I think we've figured out the cause of everyone's problems with
THRIFT-837. The patch itself is fine; however, in fixing that bug, we've exposed the fact that PHPOutputTransport erroneously calls down to the underlying PHP transport's flush() method every time the internal 8k buffer is flushed. This is fine for the buffered transport, but unacceptable for the framed transport, which should only be flushed once per RPC call.
It seems like what we need to do is separate the "internal" buffer flushes from the "external" transport flushes. If we do that, everything should work out fine.