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.IRAC_STATE;
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.NON_BLOCKING_EXECUTOR;
import static org.infinispan.persistence.manager.PersistenceManager.AccessMode.PRIVATE;
import static org.infinispan.util.concurrent.CompletionStages.handleAndCompose;
import static org.infinispan.util.concurrent.CompletionStages.ignoreValue;
import static org.infinispan.util.logging.Log.PERSISTENCE;
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.PrimitiveIterator;
import java.util.Set;
import java.util.concurrent.CompletableFuture;
import java.util.concurrent.CompletionStage;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.Executor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLong;
import java.util.function.Predicate;
import javax.transaction.Transaction;
import javax.transaction.TransactionManager;
import org.infinispan.Cache;
import org.infinispan.commands.CommandsFactory;
import org.infinispan.commands.remote.CacheRpcCommand;
import org.infinispan.commands.tx.PrepareCommand;
import org.infinispan.commands.tx.RollbackCommand;
import org.infinispan.commands.write.InvalidateCommand;
import org.infinispan.commands.write.PutKeyValueCommand;
import org.infinispan.commons.IllegalLifecycleStateException;
import org.infinispan.commons.tx.TransactionImpl;
import org.infinispan.commons.tx.XidImpl;
import org.infinispan.commons.util.EnumUtil;
import org.infinispan.commons.util.IntSet;
import org.infinispan.commons.util.IntSets;
import org.infinispan.commons.util.Util;
import org.infinispan.configuration.cache.CacheMode;
import org.infinispan.configuration.cache.Configuration;
import org.infinispan.configuration.cache.Configurations;
import org.infinispan.conflict.impl.InternalConflictManager;
import org.infinispan.container.entries.InternalCacheEntry;
import org.infinispan.container.impl.InternalDataContainer;
import org.infinispan.context.InvocationContext;
import org.infinispan.context.InvocationContextFactory;
import org.infinispan.context.impl.LocalTxInvocationContext;
import org.infinispan.distribution.DistributionInfo;
import org.infinispan.distribution.DistributionManager;
import org.infinispan.distribution.TriangleOrderManager;
import org.infinispan.distribution.ch.ConsistentHash;
import org.infinispan.distribution.ch.KeyPartitioner;
import org.infinispan.executors.LimitedExecutor;
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.factories.impl.ComponentRef;
import org.infinispan.factories.scopes.Scope;
import org.infinispan.factories.scopes.Scopes;
import org.infinispan.interceptors.AsyncInterceptorChain;
import org.infinispan.metadata.impl.InternalMetadataImpl;
import org.infinispan.notifications.cachelistener.CacheNotifier;
import org.infinispan.notifications.cachelistener.annotation.DataRehashed;
import org.infinispan.notifications.cachelistener.cluster.ClusterListenerReplicateCallable;
import org.infinispan.persistence.manager.PersistenceManager;
import org.infinispan.reactive.publisher.impl.LocalPublisherManager;
import org.infinispan.remoting.inboundhandler.DeliverOrder;
import org.infinispan.remoting.inboundhandler.PerCacheInboundInvocationHandler;
import org.infinispan.remoting.responses.CacheNotFoundResponse;
import org.infinispan.remoting.responses.Response;
import org.infinispan.remoting.responses.SuccessfulResponse;
import org.infinispan.remoting.responses.ValidResponse;
import org.infinispan.remoting.rpc.RpcManager;
import org.infinispan.remoting.rpc.RpcOptions;
import org.infinispan.remoting.transport.Address;
import org.infinispan.remoting.transport.impl.PassthroughSingleResponseCollector;
import org.infinispan.remoting.transport.impl.SingleResponseCollector;
import org.infinispan.topology.CacheTopology;
import org.infinispan.topology.LocalTopologyManager;
import org.infinispan.transaction.impl.LocalTransaction;
import org.infinispan.transaction.impl.RemoteTransaction;
import org.infinispan.transaction.impl.TransactionTable;
import org.infinispan.transaction.xa.CacheTransaction;
import org.infinispan.transaction.xa.GlobalTransaction;
import org.infinispan.util.concurrent.AggregateCompletionStage;
import org.infinispan.util.concurrent.CommandAckCollector;
import org.infinispan.util.concurrent.CompletableFutures;
import org.infinispan.util.concurrent.CompletionStages;
import org.infinispan.util.logging.Log;
import org.infinispan.util.logging.LogFactory;
import org.reactivestreams.Publisher;
import io.reactivex.rxjava3.core.Completable;
import io.reactivex.rxjava3.core.Flowable;
import net.jcip.annotations.GuardedBy;
/**
* {@link StateConsumer} implementation.
