This is connected mainly to RollingUpdates. (See https://community.jboss.org/wiki/RollingUpgradesInInfinispan, part 1, step 3)

Unmarshalling process seems not to be triggered when issuing get on NEW cluster cache to get OLD entry which was stored into OLD cluster cache before new cluster starts. (this OLD is configured to be RemoteCacheStore for NEW now)

Remote cache store seems to be configured well. When NEW cache is conf. with eviction, entries are successfully evicted to OLD cache = RCS now. Then they are accessible from NEW cache ok. (Consulting remote cache store = OLD cache) See upper part of trace log fragment [1].

OLD entries which was stored before NEW cluster starts seems NOT to be accessible. See lower part of [1] with asterisk in-line comments.

There arise some questions to answer here:
1) Why there is no unmarshalling process there?
2) Then is case of getting old entry, why is there no question asking: Entry exists in loader?

Note: Cache misses on new as well as old cache increase. It looks like new cache is consulting its cache store for the entry but for the WRONG entry (key which is not unmarshalled correctly).

Feel free to comment, mail me, ping me for any other info which is needed.

[1]: http://pastebin.com/ep0iziyz