I am very interested in why these improvements somehow only affect the CQL interface and not the Thrift interface: they sound like things that would affect core parts of query execution, well below the interface used to describe the query.

The most important change for this post was the reduction of overhead for processing each message off the wire, through improved thread scheduling/communication. The changes do not affect CQL, so much as the 'native' protocol - CQL over thrift is no faster/slower.

Ok, thank you: saying it affects the native protocol networking layer (and would not affect CQL over Thrift) totally makes it click for me; the post made it sound like CQL itself was affected. Again, thanks!