Is that really the case? That the server will somehow migrate sessions across versions? Or will it be more of a "current users use version 1 of the app, new users route to version 2, when all users are done with version 1 it can be safely undeployed". That's a different approach to the problem.
The first technique also implies some kind of migration lifecycle event that is going to allow the application to migrate the session (otherwise, it's a Bad Idea), the second technique mitigates that problem since it will only see "new" sessions and not have to convert old ones.
Finally, of course, there is the issue of having to overscale the server to handle both workloads (the old and new application simultaneously), so that's an impact that will have to be considered when using either technique. In theory the user load would "be the same", but you have two instances of the server running (and there may be substantial caching and other memory pressure).
Also, what consideration is being given to if I have two versions of an application within a cluster, version 1 on instance A and version 2 on instance B and the ramifications regarding session failover from a version 1 instance to a version 2 instance, or do we lose that ability?
[Message sent by forum member 'whartung' (whartung)]
http://forums.java.net/jive/thread.jspa?messageID=265855