Optimization to reuse preexisting buffers across requests. Doing this in a few places but there's a few more around.
Depends on #10
At this point the code is in pretty good shape. Could potentially avoid copies in a few additional places, but its really in the territory where e.g. mutex overhead for reusing a buffer across contexts would outweigh any benefit from avoiding the malloc.
In other words - this is deep in early optimization territory, lets wait until a flame chart says its a problem.