Asynchronous, Parallel Sub-Models + Hardware

[ofSingularMind]

Hello,

It might be that I would like to create in a model several sub-models or sub-graphs. I might like these to run asynchronously and in parallel. I might also like to connect these sub-graphs in various fashions, like a standard reactive MP connection, a publisher/subscriber relationship where messages flow only one way, or client/server relationship where there is a call-and-response occuring.

Has there been discussion on this topic?

I'm wondering if integrating with ROS isn't the answer... Are we familiar? I believe it offers all the features I mentioned, and additionally integrates well with hardware (sensors/actuators).

I did consider connecting all parts of the model in the same graph, but I didn't know how well multiple asynchronous inputs could be handled, may need to handle one-way message-passing, and limited to no parallelization support (?). Plus hardware connections are ad-hoc.

So, this issue opens a conversation to identify and implement the best option to enable these features.

[albertpod]

@ofSingularMind thanks for opening the thread. jfyi, issues are for bugs and concrete feature requests with clear requirements. This is exploratory conversation about potential architectural directions, which is exactly what the Discussions section is for. I will transfer this to discussions and we can proceed there.

[albertpod]

@ofSingularMind I think there's some confusion (on my side) about what problem you're trying to solve versus what tools solve that problem.

ROS is middleware, not a modeling framework. It handles inter-process communication, not probabilistic inference or model composition. RxInfer.jl is a software that in principle can sit on top of middleware, e.g., ROS.
What you're describing:

1. Sub-models/sub-graphs → This is a modeling question (how to structure your probabilistic model)
2. Asynchronous parallel execution → This is an inference/computation question (how message passing is scheduled)
3. Different communication patterns → This is already built (to some extent) into factor graphs (sum-product vs. expectation propagation, etc.)
4. Hardware connections → This is an I/O question (separate from the inference engine)

These are orthogonal concerns. ROS doesn't help you structure probabilistic models, and RxInfer doesn't need ROS to handle any of the patterns you mentioned.

Regarding current RxInfer capabilities:
RxInfer doesn't currently provide true parallelism or fully asynchronous execution. Some of the architectural concerns you're raising (distributed inference, parallelization, inter-graph communication) were in mind when developing of Cortex.jl, but that project is currently on hold.

What you should clarify:
Modeling: What probabilistic model are you trying to build? Hierarchical? Multi-agent?
Computation: Do you have a specific performance bottleneck that you think parallelization would solve?
Hardware: What hardware are you interfacing with? (This is just I/O)

If you have a specific modeling problem or computation bottleneck, describe that. "Integrating with ROS" doesn't solve any of the concerns you listed - it would just add a communication layer between things that don't need to communicate across process boundaries.