*
* @author [email protected]
* @since 5.2
*/
@Scope(Scopes.NAMED_CACHE)
public class StateConsumerImpl implements StateConsumer {
private static final Log log = LogFactory.getLog(StateConsumerImpl.class);
private static final boolean trace = log.isTraceEnabled();
protected static final int NO_STATE_TRANSFER_IN_PROGRESS = -1;
protected 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, SKIP_LOCKING, IRAC_STATE);
public static final String NO_KEY = "N/A";
@Inject protected ComponentRef> cache;
@Inject protected LocalTopologyManager localTopologyManager;
@Inject protected Configuration configuration;
@Inject protected RpcManager rpcManager;
@Inject protected TransactionManager transactionManager; // optional
@Inject protected CommandsFactory commandsFactory;
@Inject protected TransactionTable transactionTable; // optional
@Inject protected InternalDataContainer dataContainer;
@Inject protected PersistenceManager persistenceManager;
@Inject protected AsyncInterceptorChain interceptorChain;
@Inject protected InvocationContextFactory icf;
@Inject protected StateTransferLock stateTransferLock;
@Inject protected CacheNotifier cacheNotifier;
@Inject protected CommitManager commitManager;
@Inject @ComponentName(NON_BLOCKING_EXECUTOR)
protected Executor nonBlockingExecutor;
@Inject protected CommandAckCollector commandAckCollector;
@Inject protected TriangleOrderManager triangleOrderManager;
@Inject protected DistributionManager distributionManager;
@Inject protected KeyPartitioner keyPartitioner;
@Inject InternalConflictManager conflictManager;
@Inject LocalPublisherManager localPublisherManager;
@Inject PerCacheInboundInvocationHandler inboundInvocationHandler;
protected String cacheName;
protected long timeout;
protected boolean isFetchEnabled;
protected boolean isTransactional;
protected boolean isInvalidationMode;
protected volatile KeyInvalidationListener keyInvalidationListener; //for test purpose only!
protected volatile CacheTopology cacheTopology;
// The first topology in which the local node was a member. Any command with a lower
// topology id will be ignored.
private final int firstTopologyAsMember = Integer.MAX_VALUE;
/**
* 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.
*/
protected final AtomicInteger stateTransferTopologyId = new AtomicInteger(NO_STATE_TRANSFER_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).
*/
protected final AtomicBoolean waitingForState = new AtomicBoolean(false);
protected CompletableFuture stateTransferFuture = CompletableFutures.completedNull();
protected 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")
protected final Map> transfersBySegment = new HashMap<>();
/**
* Limit to one state request at a time.
*/
protected LimitedExecutor stateRequestExecutor;
private volatile boolean ownsData = false;
// Use the state transfer timeout for RPCs instead of the regular remote timeout
protected RpcOptions rpcOptions;
private volatile boolean running;
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(int topologyId) {
if (trace) log.tracef("Stop keeping track of changed keys for state transfer in topology %d", topologyId);
commitManager.stopTrack(PUT_FOR_STATE_TRANSFER);
}
public boolean hasActiveTransfers() {
synchronized (transferMapsLock) {
return !transfersBySource.isEmpty();
}
}
@Override
public boolean isStateTransferInProgress() {
return stateTransferTopologyId.get() != NO_STATE_TRANSFER_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 CompletionStage> onTopologyUpdate(final CacheTopology cacheTopology,
final boolean isRebalance) {
final ConsistentHash newWriteCh = cacheTopology.getWriteConsistentHash();
final CacheTopology previousCacheTopology = this.cacheTopology;
final ConsistentHash previousReadCh =
previousCacheTopology != null ? previousCacheTopology.getCurrentCH() : null;
final ConsistentHash previousWriteCh =
previousCacheTopology != null ? previousCacheTopology.getWriteConsistentHash() : null;
