) {
Address address = rpcManager.getAddress();
return consistentHash.getMembers().contains(address) ? consistentHash.getSegmentsForOwner(address)
: Collections.emptySet();
}
@Override
public void applyState(final Address sender, int topologyId, Collection stateChunks) {
ConsistentHash wCh = cacheTopology.getWriteConsistentHash();
// Ignore responses received after we are no longer a member
if (!wCh.getMembers().contains(rpcManager.getAddress())) {
if (trace) {
log.tracef("Ignoring received state because we are no longer a member of cache %s", cacheName);
}
return;
}
// Ignore segments that we requested for a previous rebalance
// Can happen when the coordinator leaves, and the new coordinator cancels the rebalance in progress
int rebalanceTopologyId = stateTransferTopologyId.get();
if (rebalanceTopologyId == NO_REBALANCE_IN_PROGRESS) {
log.debugf("Discarding state response with topology id %d for cache %s, we don't have a state transfer in progress",
topologyId, cacheName);
return;
}
if (topologyId < rebalanceTopologyId) {
log.debugf("Discarding state response with old topology id %d for cache %s, state transfer request topology was %b",
topologyId, cacheName, waitingForState);
return;
}
if (trace) {
log.tracef("Before applying the received state the data container of cache %s has %d keys", cacheName,
dataContainer.sizeIncludingExpired());
}
final Set mySegments = wCh.getSegmentsForOwner(rpcManager.getAddress());
final CountDownLatch countDownLatch = new CountDownLatch(stateChunks.size());
for (final StateChunk stateChunk : stateChunks) {
stateTransferExecutor.submit(() -> {
applyChunk(sender, mySegments, stateChunk);
countDownLatch.countDown();
});
}
try {
boolean await = countDownLatch.await(timeout, TimeUnit.MILLISECONDS);
if (!await) {
throw new TimeoutException("Timed out applying state");
}
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
throw new CacheException(e);
}
if (trace) {
log.tracef("After applying the received state the data container of cache %s has %d keys", cacheName,
dataContainer.sizeIncludingExpired());
synchronized (transferMapsLock) {
log.tracef("Segments not received yet for cache %s: %s", cacheName, transfersBySource);
}
}
}
private void applyChunk(Address sender, Set mySegments, StateChunk stateChunk) {
if (!mySegments.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);
return;
}
// Notify the inbound task that a chunk of cache entries was received
InboundTransferTask inboundTransfer;
synchronized (transferMapsLock) {
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 {
if (cache.getStatus().allowInvocations()) {
log.ignoringUnsolicitedState(sender, stateChunk.getSegmentId(), cacheName);
}
}
}
private void doApplyState(Address sender, int segmentId, Collection cacheEntries) {
if (trace) log.tracef("Applying new state chunk for segment %d of cache %s from node %s: received %d cache entries",
segmentId, cacheName, sender, cacheEntries.size());
// CACHE_MODE_LOCAL avoids handling by StateTransferInterceptor and any potential locks in StateTransferLock
boolean transactional = transactionManager != null;
for (InternalCacheEntry e : cacheEntries) {
try {
InvocationContext ctx;
if (transactional) {
transactionManager.begin();
ctx = icf.createInvocationContext(transactionManager.getTransaction(), true);
((TxInvocationContext) ctx).getCacheTransaction().setStateTransferFlag(PUT_FOR_STATE_TRANSFER);
} else {
// non-tx cache
ctx = icf.createSingleKeyNonTxInvocationContext();
}
PutKeyValueCommand put = commandsFactory.buildPutKeyValueCommand(
e.getKey(), e.getValue(), e.getMetadata(), STATE_TRANSFER_FLAGS);
ctx.setLockOwner(put.getKeyLockOwner());
interceptorChain.invoke(ctx, put);
if (transactionManager != null) {
transactionManager.commit();
}
} catch (Exception ex) {
if (!cache.getStatus().allowInvocations()) {
log.debugf("Cache %s is shutting down, stopping state transfer", cacheName);
break;
} else {
log.problemApplyingStateForKey(ex.getMessage(), e.getKey(), ex);
}
} finally {
try {
if (transactional && transactionManager.getTransaction() != null) {
transactionManager.rollback();
}
} catch (SystemException e1) {
// Ignore
}
}
}
if (trace) log.tracef("Finished applying chunk of segment %d of cache %s", segmentId, cacheName);
}
private void applyTransactions(Address sender, Collection transactions, int topologyId) {
log.debugf("Applying %d transactions for cache %s transferred from node %s", transactions.size(), cacheName, sender);
if (isTransactional) {
for (TransactionInfo transactionInfo : transactions) {
GlobalTransaction gtx = transactionInfo.getGlobalTransaction();
if (rpcManager.getAddress().equals(gtx.getAddress())) {
continue; // it is a transaction originated in this node. can happen with partition handling
}
// Mark the global transaction as remote. Only used for logging, hashCode/equals ignore it.
