So I'm sitting here laughing a little bit, as I started to write down code to make this work. I realized that the REST interface supports most of this already, The key difference right now, acknowledgements are based on the last message received and its fully server stateful, doesn't seem to handle a multiple server deployment.
I know it was discussed a while ago about the compatibility of the REST API, while most of its functional, there's some call outs I see: - Separation of logic. Right now a lot of the logic is embedded directly in the rest endpoint classes, makes it hard to follow. - Refactor the actual API, or maybe introduce a new API that better supports messaging strategies people are using - JSON for instance. I'm inclined to start with this codebase, but introduce it as a new module. This way if anyone's using the REST API its not broken. It would give the opportunity to also move this into a protocol implementation as well. John On Thu, Nov 10, 2016 at 10:10 PM John D. Ament <johndam...@apache.org> wrote: > I also started summarizing and jotting down things in a JIRA ticket. > > https://issues.apache.org/jira/browse/ARTEMIS-847 > > John > > > On Thu, Nov 10, 2016 at 9:46 PM John D. Ament <johndam...@apache.org> > wrote: > > So, the ack process is something essentially there already in most > protocols. The difference is that you have a persistent TCP connection > from client to server. With HTTP, its a new socket each time. > > I suspect there's some level of node affinity that may exist in the code > that may be a problem for this. For instance, a "session" is no longer > based on that connection, but instead probably based on some subscriber ID > field being shared. I might even expect that this continues to work > independent of server you access. If you're in a cluster, the entire > cluster should be aware of the messages you're processing. An > acknowledgement sent to any node should work. > > With regard to leveraging the existing REST API. I'd be ok with this new > REST being its own independent protocol. It's a larger dev effort, but > ideally we could even move away from the very weird REST impl that's there > today. Basically, the way I see it, this feature adds acknowledgement to > the REST API, without assuming each message is auto acknowledged (at least > last I looked at the REST API, it was auto-ack). > > John > > > On Thu, Nov 10, 2016 at 7:23 PM Matt Pavlovich <mattr...@gmail.com> wrote: > > Sounds good, this sounds like an interesting use case. The one upside I > see of being a subscription strategy is that you could theoretically add > that behavior to all clients (that support some sort of ack) using any > protocol vs having it be REST-only. > > On 11/10/16 4:09 PM, John D. Ament wrote: > > What Martyn is describing is closer to what I'm thinking. It may even > make > > sense to refactor rest into a protocol as a part of this. > > > > I'll respond tonight with a few more details of what I was thinking. > > Thanks for the input so far guys! > > > > John > > > > On Nov 10, 2016 16:20, "Martyn Taylor" <mtay...@redhat.com> wrote: > > > > The Artemis REST API is already an independent service that layers on top > > of JMS. If we're adding this API to the REST module it'd be independent > > > > That said... this could be done as a protocol module (the protocol > modules > > are pluggable) when deployed it'd then be managed by the broker. Just > like > > AMQP or any other protocol. The benefit of doing it this way is that > you'd > > have more fine grained control via the internal CORE API. Also means you > > can plug in to the security layer etc... it does mean starting from > > scratch though... > > > > On 10 Nov 2016 21:01, "Matt Pavlovich" <mattr...@gmail.com> wrote: > > > >> How do you see the service layering on top of Artemis? A fully > > independent service with a "seen message id" repository, or a > subscription > > recovery-style policy within the broker with a REST service on top? > >> > >> On 11/9/16 11:54 AM, John D. Ament wrote: > >>> All, > >>> > >>> One thing I see come up quite often when looking at cloud based > messaging > >>> systems is the concept of a reservation (there's a couple of terms out > >>> there, reservation seems to describe it best). The reservation acts > like > >>> this: > >>> > >>> - Client polls for messages and get some number of messages back. > >>> - When a client polls again, those messages are not returned for some > >>> duration since it read them. > >>> - The messages are not auto-acknowledged. > >>> - A second API is invoked indicating that the client has acknowledged > > that > >>> message, typically using some message id or reservation id. > >>> - If after some duration, a message was not acknowledged, it becomes > >>> eligible for reception again. > >>> > >>> I'd like to add this type of capability to the REST API for Artemis. > > What > >>> do others think? > >>> > >>> John > >>> > >