IntSet newWriteSegments = getOwnedSegments(newWriteCh);
final boolean isMember = cacheTopology.getMembers().contains(rpcManager.getAddress());
final boolean wasMember = previousWriteCh != null &&
previousWriteCh.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 addedPendingCH = cacheTopology.getPendingCH() != null && wasMember &&
previousCacheTopology.getPendingCH() == null;
boolean startConflictResolution =
!isRebalance && cacheTopology.getPhase() == CacheTopology.Phase.CONFLICT_RESOLUTION;
boolean startRebalance = isRebalance || (addedPendingCH && !startConflictResolution);
if (startRebalance && !isRebalance) {
if (trace) log.tracef("Forcing startRebalance = true");
}
CompletionStage stage = CompletableFutures.completedNull();
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_STATE_TRANSFER_IN_PROGRESS, cacheTopology.getTopologyId());
conflictManager.cancelVersionRequests();
if (cacheNotifier.hasListener(DataRehashed.class)) {
stage = cacheNotifier.notifyDataRehashed(cacheTopology.getCurrentCH(), cacheTopology.getPendingCH(),
cacheTopology.getUnionCH(), cacheTopology.getTopologyId(), true);
}
}
stage = stage.thenCompose(ignored -> {
if (startConflictResolution) {
// This stops state being applied from a prior rebalance and also prevents tracking from being stopped
stateTransferTopologyId.set(NO_STATE_TRANSFER_IN_PROGRESS);
}
// 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);
stateTransferFuture = new CompletableFuture<>();
beforeTopologyInstalled(cacheTopology.getTopologyId(), startRebalance, previousWriteCh, newWriteCh);
if (!configuration.clustering().cacheMode().isInvalidation()) {
// Owned segments
dataContainer.addSegments(newWriteSegments);
// TODO Should we throw an exception if addSegments() returns false?
return ignoreValue(persistenceManager.addSegments(newWriteSegments));
}
return CompletableFutures.completedNull();
});
stage = stage.thenCompose(ignored -> {
// We need to track changes so that user puts during conflict resolution are prioritised over
// state transfer or conflict resolution updates
// Tracking is stopped once the subsequent rebalance completes
if (startRebalance || startConflictResolution) {
if (trace) log.tracef("Start keeping track of keys for rebalance");
commitManager.stopTrack(PUT_FOR_STATE_TRANSFER);
commitManager.startTrack(PUT_FOR_STATE_TRANSFER);
}
// Ensures writes to the data container use the right consistent hash
// Writers block on the state transfer shared lock, so we keep the exclusive lock as short as possible
stateTransferLock.acquireExclusiveTopologyLock();
try {
this.cacheTopology = cacheTopology;
distributionManager.setCacheTopology(cacheTopology);
} finally {
stateTransferLock.releaseExclusiveTopologyLock();
}
stateTransferLock.notifyTopologyInstalled(cacheTopology.getTopologyId());
inboundInvocationHandler.checkForReadyTasks();
if (!wasMember && isMember) {
return fetchClusterListeners(cacheTopology);
}
return CompletableFutures.completedNull();
});
stage = stage.thenCompose(ignored -> {
// fetch transactions and data segments from other owners if this is enabled
if (startConflictResolution || (!isTransactional && !isFetchEnabled)) {
return CompletableFutures.completedNull();
}
IntSet addedSegments, removedSegments;
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.
removedSegments = IntSets.immutableEmptySet();
addedSegments = IntSets.immutableEmptySet();
if (trace) {
log.tracef("On cache %s we have: added segments: %s", cacheName, addedSegments);
}
} else {
IntSet previousSegments = getOwnedSegments(previousWriteCh);
if (newWriteSegments.size() == newWriteCh.getNumSegments()) {
// Optimization for replicated caches
removedSegments = IntSets.immutableEmptySet();
} else {
removedSegments = IntSets.mutableCopyFrom(previousSegments);
removedSegments.removeAll(newWriteSegments);
}
// This is a rebalance, we need to request the segments we own in the new CH.
