In reflecting on the proposed Hyades protocol design, I’ve
got a couple of concerns about which I wanted to solicit feedback.
First, although we allow messages to be sent to the HCE
which contain a number of commands bundled together, there is no corresponding
way for multiple replies to be bundled together and sent out from the HCE, nor
can the commands be delivered in a single bundle, for instance, if they are all
directed to the same agent. All commands going out from the HCE go out
one-by-one. This is a consequence of the architecture we laid out for
separating transport layer, payload normalizer and command extractor. It
also seems to be a limitation of the current RAC implementation. I don’t
think this is a huge limitation, and it may be worth accepting in order to keep
the proposed architecture, but I wanted to specifically bring it up and see
what other people thought.
Second, we have so far not proposed a way for agents to
communicate with each other except by routing messages through the HCE,
although I think we’ve talked about it. Since agents will generally
be running on the same computer, it seems like it would sometimes be desirable
to have them talk directly to one another. Agents could negotiate such a
direct connection using their own proprietary commands, but if this sort of
thing will be common, then I think we might prefer to have the HCE broker the
conversation.
I would appreciate whatever perspectives on these issues any
of you can offer.
Thanks,
Andy