Truly, I don't want to re-invent FTP. The key here is that the only channel I can rely on is SSH, and it is going to be way too much overhead to open up separate data channels for each binary object.

Objects vary in size from a few bytes to about 1GB, and there could easily be thousands in a single session.

I guess I could open up one extra data channel when the original connection is initiated, and then use syswrite/read on that, approximately like this:

Control channel:

C: STORE Name=<name> Content-Length=<length> MD5=<hash>
S: OK, go ahead
[download]

Then the client transmits the object in a series of syswrite() calls to the separate SSH. The separate process on the server would do sysread() only.

However, I would prefer to in-line it in one channel, to avoid the complexity of the extra connection and extra processes/threads. HTTP transmits a mixture of text and binary data pretty readily through a single socket.