addedSegments = IntSets.mutableCopyFrom(newWriteSegments);
addedSegments.removeAll(previousSegments);
if (trace) {
log.tracef("On cache %s we have: new segments: %s; old segments: %s", cacheName, newWriteSegments,
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);
// Scattered cache gets added segments on the first CH_UPDATE, and we want to keep these
if (!startRebalance && !addedSegments.isEmpty() &&
!configuration.clustering().cacheMode().isScattered()) {
// 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);
}
return handleSegments(startRebalance, addedSegments, removedSegments);
});
stage = stage.thenCompose(ignored -> {
int stateTransferTopologyId = this.stateTransferTopologyId.get();
if (trace)
log.tracef("Topology update processed, stateTransferTopologyId = %d, startRebalance = %s, pending CH = %s",
(Object) stateTransferTopologyId, startRebalance, cacheTopology.getPendingCH());
if (stateTransferTopologyId != NO_STATE_TRANSFER_IN_PROGRESS && !startRebalance &&
!cacheTopology.getPhase().isRebalance()) {
// we have received a topology update without a pending CH, signalling the end of the rebalance
boolean changed = this.stateTransferTopologyId.compareAndSet(stateTransferTopologyId,
NO_STATE_TRANSFER_IN_PROGRESS);
if (changed) {
stopApplyingState(stateTransferTopologyId);
// if the coordinator changed, we might get two concurrent topology updates,
// but we only want to notify the @DataRehashed listeners once
ConsistentHash nextConsistentHash = cacheTopology.getPendingCH();
if (nextConsistentHash == null) {
nextConsistentHash = cacheTopology.getCurrentCH();
}
if (cacheNotifier.hasListener(DataRehashed.class)) {
return cacheNotifier.notifyDataRehashed(previousReadCh, nextConsistentHash, previousWriteCh,
cacheTopology.getTopologyId(), false);
}
}
}
return CompletableFutures.completedNull();
});
return handleAndCompose(stage, (ignored, throwable) -> {
if (trace) {
log.tracef("Unlock State Transfer in Progress for topology ID %s", cacheTopology.getTopologyId());
}
stateTransferLock.notifyTransactionDataReceived(cacheTopology.getTopologyId());
inboundInvocationHandler.checkForReadyTasks();
// Only set the flag here, after all the transfers have been added to the transfersBySource map
if (stateTransferTopologyId.get() != NO_STATE_TRANSFER_IN_PROGRESS && isMember) {
waitingForState.set(true);
}
notifyEndOfStateTransferIfNeeded();
// 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.
try {
if (transactionTable != null) {
transactionTable.cleanupLeaverTransactions(rpcManager.getTransport().getMembers());
}
} catch (Exception e) {
// Do not fail state transfer when the cleanup fails. See ISPN-7437 for details.
log.transactionCleanupError(e);
}
commandAckCollector.onMembersChange(newWriteCh.getMembers());
// The rebalance (READ_OLD_WRITE_ALL/TRANSITORY) is completed through notifyEndOfStateTransferIfNeeded
// and STABLE does not have to be confirmed at all
switch (cacheTopology.getPhase()) {
case READ_ALL_WRITE_ALL:
case READ_NEW_WRITE_ALL:
stateTransferFuture.complete(null);
}
// 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.
if ((isMember || wasMember) && cacheTopology.getPhase() == CacheTopology.Phase.NO_REBALANCE) {
int numSegments = newWriteCh.getNumSegments();
IntSet removedSegments = IntSets.mutableEmptySet(numSegments);
IntSet newSegments = getOwnedSegments(newWriteCh);
for (int i = 0; i < numSegments; ++i) {
if (!newSegments.contains(i)) {
removedSegments.add(i);
}
}
return removeStaleData(removedSegments)
.thenApply(ignored1 -> {
conflictManager.restartVersionRequests();
// rethrow the original exception, if any
CompletableFutures.rethrowExceptionIfPresent(throwable);
return stateTransferFuture;
});
}
CompletableFutures.rethrowExceptionIfPresent(throwable);
return CompletableFuture.completedFuture(stateTransferFuture);
});
}
private CompletionStage fetchClusterListeners(CacheTopology cacheTopology) {
if (!configuration.clustering().cacheMode().isDistributed() &&
!configuration.clustering().cacheMode().isScattered()) {
return CompletableFutures.completedNull();
}
return getClusterListeners(cacheTopology.getTopologyId(), cacheTopology.getMembers()).thenAccept(callables -> {
Cache cache = this.cache.wired();
for (ClusterListenerReplicateCallable callable : callables) {
try {
// TODO: need security check?
