I think TokenAwarePolicy has a typo introducing a bug, added a comment.

Metadata#getReplicas(String keyspace, TokenRange range) now creates a new HashSet copy rather than return directly. Metadata#getReplicas(String keyspace, Token token) now returns a copy too TokenAwarePolicy#newQueryPlan(final String loggedKeyspace, final Statement) Looks like for token map does a copy and for tablet map does not. Previously it would make a copy for both cases so this is partial improvement.

Overall net -1 collection creations for tablet map but +1 for token map (plus O(1) instanceof check for both) when we call newQueryPlan. For other policies it should be net 0 or +1 in that case. If we don't care as much about token map performance then this is maybe an ok sacrifice to make.

It would also be good to update the javadocs to mention where methods return a shallow copy and where not. Previously they would return underlying token maps collections.

I don't think that it makes sense, it is only confuse API user, let's just state that you are not suppose to write to it, which we alredy do.

getReplicas - marked as depricated in favor of getReplicasList it is not used in the driver internals, only tests.
TabletMap does not return copy, it recreates it, later when/if we change it to contain host instance, it will return valu from tabletmap.

getReplicas - marked as depricated in favor of getReplicasList it is not used in the driver internals, only tests.

You're right, this is actually not on the path to newQueryPlan, so this does not impact it. It uses getReplicas from TokenMap which is defined inside Metadata too, but those are different methods.

@Bouncheck , please check on it.

See comments

All adressed

Generally LGTM, although this does not avoid converting into ArrayList in the end, which i thought was the purpose.

Thanks, it will be addressed in another PR after we introduce bench tests