[zeromq-dev] Feedback on new PATCH socket

[Martin Sustrik]

On 05/08/2011 07:25 AM, Fabien Ninoles wrote:

> I'm not sure about how 190 can be fixed by separating the semantic of
> the ROUTER from the req/rep socket; propagating a disconnection from
> one edge to another could be quite hard and probably hardly possible
> if you have any fairqueuing in-between.

That's not what 190 asks for. It's more of a heuristic: If next *hop* 
downstream disconnects, immediately drop any queued requests you've got 
from it. The rationale being that there's noone to send the replies to 
anymore anyway.

> However, one way to be more transparent is to mark a frame as being
> "added" to the original message instead of using the empty delimiter.
> This will allow REQ/REP to work directly without change.  The flag
> need to be pass between sockets however (just like the MORE flag) and
> so this is not compatible with the current queue device.  However, if
> the API were setting the flag inside the msg_t struct, this would
> have been completly transparent, will remove some potential beginner
> bug (like sending a RCVMORE msg without the SNDMORE flag in a device)
> and, I'm pretty sure we can make it semantically meaningful (by
> calling it FOLLOWED for example).

I would love to replace multi-part messages by message+labels system, 
however, this is not backward compatible and thus unlikely to happen in 
0MQ. You will probably find it in the future implementations such as 
Linux kernel implementation.

Martin