Status

DECIDED 

StakeholdersMakoto Takemiya
OutcomeCustom minimalistic actor framework will be introduced and integrated
Due date
Owner

Background

To support a various number of use cases Iroha2 must implement logic to plug additional logic as an actor. The design should allow an arbitrary actor to receive messages from other actors and share messages about his work.

Problem

The current runtime design is implemented using mpsc. The single receiver channel architecture turns into a chaos of strict dependencies and creates constraints for extending logic with new modules.

Solution

Rework core architecture with actor model and single message bus. Bus receives all messages during execution of actors and broadcast them to others. Message trait should allow identify nature of message to filter it. To prevent any blocking problems, messages must reference an immutable value or a cloned value.

According to the requirement of supporting different use cases, actors should be implemented as compile-time features (at least) or as pluggable extensions (point to discuss and future research).


Decisions

Custom minimalistic actor framework will be introduced and integrated.

Alternatives

  1. Use available actor frameworks such as Actix or different

Concerns

  1. Tradeoff between extensibility and overhead with maintenance global bus and filter messages

Assumptions

  1. Filtering messages on message bus layer against skipping on actors
  2. Research pros and cons of actors as separate binaries with external API to extend pluggability

Risks

Additional Information

https://github.com/jonhoo/bus
Simple implementation of a lock-free, bounded, single-producer, multi-consumer, broadcast channel

https://github.com/actix/actix
Popular actor framework for Rust

https://video.ethz.ch/events/2017/rust/e2d74a7b-7cfd-4bd3-8ce5-23c99ad775e2.html
Type-safe & high-perf distributed actor systems with Rust by Eickhoff, Anselm


  • No labels

1 Comment

  1. Aler Denisov What are your thoughts about creating an actor framework which would require sending all of messages also to some supervisor? So we could always observe status of everything inside of it.


    Also we can't use actix framework as it has its own runtime build on top of tokio. May be we could take a look at xtra framework, it is only 1k lines of code and could be easily forked and changed. Could you look at it and share your thoughts?


    xtra – https://docs.rs/xtra/0.5.1/xtra/index.html