new cluster design: folded active-passive debate into main design doc.

git-svn-id: https://svn.apache.org/repos/asf/qpid/trunk@1029679 13f79535-47bb-0310-9956-ffa450edef68

2 files changed, 41 insertions, 65 deletions

diff --git a/qpid/cpp/design/new-cluster-active-passive.txt b/qpid/cpp/design/new-cluster-active-passive.txt
deleted file mode 100644
index f6d16d751f..0000000000
--- a/qpid/cpp/design/new-cluster-active-passive.txt
+++ /dev/null
@@ -1,64 +0,0 @@
--*-org-*-
-# Licensed to the Apache Software Foundation (ASF) under one
-# or more contributor license agreements.  See the NOTICE file
-# distributed with this work for additional information
-# regarding copyright ownership.  The ASF licenses this file
-# to you under the Apache License, Version 2.0 (the
-# "License"); you may not use this file except in compliance
-# with the License.  You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing,
-# software distributed under the License is distributed on an
-# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
-# KIND, either express or implied.  See the License for the
-# specific language governing permissions and limitations
-# under the License.
-
-* Active/passive cluster implementation
-
-The active-active cluster can be used in an active-passive mode. In
-this mode we would like the cluster to be as efficient as a strictly
-active-passive implementation.
-
-An active/passive implementation requires a subset of the active/active approach:
-- drop Queue ownership and locking
-- replicate subset of Cluster interface, don't need to know all message disposition.
-
-Can re-use:
-- cluster membership
-- new member updates
-- store integration
-
-Simpler implementation of broker::Cluster:
-- act like distributed MessageStore. Don't need acquisition details.
-- can do immediate local enqueue and still guarantee order.
-- can use smaller message IDs: just sequence number. Can be implicit.
-
-Extra requirements:
-- Exactly one broker hast to take over if primary fails.
-- Passive members refuse client connections and redirect to active member.
-- On failover, clients keep trying till they find the active member.
-
-** Active/active vs. active passive
-
-Active/active benefits:
-- Total # connections: practical 60k limit per node.
-- Handle client losing connectivity to one cluster node - can fail over to any.
-- Some load sharing: reading from client + multicast only done on direct node.
-- Clients can switch to any broker on failover
-- Failure affects a subset of the clients, not all the clients.
-
-Active/active drawbacks:
-- Co-ordinating message allocation impacts performance.
-
-Active/passive benefits:
-- Don't need message allocation strategy, can feed consumers at top speed.
-
-Active/passive drawbacks:
-- All clients on one node so a failure affects every client in the system.
-- After a failure there is a "reconnect storm" as every client reconnects to the new active node.
-- After a failure there may be a period where no broker is active.
-- Can't help clients with no connectivity to the active member.
-- Clients must find the single active to fail-over.
diff --git a/qpid/cpp/design/new-cluster-design.txt b/qpid/cpp/design/new-cluster-design.txt
index d8130b31db..7adb46fee3 100644
--- a/qpid/cpp/design/new-cluster-design.txt
+++ b/qpid/cpp/design/new-cluster-design.txt
@@ -371,14 +371,53 @@ Clustering and scalability: new design may give us the flexibility to
 address scalability as part of cluster design. Think about
 relationship to federation and "fragmented queues" idea.
 
-* Design questions/descisions
+* Design debates/descisions
+
+** Active/active vs. active passive
+
+An active-active cluster can be used in an active-passive mode. In
+this mode we would like the cluster to be as efficient as a strictly
+active-passive implementation.
+
+An active/passive implementation allows some simplifications over active/active:
+- drop Queue ownership and locking
+- don't need to replicate message acquisition.
+- can do immediate local enqueue and still guarantee order.
+
+Active/passive introduces a few extra requirements:
+- Exactly one broker hast to take over if primary fails.
+- Passive members must refuse client connections.
+- On failover, clients must re-try all known addresses till they find the active member.
+
+Active/active benefits:
+- A broker failure only affects the subset of clients connected to that broker.
+- Clients can switch to any other broker on failover
+- Backup brokers are immediately available on failover.
+- Some load sharing: reading from client + multicast only done on direct node.
+
+Active/active drawbacks:
+- Co-ordinating message acquisition may impact performance (not tested)
+- Code may be more complex that active/passive.
+
+Active/passive benefits:
+- Don't need message allocation strategy, can feed consumers at top speed.
+- Code may be simpler than active/active.
+
+Active/passive drawbacks:
+- All clients on one node so a failure affects every client in the system.
+- After a failure there is a "reconnect storm" as every client reconnects to the new active node.
+- After a failure there is a period where no broker is active, until the other brokers realize the primary is gone and agree on the new primary.
+- Clients must find the single active node, may involve multiple connect attempts.
+
 ** Total ordering.
+
 Initial thinking: allow message ordering to differ between brokers.
 New thinking: use CPG total ordering, get identical ordering on all brokers.
 - Allowing variation in order introduces too much chance of unexpected behavior.
 - Usign total order allows other optimizations, see Message Identifiers below.
 
 ** Message identifiers.
+
 Initial thinking: message ID = CPG node id + 64 bit sequence number.
 This involves a lot of mapping between cluster IDs and broker messsages.
 
@@ -387,6 +426,7 @@ New thinking: message ID = queue name + queue position.
 - Requires total ordering of messages (see above)
 
 ** Message rejection
+
 Initial thinking: add special reject/rejected points to cluster interface so
 rejected messages could be re-queued without multicast.