gtx.setRemote(true);
CacheTransaction tx = transactionTable.getLocalTransaction(gtx);
if (tx == null) {
tx = transactionTable.getRemoteTransaction(gtx);
if (tx == null) {
tx = transactionTable.getOrCreateRemoteTransaction(gtx, transactionInfo.getModifications());
// Force this node to replay the given transaction data by making it think it is 1 behind
((RemoteTransaction) tx).setLookedUpEntriesTopology(topologyId - 1);
}
}
// TODO Shouldn't this be done for transactions originated locally as well?
transactionInfo.getLockedKeys().forEach(tx::addBackupLockForKey);
}
}
}
// Must run after the PersistenceManager
@Start(priority = 20)
public void start() {
CacheMode mode = configuration.clustering().cacheMode();
isFetchEnabled = (mode.isDistributed() || mode.isReplicated()) &&
(configuration.clustering().stateTransfer().fetchInMemoryState() || configuration.persistence().fetchPersistentState());
//rpc options does not changes in runtime. we can use always the same instance.
synchronousRpcOptions = rpcManager.getRpcOptionsBuilder(ResponseMode.SYNCHRONOUS)
.timeout(timeout, TimeUnit.MILLISECONDS).build();
synchronousIgnoreLeaversRpcOptions = rpcManager.getRpcOptionsBuilder(ResponseMode.SYNCHRONOUS_IGNORE_LEAVERS)
.timeout(timeout, TimeUnit.MILLISECONDS).build();
stateRequestExecutor = new LimitedExecutor("StateRequest-" + cacheName, stateTransferExecutor, 1);
}
@Stop(priority = 0)
@Override
public void stop() {
if (trace) {
log.tracef("Shutting down StateConsumer of cache %s on node %s", cacheName, rpcManager.getAddress());
}
try {
synchronized (transferMapsLock) {
// cancel all inbound transfers
stateRequestExecutor.cancelQueuedTasks();
for (List inboundTransfers : transfersBySource.values()) {
inboundTransfers.forEach(InboundTransferTask::cancel);
}
transfersBySource.clear();
transfersBySegment.clear();
}
} catch (Throwable t) {
log.errorf(t, "Failed to stop StateConsumer of cache %s on node %s", cacheName, rpcManager.getAddress());
}
}
@Override
public CacheTopology getCacheTopology() {
return cacheTopology;
}
public void setKeyInvalidationListener(KeyInvalidationListener keyInvalidationListener) {
this.keyInvalidationListener = keyInvalidationListener;
}
private void addTransfers(Set segments) {
log.debugf("Adding inbound state transfer for segments %s", segments);
// the set of nodes that reported errors when fetching data from them - these will not be retried in this topology
Set excludedSources = new HashSet<>();
// the sources and segments we are going to get from each source
Map> sources = new HashMap<>();
if (isTransactional && !isTotalOrder) {
requestTransactions(segments, sources, excludedSources);
}
if (isFetchEnabled) {
requestSegments(segments, sources, excludedSources);
}
if (trace) log.tracef("Finished adding inbound state transfer for segments %s", segments,
cacheName);
}
private void findSources(Set segments, Map> sources, Set excludedSources) {
if (cache.getStatus().isTerminated())
return;
SmallIntSet segmentsWithoutSource = new SmallIntSet(configuration.clustering().hash().numSegments());
for (Integer segmentId : segments) {
Address source = findSource(segmentId, excludedSources);
// ignore all segments for which there are no other owners to pull data from.
// these segments are considered empty (or lost) and do not require a state transfer
if (source != null) {
Set segmentsFromSource = sources.computeIfAbsent(source, k -> new SmallIntSet());
segmentsFromSource.add(segmentId);
} else {
segmentsWithoutSource.set(segmentId);
}
}
if (!segmentsWithoutSource.isEmpty()) {
log.noLiveOwnersFoundForSegments(segmentsWithoutSource, cacheName, excludedSources);
}
}
private Address findSource(int segmentId, Set excludedSources) {
List owners = cacheTopology.getReadConsistentHash().locateOwnersForSegment(segmentId);
if (!owners.contains(rpcManager.getAddress())) {
// We prefer that transactions are sourced from primary owners.
