Summary of the discussion:
- No pages or at least they are not defined;
- People from CRU to be involved;
- No difference between single and multipart header. With single header the routing is simpler since we do not have to combine the global header.
MA-T: There is interesting discussion why the multipart messages are not implemented in nanomessage. The idea is that they belong to the application and not to the framework. They are based on protocol and not on transport. We are considering messagepack (something like protobuf) and mad eit available in the ALFA framework. It is faster than anything else. It implements the protocol and not the transport. In January Mohammad and Alexey will make detailed presentation.
Several questions:
- Do we want to connect to the transport layer? - No.
- Do we want monolithic application? - No, based on transport.
- Does message pack save a copy? - No, zero copy.
MR: I think this is very good idea. How would it look like in practice? For example if I want to send multipart message.
You should create a message pack and add all the parts to it. However the integration of message pack in FairMQ is not yet available.
MR: If you integrate message pack as new option in FairMQ, there will be no distinction between ZeroMQ and nanomsg. The disadvantage is that we do not have schema evolution. On the other hand, we do not need schema evolution for the headers and the payloads are treated separately.