// We have to invoke a separate method as we can't retrieve the cache as it is still starting
callable.accept(cache.getCacheManager(), cache);
} catch (Exception e) {
log.clusterListenerInstallationFailure(e);
}
}
});
}
protected void beforeTopologyInstalled(int topologyId, boolean startRebalance, ConsistentHash previousWriteCh,
ConsistentHash newWriteCh) {
}
protected CompletionStage handleSegments(boolean startRebalance, IntSet addedSegments,
IntSet removedSegments) {
if (addedSegments.isEmpty()) {
return CompletableFutures.completedNull();
}
// add transfers for new or restarted segments
return addTransfers(addedSegments);
}
protected boolean notifyEndOfStateTransferIfNeeded() {
if (waitingForState.get()) {
if (hasActiveTransfers()) {
if (trace)
log.tracef("No end of state transfer notification, active transfers still exist");
return false;
}
if (waitingForState.compareAndSet(true, false)) {
int topologyId = stateTransferTopologyId.get();
log.debugf("Finished receiving of segments for cache %s for topology %d.", cacheName, topologyId);
stopApplyingState(topologyId);
stateTransferFuture.complete(null);
}
if (trace)
log.tracef("No end of state transfer notification, waitingForState already set to false by another thread");
return false;
}
if (trace)
log.tracef("No end of state transfer notification, waitingForState already set to false by another thread");
return true;
}
protected IntSet getOwnedSegments(ConsistentHash consistentHash) {
Address address = rpcManager.getAddress();
return consistentHash.getMembers().contains(address) ? IntSets.from(consistentHash.getSegmentsForOwner(address))
: IntSets.immutableEmptySet();
}
@Override
public CompletionStage applyState(final Address sender, int topologyId, boolean pushTransfer,
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 CompletableFutures.completedNull();
}
// 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_STATE_TRANSFER_IN_PROGRESS && !pushTransfer) {
log.debugf("Discarding state response with topology id %d for cache %s, we don't have a state transfer in progress",
topologyId, cacheName);
return CompletableFutures.completedNull();
}
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 CompletableFutures.completedNull();
}
if (trace) {
log.tracef("Before applying the received state the data container of cache %s has %d keys", cacheName,
dataContainer.sizeIncludingExpired());
}
IntSet mySegments = IntSets.from(wCh.getSegmentsForOwner(rpcManager.getAddress()));
Iterator iterator = stateChunks.iterator();
return applyStateIteration(sender, pushTransfer, mySegments, iterator).whenComplete((v, t) -> {
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 CompletionStage applyStateIteration(Address sender, boolean pushTransfer, IntSet mySegments,
Iterator iterator) {
CompletionStage chunkStage = CompletableFutures.completedNull();
// Replace recursion with iteration if the state was applied synchronously
while (iterator.hasNext() && CompletionStages.isCompletedSuccessfully(chunkStage)) {
StateChunk stateChunk = iterator.next();
if (pushTransfer) {
// push-transfer is specific for scattered cache but this is the easiest way to integrate it
chunkStage = doApplyState(sender, stateChunk.getSegmentId(), stateChunk.getCacheEntries());
} else {
chunkStage = applyChunk(sender, mySegments, stateChunk);
}
}
if (!iterator.hasNext())
return chunkStage;
return chunkStage.thenCompose(v -> applyStateIteration(sender, pushTransfer, mySegments, iterator));
}
private CompletionStage applyChunk(Address sender, IntSet mySegments, StateChunk stateChunk) {
if (!mySegments.contains(stateChunk.getSegmentId())) {
log.debugf("Discarding received cache entries for segment %d of cache %s because they do not belong to this node.", stateChunk.getSegmentId(), cacheName);
return CompletableFutures.completedNull();
}
// Notify the inbound task that a chunk of cache entries was received
InboundTransferTask inboundTransfer;
synchronized (transferMapsLock) {
List inboundTransfers = transfersBySegment.get(stateChunk.getSegmentId());
if (inboundTransfers != null) {
inboundTransfer = inboundTransfers.stream().filter(task -> task.getSource().equals(sender)).findFirst().orElse(null);
} else {
inboundTransfer = null;
}
}
if (inboundTransfer != null) {
return doApplyState(sender, stateChunk.getSegmentId(), stateChunk.getCacheEntries())
.thenAccept(
v -> inboundTransfer.onStateReceived(stateChunk.getSegmentId(), stateChunk.isLastChunk()));
} else {
if (cache.wired().getStatus().allowInvocations()) {
