I think the cassandra team did not re-implement their own GC. I guess what they meant is the less heap being used, the better GC performance. 

AFAIK, only data that is not been updated frequently can benefit from off-heap row cache, because when a row is modified, the row inside cache need be invalidated.

On Mon, Sep 19, 2011 at 9:56 AM, Yang <teddyyyy123@gmail.com> wrote:
this comment in JIRA mentions it

but in the end,  it's not immediately clear. could someone give a
summary of its advantages?
(if it's just re-implement your own GC, why would your own GC be better?)