To make sure I understand what you are saying, to rephase what we have are the following translators:
What you are getting at, which we want is:
Infinispan - likely an extension of Map where whether to use lucence is based upon the table metadata and perhaps even the query (for example if only equi primary key criteria is specified, then it should be better to query the cache rather than use lucene).
All of the other concerns about cache lookup should be abstracted - either through a ResourceAdapter, JNDI, or container service injection.
Currently, to my knowledge, there isn't a ResourceAdapter , which is what I'm proposing creating (i.e., infinispan-connector).
So now the current state reflects two translator types using a common connection interface ObjectConnection. We'll develop the connection strategies for that on the TEIID-2210 issue.