This change is a partial fix for the problem of long cluster failovers. This change addresses part of that problem: long newbie-broker catchup times.

At this time, it looks as though there are two distinct mechanisms causing these long catchup times.  They roughly divide into a population of times less than 20 seconds, and a population of much longer times (up to hundreds of seconds).

This fix addresses only the shorter times.  In a test of 25 failovers before and after the change, it reduced those times by an average of nearly 50%.  A T-test on the times indicates that the likelihood of the observed change happening randomly is less than 2%.


git-svn-id: https://svn.apache.org/repos/asf/qpid/trunk/qpid@954895 13f79535-47bb-0310-9956-ffa450edef68

1 files changed, 3 insertions, 1 deletions

diff --git a/cpp/src/qpid/cluster/Cluster.cpp b/cpp/src/qpid/cluster/Cluster.cpp
index 6b9fceccd9..9ca6fbf2bf 100644
--- a/cpp/src/qpid/cluster/Cluster.cpp
+++ b/cpp/src/qpid/cluster/Cluster.cpp
@@ -444,8 +444,10 @@ void Cluster::deliveredEvent(const Event& e) {
         EventFrame ef(e, e.getFrame());
         // Stop the deliverEventQueue on update offers.
         // This preserves the connection decoder fragments for an update.
+        // Only do this for the two brokers that are directly involved in this 
+        // offer: the one making the offer, or the one receiving it.
         const ClusterUpdateOfferBody* offer = castUpdateOffer(ef.frame.getBody());
-        if (offer) {
+        if (offer && ( e.getMemberId() == self || MemberId(offer->getUpdatee()) == self) ) {
             QPID_LOG(info, *this << " stall for update offer from " << e.getMemberId()
                      << " to " << MemberId(offer->getUpdatee()));
             deliverEventQueue.stop();