When evaluation begins, RuntimeManager asks to the plugin the plan of execution.
Then, RuntimeManager use the execution plan to coordinate different execution steps

Considerations:

1. happy path:
   1. module "A" create a pre-process result "pA" only based on original input and send it back to get result from "B"
   2. module "B" receives "pA", create its final result "rB" only based on "pA", and send it back
   3. module "A" receives result "rB" and create its final result "rA" only based on "rB" (e.g. DMN -> PMML)

1. possible issues
   1. module "A" create a pre-process result "pA" and send it back to get result from "B"
   2. module "B" receives "pA", create its final result "rB" only based on "pA", and send it back
   3. module "A" receives result "rB" and create its final result "rA" based on both "rB" and original input (e.g. PMML -> DRL)

1. other
   1. some models execute a pre-process on input and send this preprocessed input for redirected evaluation (e.g. PMML -> DRL)
   2. when the runtimemanager receives the redirected-evaluation (e.g from DRL) it must send it back to the exact instance that executed the preprocessing
   3. FRI currently does not have a knowledge of "instance", but only of "model"
   4. that would require
      1. modification of FRI to allow instance addressing (easy to do)
      2. implement some sort of instance status for models that - otherwise - could be happily stateless

pbizzarr@redhat.com ^ ^