log.ignoringUnsolicitedState(sender, stateChunk.getSegmentId(), cacheName);
}
}
return CompletableFutures.completedNull();
}
private CompletionStage doApplyState(Address sender, int segmentId,
Collection> cacheEntries) {
if (cacheEntries == null || cacheEntries.isEmpty())
return CompletableFutures.completedNull();
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;
if (transactional) {
Object key = NO_KEY;
Transaction transaction = new ApplyStateTransaction();
InvocationContext ctx = icf.createInvocationContext(transaction, false);
LocalTransaction localTransaction = ((LocalTxInvocationContext) ctx).getCacheTransaction();
try {
localTransaction.setStateTransferFlag(PUT_FOR_STATE_TRANSFER);
for (InternalCacheEntry e : cacheEntries) {
// CallInterceptor will preserve the timestamps if the metadata is an InternalMetadataImpl instance
key = e.getKey();
CompletableFuture future = invokePut(segmentId, ctx, e);
if (!future.isDone()) {
throw new IllegalStateException("State transfer in-tx put should always be synchronous");
}
}
} catch (Throwable t) {
logApplyException(t, key);
return invokeRollback(localTransaction).handle((rv, t1) -> {
transactionTable.removeLocalTransaction(localTransaction);
if (t1 != null) {
t.addSuppressed(t1);
}
return null;
});
}
return invoke1PCPrepare(localTransaction).whenComplete((rv, t) -> {
transactionTable.removeLocalTransaction(localTransaction);
if (t != null) {
logApplyException(t, NO_KEY);
}
});
} else {
// non-tx cache
AggregateCompletionStage aggregateStage = CompletionStages.aggregateCompletionStage();
for (InternalCacheEntry e : cacheEntries) {
InvocationContext ctx = icf.createSingleKeyNonTxInvocationContext();
CompletionStage putStage = invokePut(segmentId, ctx, e);
aggregateStage.dependsOn(putStage.exceptionally(t -> {
logApplyException(t, e.getKey());
return null;
}));
}
return aggregateStage.freeze();
}
}
private CompletionStage invoke1PCPrepare(LocalTransaction localTransaction) {
PrepareCommand prepareCommand;
if (Configurations.isTxVersioned(configuration)) {
prepareCommand = commandsFactory.buildVersionedPrepareCommand(localTransaction.getGlobalTransaction(),
localTransaction.getModifications(), true);
} else {
prepareCommand = commandsFactory.buildPrepareCommand(localTransaction.getGlobalTransaction(),
localTransaction.getModifications(), true);
}
LocalTxInvocationContext ctx = icf.createTxInvocationContext(localTransaction);
return interceptorChain.invokeAsync(ctx, prepareCommand);
}
private CompletionStage invokeRollback(LocalTransaction localTransaction) {
RollbackCommand prepareCommand = commandsFactory.buildRollbackCommand(localTransaction.getGlobalTransaction());
LocalTxInvocationContext ctx = icf.createTxInvocationContext(localTransaction);
return interceptorChain.invokeAsync(ctx, prepareCommand);
}
private CompletableFuture invokePut(int segmentId, InvocationContext ctx, InternalCacheEntry e) {
// CallInterceptor will preserve the timestamps if the metadata is an InternalMetadataImpl instance
InternalMetadataImpl metadata = new InternalMetadataImpl(e);
PutKeyValueCommand put = commandsFactory.buildPutKeyValueCommand(e.getKey(), e.getValue(), segmentId,
metadata, STATE_TRANSFER_FLAGS);
put.setInternalMetadata(e.getInternalMetadata());
ctx.setLockOwner(put.getKeyLockOwner());
return interceptorChain.invokeAsync(ctx, put);
}
private void logApplyException(Throwable t, Object key) {
if (!cache.wired().getStatus().allowInvocations()) {
log.tracef("Cache %s is shutting down, stopping state transfer", cacheName);
} else {
log.problemApplyingStateForKey(t.getMessage(), key, t);
}
}
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) {
try {
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);
} catch (Throwable t) {
if (trace)
log.tracef(t, "Failed to create remote transaction %s", gtx);
}
}
}
if (tx != null) {
transactionInfo.getLockedKeys().forEach(tx::addBackupLockForKey);
}
}
}
}
// Must run after the PersistenceManager
@Start(priority = 20)
public void start() {
cacheName = cache.wired().getName();
isInvalidationMode = configuration.clustering().cacheMode().isInvalidation();
isTransactional = configuration.transaction().transactionMode().isTransactional();
timeout = configuration.clustering().stateTransfer().timeout();
CacheMode mode = configuration.clustering().cacheMode();
isFetchEnabled = mode.needsStateTransfer() &&
(configuration.clustering().stateTransfer().fetchInMemoryState() || configuration.persistence().fetchPersistentState());