// Needed in pessimistic mode, if the originator is the primary owner of the key than the lock
// command is not replicated to the backup owners. See PessimisticDistributionInterceptor.acquireRemoteIfNeeded.
for (Address o : owners) {
if (!o.equals(rpcManager.getAddress()) && !excludedSources.contains(o)) {
return o;
}
}
}
return null;
}
private void requestTransactions(Set segments, Map> sources, Set excludedSources) {
findSources(segments, sources, excludedSources);
boolean seenFailures = false;
while (true) {
SmallIntSet failedSegments = new SmallIntSet();
int topologyId = cacheTopology.getTopologyId();
for (Map.Entry> sourceEntry : sources.entrySet()) {
Address source = sourceEntry.getKey();
Set segmentsFromSource = sourceEntry.getValue();
boolean failed = false;
boolean exclude = false;
try {
Response response = getTransactions(source, segmentsFromSource, topologyId);
if (response instanceof SuccessfulResponse) {
List transactions = (List) ((SuccessfulResponse) response).getResponseValue();
applyTransactions(source, transactions, topologyId);
} else if (response instanceof CacheNotFoundResponse) {
log.debugf("Cache %s was stopped on node %s before sending transaction information", cacheName, source);
failed = true;
exclude = true;
} else {
log.unsuccessfulResponseRetrievingTransactionsForSegments(source, response);
failed = true;
}
} catch (SuspectException e) {
log.debugf("Node %s left the cluster before sending transaction information", source);
failed = true;
exclude = true;
} catch (Exception e) {
if (cache.getStatus().isTerminated()) {
log.debugf("Cache %s has stopped while requesting transactions", cacheName);
sources.clear();
return;
} else {
log.failedToRetrieveTransactionsForSegments(cacheName, source, segments, e);
}
// The primary owner is still in the cluster, so we can't exclude it - see ISPN-4091
failed = true;
}
// If requesting the transactions failed we need to retry
if (failed) {
failedSegments.addAll(segmentsFromSource);
}
// If the primary owner is no longer running, we can retry on a backup owner
if (exclude) {
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 Collection getClusterListeners(CacheTopology topology) {
for (Address source : topology.getMembers()) {
// Don't send to ourselves
if (!source.equals(rpcManager.getAddress())) {
if (trace) {
log.tracef("Requesting cluster listeners of cache %s from node %s", cacheName, source);
}
// get cluster listeners
try {
StateRequestCommand cmd = commandsFactory.buildStateRequestCommand(StateRequestCommand.Type.GET_CACHE_LISTENERS,
rpcManager.getAddress(), topology.getTopologyId(), null);
Map responses = rpcManager.invokeRemotely(Collections.singleton(source), cmd, synchronousIgnoreLeaversRpcOptions);
Response response = responses.get(source);
if (response instanceof SuccessfulResponse) {
return (Collection) ((SuccessfulResponse) response).getResponseValue();
} else {
log.unsuccessfulResponseForClusterListeners(source, response);
}
} catch (CacheException e) {
log.exceptionDuringClusterListenerRetrieval(source, e);
}
}
}
if (trace) log.trace("Unable to acquire cluster listeners from other members, assuming none are present");
return Collections.emptySet();
}
private Response getTransactions(Address source, Set segments, int topologyId) {
if (trace) {
log.tracef("Requesting transactions from node %s for segments %s", source, segments);
}
// get transactions and locks
StateRequestCommand cmd = commandsFactory.buildStateRequestCommand(StateRequestCommand.Type.GET_TRANSACTIONS, rpcManager.getAddress(), topologyId, segments);
Map responses = rpcManager.invokeRemotely(Collections.singleton(source), cmd, synchronousRpcOptions);
return responses.get(source);
}
private void requestSegments(Set segments, Map> sources, Set excludedSources) {
if (sources.isEmpty()) {
findSources(segments, sources, excludedSources);
}
for (Map.Entry> e : sources.entrySet()) {
addTransfer(e.getKey(), e.getValue());
}
}
/**
* Cancel transfers for segments we no longer own.
*
* @param removedSegments segments to be cancelled
*/
private void cancelTransfers(Set removedSegments) {
synchronized (transferMapsLock) {
List segmentsToCancel = new ArrayList<>(removedSegments);
while (!segmentsToCancel.isEmpty()) {
int segmentId = segmentsToCancel.remove(0);
InboundTransferTask inboundTransfer = transfersBySegment.get(segmentId);
if (inboundTransfer != null) { // we need to check the transfer was not already completed
Set cancelledSegments = new SmallIntSet(removedSegments);
cancelledSegments.retainAll(inboundTransfer.getSegments());
segmentsToCancel.removeAll(cancelledSegments);
transfersBySegment.keySet().removeAll(cancelledSegments);
//this will also remove it from transfersBySource if the entire task gets cancelled
inboundTransfer.cancelSegments(cancelledSegments);
if (inboundTransfer.isCancelled()) {
removeTransfer(inboundTransfer);
}
}
}
}
}
private void removeStaleData(final Set removedSegments) throws InterruptedException {
log.debugf("Removing no longer owned entries for cache %s", cacheName);
if (keyInvalidationListener != null) {
keyInvalidationListener.beforeInvalidation(removedSegments, Collections.emptySet());
}
if (removedSegments.isEmpty())
return;
// Keys that we used to own, and need to be removed from the data container AND the cache stores
final ConcurrentHashSet © 2015 - 2025 Weber Informatics LLC | Privacy Policy