A possible outcome of this is to stick with Avro.

My main reason is that AFAIK we choose Avro because Infinispan had chosen it too; still Infinispan doesn't use it any longer and so because of our Serialization needs the Infinispan distribution is becoming a collection of various marshallers.

Still, I wouldn't propose to rethink our strategy as it has been working fine, but we need to revisit that code during the 5.0 works anyway: after the Lucene4 migration it's using a lot of deprecated code which needs to be removed soon, and I want to add several tests too. Even the performance tests need to be rewritten.