rpcOptions = new RpcOptions(DeliverOrder.NONE, timeout, TimeUnit.MILLISECONDS);
stateRequestExecutor = new LimitedExecutor("StateRequest-" + cacheName, nonBlockingExecutor, 1);
running = true;
}
@Stop(priority = 0)
@Override
public void stop() {
if (trace) {
log.tracef("Shutting down StateConsumer of cache %s on node %s", cacheName, rpcManager.getAddress());
}
running = false;
try {
synchronized (transferMapsLock) {
// cancel all inbound transfers
// make a copy and then clear both maps so that cancel doesn't interfere with the iteration
Collection> transfers = new ArrayList<>(transfersBySource.values());
transfersBySource.clear();
transfersBySegment.clear();
for (List inboundTransfers : transfers) {
inboundTransfers.forEach(InboundTransferTask::cancel);
}
}
stateRequestExecutor.shutdownNow();
} catch (Throwable t) {
log.errorf(t, "Failed to stop StateConsumer of cache %s on node %s", cacheName, rpcManager.getAddress());
}
}
public void setKeyInvalidationListener(KeyInvalidationListener keyInvalidationListener) {
this.keyInvalidationListener = keyInvalidationListener;
}
// not used in scattered cache
private CompletionStage addTransfers(IntSet 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<>();
CompletionStage stage = CompletableFutures.completedNull();
if (isTransactional) {
stage = requestTransactions(segments, sources, excludedSources);
}
if (isFetchEnabled) {
stage = stage.thenRun(() -> requestSegments(segments, sources, excludedSources));
}
return stage;
}
private void findSources(IntSet segments, Map sources, Set excludedSources) {
if (cache.wired().getStatus().isTerminated())
return;
int numSegments = configuration.clustering().hash().numSegments();
IntSet segmentsWithoutSource = IntSets.mutableEmptySet(numSegments);
for (PrimitiveIterator.OfInt iter = segments.iterator(); iter.hasNext(); ) {
int segmentId = iter.nextInt();
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) {
IntSet segmentsFromSource = sources.computeIfAbsent(source, k -> IntSets.mutableEmptySet(numSegments));
segmentsFromSource.set(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 CompletionStage requestTransactions(IntSet segments, Map sources,
Set excludedSources) {
// TODO Remove excludedSources and always only for transactions/segments from the primary owner
findSources(segments, sources, excludedSources);
AggregateCompletionStage aggregateStage = CompletionStages.aggregateCompletionStage();
IntSet failedSegments = IntSets.concurrentSet(configuration.clustering().hash().numSegments());
Set sourcesToExclude = ConcurrentHashMap.newKeySet();
int topologyId = cacheTopology.getTopologyId();
for (Map.Entry sourceEntry : sources.entrySet()) {
Address source = sourceEntry.getKey();
IntSet segmentsFromSource = sourceEntry.getValue();
CompletionStage sourceStage = getTransactions(source, segmentsFromSource, topologyId)
.whenComplete((response, throwable) -> processTransactionsResponse(segments, sources, failedSegments, sourcesToExclude, topologyId,
source, segmentsFromSource, response, throwable));
aggregateStage.dependsOn(sourceStage);
}
return aggregateStage.freeze().thenCompose(ignored -> {
if (failedSegments.isEmpty()) {
return CompletableFutures.completedNull();
}
excludedSources.addAll(sourcesToExclude);
// look for other sources for all failed segments
sources.clear();
findSources(failedSegments, sources, excludedSources);
return requestTransactions(segments, sources, excludedSources);
});
}
private void processTransactionsResponse(IntSet segments, Map sources, IntSet failedSegments,
Set sourcesToExclude, int topologyId, Address source,
IntSet segmentsFromSource, Response response, Throwable throwable) {
boolean failed = false;
boolean exclude = false;
if (throwable != null) {
if (cache.wired().getStatus().isTerminated()) {
log.debugf("Cache %s has stopped while requesting transactions", cacheName);
sources.clear();
return;
} else {
log.failedToRetrieveTransactionsForSegments(cacheName, source, segments, throwable);
}
// The primary owner is still in the cluster, so we can't exclude it - see ISPN-4091
failed = true;
}
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;
}
// If requesting the transactions failed we need to retry
if (failed) {
failedSegments.addAll(segmentsFromSource);
}
if (exclude) {
sourcesToExclude.add(source);
}
}
private CompletionStage>> getClusterListeners(
int topologyId, List sources) {
// Try the first member. If the request fails, fall back to the second member and so on.
