Github user Jens-G commented on the issue:
https://github.com/apache/thrift/pull/1402
> Why not actually use (2^16)-1 which is the limit?
Several reasons. First, aligned memory access is always faster. If we
subtract 1 byte, we get the worst case. Next, at least on Windows a number of
system resources use multiples of 4096, so it is probably not a bad idea to
follow that model. That's why I picked this particular value.---
