[zeromq-dev] HWM ignored on inproc connection

[Martin Sustrik]

On 01/07/2011 02:14 AM, Douglas Creager wrote:
>>> Please, do check whether everyting works as expected.
>>
>> Hm, presumably, some more patching is needed to cover the cases
>> when one or both hwms/swaps is infinite. Can you possibly do that?
>
> I tried this patch out, and it works as expected.  For infinite HWMs,
> I think it works, depending on what we want the behavior to be.  If
> one side has a finite HWM, and the other an infinite, and we want the
> finite one to take precedence, then it works: X+0 = X.  And two
> infinite HWMs stay infinite: 0+0=0.
>
> That said, the more I think about this, the more I'm not sure about
> it.  With this patch in place, a PULL inproc socket can affect the
> HWM buffer that's seen by the PUSH socket on the other side.  For a
> TCP connection, it can't: there's an HWM buffer on the PULL side, but
> since it's a PULL socket, it's ignored.  And I don't think that's
> something that breaking HWM into SNDHWM/RCVHWM would fix, because the
> PULL socket could still affect the PUSH's SNDHWM by setting its
> RCVHWM.  Am I missing something obvious?  Or does the processing need
> to be aware of the socket types being connected to do the right
> thing?

I would say that infinite + X should be infinite rather than X.

Martin