A brief performance analysis highlighted a significant impact on performance because of the use of ThreadLocals to hold the enabled flags in the current context.

We should check these findings with a proper test and provide an alternative way, ideally one which doesn't need to lookup in the flagHolder either (avoid both the put and the get in the threadlocal).

We'll need a method which returns a new Cache instance having the mentioned Flags always implicitly added to the invocation context. I think it would be fine in such an alternative implementation to not support dynamic flags.