import com.google.common.annotations.Beta;
import com.google.common.base.Preconditions;
+import com.google.common.base.Stopwatch;
import com.google.common.base.Verify;
+import java.util.Collection;
import java.util.Map;
+import java.util.Optional;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.TimeoutException;
import javax.annotation.Nullable;
import javax.annotation.concurrent.GuardedBy;
import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
+import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
import org.opendaylight.controller.cluster.access.concepts.FailureEnvelope;
import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
* @param enqueuedEntries Previously-enqueued entries
* @return A {@link ReconnectForwarder} to handle any straggler messages which arrive after this method returns.
*/
- @Nonnull ReconnectForwarder finishReconnect(@Nonnull Iterable<ConnectionEntry> enqueuedEntries);
+ @Nonnull ReconnectForwarder finishReconnect(@Nonnull Collection<ConnectionEntry> enqueuedEntries);
}
private static final Logger LOG = LoggerFactory.getLogger(ClientActorBehavior.class);
- private static final FiniteDuration RESOLVE_RETRY_DURATION = FiniteDuration.apply(5, TimeUnit.SECONDS);
+ private static final FiniteDuration RESOLVE_RETRY_DURATION = FiniteDuration.apply(1, TimeUnit.SECONDS);
/**
* Map of connections to the backend. This map is concurrent to allow lookups, but given complex operations
}
private ClientActorBehavior<T> internalOnRequestFailure(final FailureEnvelope command) {
+ final AbstractClientConnection<T> conn = getConnection(command);
+ if (conn != null) {
+ /*
+ * We are talking to multiple actors, which may be lagging behind our state significantly. This has
+ * the effect that we may be receiving responses from a previous connection after we have created a new
+ * one to a different actor.
+ *
+ * Since we are already replaying requests to the new actor, we want to ignore errors reported on the old
+ * connection -- for example NotLeaderException, which must not cause a new reconnect. Check the envelope's
+ * sessionId and if it does not match our current connection just ignore it.
+ */
+ final Optional<T> optBackend = conn.getBackendInfo();
+ if (optBackend.isPresent() && optBackend.get().getSessionId() != command.getSessionId()) {
+ LOG.debug("{}: Mismatched current connection {} and envelope {}, ignoring response", persistenceId(),
+ conn, command);
+ return this;
+ }
+ }
+
final RequestFailure<?, ?> failure = command.getMessage();
final RequestException cause = failure.getCause();
if (cause instanceof RetiredGenerationException) {
return null;
}
if (cause instanceof NotLeaderException) {
- final AbstractClientConnection<T> conn = getConnection(command);
if (conn instanceof ReconnectingClientConnection) {
// Already reconnecting, do not churn the logs
return this;
} else if (conn != null) {
LOG.info("{}: connection {} indicated no leadership, reconnecting it", persistenceId(), conn, cause);
- return conn.reconnect(this);
+ return conn.reconnect(this, cause);
+ }
+ }
+ if (cause instanceof OutOfSequenceEnvelopeException) {
+ if (conn instanceof ReconnectingClientConnection) {
+ // Already reconnecting, do not churn the logs
+ return this;
+ } else if (conn != null) {
+ LOG.info("{}: connection {} indicated sequencing mismatch on {} sequence {} ({}), reconnecting it",
+ persistenceId(), conn, failure.getTarget(), failure.getSequence(), command.getTxSequence(), cause);
+ return conn.reconnect(this, cause);
}
}
@GuardedBy("connectionsLock")
@Nonnull protected abstract ConnectionConnectCohort connectionUp(@Nonnull ConnectedClientConnection<T> newConn);
- private void backendConnectFinished(final Long shard, final AbstractClientConnection<T> conn,
+ private void backendConnectFinished(final Long shard, final AbstractClientConnection<T> oldConn,
final T backend, final Throwable failure) {
if (failure != null) {
if (failure instanceof TimeoutException) {
- if (!conn.equals(connections.get(shard))) {
+ if (!oldConn.equals(connections.get(shard))) {
// AbstractClientConnection will remove itself when it decides there is no point in continuing,
// at which point we want to stop retrying
- LOG.info("{}: stopping resolution of shard {} on stale connection {}", persistenceId(), shard, conn,
- failure);
+ LOG.info("{}: stopping resolution of shard {} on stale connection {}", persistenceId(), shard,
