On Sat, May 3, 2014 at 8:32 PM, Garrett D'Amore <garrett@xxxxxxxxxx> wrote: > There are some good reasons for this to be that way — not least of which is > that subscriber filtering is CPU intensive, and is frankly incompatible with You meant publisher filtering here? > multicasting on a medium that supports multicast. (Note that nanomsg > currently lacks multicast support, but I consider that possibly a temporary > limitation. While inappropriate for req/rep or push/pull or pair, > multicast media would be useful for pub/sub, bus, or my experimental star > protocol. Anyway, this isn’t about multicast — at least not really.) > Certainly, the more I think about this, the more I see the simple elegance > of just subscriber filtering, and the more I want to simply dispense with > publisher side filtering. But people are asking… > > Thoughts? It was implemented in crossroads and zeromq, I think. One of the use cases where I really liked it was that it was possible to do a PUB/SUB device (i.e. bridge) for e.g. a WAN linking several LANs, and it really made a lot of sense there. In my mind, publisher-side filtering is a better conceptual fit with the "publish-subscribe" idea; i.e. I think the natural expectation for publish/subscribe systems is that all of the unsubscribed messages should not traverse the link. Cheers, Dirkjan