BUG-8445: check sessionId before propagating failures

When we have leader movement ocurring, based on timing details we
can re-establish a connection to the new leader and then start
receiving responses from the old leader telling us it no longer
is the leader.

To stop this from happening we need to check connection session ID
against the incoming failure.

Change-Id: If9a891016c7f213f2552283e3ec13485e598f5a4
Signed-off-by: Robert Varga <robert.varga@pantheon.tech>

diff --git a/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java b/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java
index 554ffe97c77ae7c3011daca2a28c22f660a20e78..ccfbba6cd3d94f92ef9757999352893e1b0f7270 100644 (file)
--- a/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java
+++ b/opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java
@@ -13,6 +13,7 @@ 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;
@@ -160,6 +161,25 @@ public abstract class ClientActorBehavior<T extends BackendInfo> extends
     }
 
     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) {
@@ -169,7 +189,6 @@ public abstract class ClientActorBehavior<T extends BackendInfo> extends
             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;
@@ -179,7 +198,6 @@ public abstract class ClientActorBehavior<T extends BackendInfo> extends
             }
         }
         if (cause instanceof OutOfSequenceEnvelopeException) {
-            final AbstractClientConnection<T> conn = getConnection(command);
             if (conn instanceof ReconnectingClientConnection) {
                 // Already reconnecting, do not churn the logs
                 return this;