While these ideas was brewing, I ventured down into many libraries, Camel was one of them. It did not solve any of my problems.

Mats is messaging with a call stack. One messaging endpoint can invoke another messaging endpoint, and get the reply supplied to its next stage, with the state from the previous stage magically present: https://mats3.io/docs/message-oriented-rpc/#isc-using-mats3 and other pages.

Camel has, AFAIK, nothing of the kind.

It's interesting how we can have such different takes on it. Still everything you say maps naturally onto Camel for me.

Your example using Camel would look something like:

    from('rest:/api/myprocess') // using REST API endpoint as an example starting point
      .inOut()
      .to('activemq:main')
      .to('activemq:mid')
      .to('activemq:leaf')
    .end()

Then you would define the handlers:

   from('activemq:main')
      .transform { e ->
            return new State(...)
      }
      
   from('activemq:mid')
      .transform { e ->
            e.body.number1 = 
      }

All the processing stages are happening asynchronously and results are passed back and forth by Camel. If things error out half way through you get a printout from Camel of the whole workflow including each step, where exactly it failed and what the content of the state was at that point (and it'll throw the lost message onto DLQ if you want to re-do the process).

Mats does not explicitly define a set "route" through a bunch of stages. I view Camel more like a Workflow system, which is pointed out elsewhere in the discussion threads here - the workflows are defined external to the actual processing steps.

With Mats, you define Endpoints, which can be generic like a REST endpoint can be generic: "AccountService.getAccountListForCustomer" - which can be used (by invocation) by sevaral other endpoints.

Also, IIRC, the process you show with Camel is now effectively thread bound - or at least JVM bound. If the node running that process goes down, it takes all unfinished, mid-flow, processes with it. The steps are not transactional.

With Mats, the "life" of the process lives in the message. You can literally bring down your entire system from GCP, and rebuild it on Azure, and then when you start it up again, the mid-flow processes will just continue and finish as if nothing happened - as long as you brought along the state from the message broker (in addition to your databases). Viewed "from within the process", the only thing visible is that there was a bit more latency between one step and the next than usual. AFAIR, you cannot get anything like this with Camel. The idea is of course not the ability to move clouds, but that the result is exceptionally robust and stable.