if (sources.isEmpty()) {
if (trace) // TODO Ignore self again
log.trace("Unable to acquire cluster listeners from other members, assuming none are present");
return CompletableFuture.completedFuture(Collections.emptySet());
}
Address source = sources.get(0);
// Don't send the request to self
if (sources.get(0).equals(rpcManager.getAddress())) {
return getClusterListeners(topologyId, sources.subList(1, sources.size()));
}
if (trace)
log.tracef("Requesting cluster listeners of cache %s from node %s", cacheName, sources);
CacheRpcCommand cmd = commandsFactory.buildStateTransferGetListenersCommand(topologyId);
CompletionStage remoteStage =
rpcManager.invokeCommand(source, cmd, SingleResponseCollector.validOnly(), rpcOptions);
return handleAndCompose(remoteStage, (response, throwable) -> {
if (throwable != null) {
log.exceptionDuringClusterListenerRetrieval(source, throwable);
}
if (response instanceof SuccessfulResponse) {
return CompletableFuture.completedFuture(
(Collection>) response.getResponseValue());
} else {
log.unsuccessfulResponseForClusterListeners(source, response);
return getClusterListeners(topologyId, sources.subList(1, sources.size()));
}
});
}
private CompletionStage getTransactions(Address source, IntSet segments, int topologyId) {
if (trace) {
log.tracef("Requesting transactions from node %s for segments %s", source, segments);
}
// get transactions and locks
CacheRpcCommand cmd = commandsFactory.buildStateTransferGetTransactionsCommand(topologyId, segments);
return rpcManager.invokeCommand(source, cmd, PassthroughSingleResponseCollector.INSTANCE, rpcOptions);
}
// not used in scattered cache
private void requestSegments(IntSet segments, Map sources, Set excludedSources) {
if (sources.isEmpty()) {
findSources(segments, sources, excludedSources);
}
for (Map.Entry e : sources.entrySet()) {
addTransfer(e.getKey(), e.getValue());
}
if (trace) log.tracef("Finished adding inbound state transfer for segments %s", segments, cacheName);
}
/**
* Cancel transfers for segments we no longer own.
*
* @param removedSegments segments to be cancelled
*/
protected void cancelTransfers(IntSet removedSegments) {
synchronized (transferMapsLock) {
List segmentsToCancel = new ArrayList<>(removedSegments);
while (!segmentsToCancel.isEmpty()) {
int segmentId = segmentsToCancel.remove(0);
List inboundTransfers = transfersBySegment.get(segmentId);
if (inboundTransfers != null) { // we need to check the transfer was not already completed
for (InboundTransferTask inboundTransfer : inboundTransfers) {
IntSet cancelledSegments = IntSets.mutableCopyFrom(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);
}
}
}
}
}
}
protected CompletionStage removeStaleData(final IntSet removedSegments) {
// Invalidation doesn't ever remove stale data
if (configuration.clustering().cacheMode().isInvalidation()) {
return CompletableFutures.completedNull();
}
log.debugf("Removing no longer owned entries for cache %s", cacheName);
if (keyInvalidationListener != null) {
keyInvalidationListener.beforeInvalidation(removedSegments, IntSets.immutableEmptySet());
}
// This has to be invoked before removing the segments on the data container
localPublisherManager.segmentsLost(removedSegments);
dataContainer.removeSegments(removedSegments);
// We have to invoke removeSegments above on the data container. This is always done in case if L1 is enabled. L1
// store removes all the temporary entries when removeSegments is invoked. However there is no reason to mess
// with the store if no segments are removed, so just exit early.
if (removedSegments.isEmpty())
return CompletableFutures.completedNull();
return persistenceManager.removeSegments(removedSegments)
.thenCompose(removed -> invalidateStaleEntries(removedSegments, removed));
}
private CompletionStage invalidateStaleEntries(IntSet removedSegments, Boolean removed) {
// If there are no stores that couldn't remove segments, we don't have to worry about invaliding entries
if (removed) {
return CompletableFutures.completedNull();
}
// All these segments have been removed from the data container, so we only care about private stores
AtomicLong removedEntriesCounter = new AtomicLong();
Predicate © 2015 - 2025 Weber Informatics LLC | Privacy Policy