+ oldConn, failure);
return;
}
LOG.debug("{}: timed out resolving shard {}, scheduling retry in {}", persistenceId(), shard,
RESOLVE_RETRY_DURATION, failure);
context().executeInActor(b -> {
- resolveConnection(shard, conn);
+ resolveConnection(shard, oldConn);
return b;
}, RESOLVE_RETRY_DURATION);
return;
cause = new RuntimeRequestException("Failed to resolve shard " + shard, failure);
}
- conn.poison(cause);
+ oldConn.poison(cause);
return;
}
LOG.info("{}: resolved shard {} to {}", persistenceId(), shard, backend);
final long stamp = connectionsLock.writeLock();
try {
+ final Stopwatch sw = Stopwatch.createStarted();
+
// Create a new connected connection
- final ConnectedClientConnection<T> newConn = new ConnectedClientConnection<>(conn.context(),
- conn.cookie(), backend);
- LOG.info("{}: resolving connection {} to {}", persistenceId(), conn, newConn);
+ final ConnectedClientConnection<T> newConn = new ConnectedClientConnection<>(oldConn, backend);
+ LOG.info("{}: resolving connection {} to {}", persistenceId(), oldConn, newConn);
// Start reconnecting without the old connection lock held
final ConnectionConnectCohort cohort = Verify.verifyNotNull(connectionUp(newConn));
// Lock the old connection and get a reference to its entries
- final Iterable<ConnectionEntry> replayIterable = conn.startReplay();
+ final Collection<ConnectionEntry> replayIterable = oldConn.startReplay();
// Finish the connection attempt
final ReconnectForwarder forwarder = Verify.verifyNotNull(cohort.finishReconnect(replayIterable));
+ // Cancel sleep debt after entries were replayed, before new connection starts receiving.
+ newConn.cancelDebt();
+
// Install the forwarder, unlocking the old connection
- conn.finishReplay(forwarder);
+ oldConn.finishReplay(forwarder);
// Make sure new lookups pick up the new connection
- connections.replace(shard, conn, newConn);
- LOG.info("{}: replaced connection {} with {}", persistenceId(), conn, newConn);
+ if (!connections.replace(shard, oldConn, newConn)) {
+ final AbstractClientConnection<T> existing = connections.get(oldConn.cookie());
+ LOG.warn("{}: old connection {} does not match existing {}, new connection {} in limbo",
+ persistenceId(), oldConn, existing, newConn);
+ } else {
+ LOG.info("{}: replaced connection {} with {} in {}", persistenceId(), oldConn, newConn, sw);
+ }
} finally {
connectionsLock.unlockWrite(stamp);
}
}
void removeConnection(final AbstractClientConnection<?> conn) {
- connections.remove(conn.cookie(), conn);
- LOG.debug("{}: removed connection {}", persistenceId(), conn);
+ final long stamp = connectionsLock.writeLock();
+ try {
+ if (!connections.remove(conn.cookie(), conn)) {
+ final AbstractClientConnection<T> existing = connections.get(conn.cookie());
+ if (existing != null) {
+ LOG.warn("{}: failed to remove connection {}, as it was superseded by {}", persistenceId(), conn,
+ existing);
+ } else {
+ LOG.warn("{}: failed to remove connection {}, as it was not tracked", persistenceId(), conn);
+ }
+ } else {
+ LOG.info("{}: removed connection {}", persistenceId(), conn);
+ }
+ } finally {
+ connectionsLock.unlockWrite(stamp);
+ }
}
@SuppressWarnings("unchecked")
final ReconnectingClientConnection<T> conn = (ReconnectingClientConnection<T>)newConn;
LOG.info("{}: connection {} reconnecting as {}", persistenceId(), oldConn, newConn);
- final boolean replaced = connections.replace(oldConn.cookie(), (AbstractClientConnection<T>)oldConn, conn);
- if (!replaced) {
- final AbstractClientConnection<T> existing = connections.get(oldConn.cookie());
- LOG.warn("{}: old connection {} does not match existing {}, new connection {} in limbo", persistenceId(),
- oldConn, existing, newConn);
+ final long stamp = connectionsLock.writeLock();
+ try {
+ final boolean replaced = connections.replace(oldConn.cookie(), (AbstractClientConnection<T>)oldConn, conn);
+ if (!replaced) {
+ final AbstractClientConnection<T> existing = connections.get(oldConn.cookie());
+ if (existing != null) {
+ LOG.warn("{}: failed to replace connection {}, as it was superseded by {}", persistenceId(), conn,
+ existing);
+ } else {
+ LOG.warn("{}: failed to replace connection {}, as it was not tracked", persistenceId(), conn);
+ }
+ }
+ } finally {
+ connectionsLock.unlockWrite(stamp);
}
final Long shard = oldConn.cookie();
Code Review / controller.git / blobdiff