37 releases (7 breaking)
0.13.0 | Oct 25, 2024 |
---|---|
0.11.2 | Sep 25, 2024 |
0.10.3 | May 24, 2024 |
0.9.7 | Mar 16, 2024 |
0.7.5 | Mar 24, 2023 |
#472 in Asynchronous
528 downloads per month
760KB
16K
SLoC
ractor_cluster
A companion crate to ractor
for supporting remote actors
This crate contains extensions to ractor
, a pure-Rust actor framework. Inspired from Erlang's gen_server
.
Website Ractor has a companion website for more detailed getting-started guides along with some best practices and is updated regularly. Api docs will still be available at docs.rs however this will be a supplimentary site for ractor
. Try it out! https://slawlor.github.io/ractor/
About
ractor_cluster
expands upon ractor
actors to support transmission over a network link and synchronization of actors on networked clusters of actors.
Installation
Install ractor_cluster
by adding the following to your Cargo.toml dependencies
[dependencies]
ractor = { version = "0.13", features = ["cluster"] }
ractor_cluster = "0.13"
Ractor in distribucted clusters
Ractor actors can be built in a network-distributed pool of actors, similar to Erlang's EPMD which manages inter-node connections + node naming. In our implementation, we have ractor_cluster
in order to facilitate distributed ractor
actors.
ractor_cluster
has a single main type in it, namely the NodeServer
which represents a host of a node()
process. It additionally has some macros and a procedural macros to facilitate developer efficiency when building distributed actors. The NodeServer
is responsible for:
- Managing all incoming and outgoing
NodeSession
actors which represent a remote node connected to this host. - Managing the
TcpListener
which hosts the server socket to accept incoming session requests (with or without encryption).
The bulk of the logic for node interconnections however is held in the NodeSession
which manages
- The underlying TCP connection managing reading and writing to the stream.
- The authentication between this node and the connection to the peer
- Managing actor lifecycle for actors spawned on the remote system.
- Transmitting all inter-actor messages between nodes.
- Managing PG group synchronization
etc..
The NodeSession
makes local actors available on a remote system by spawning RemoteActor
s which are essentially untyped actors that only handle serialized messages, leaving message deserialization up to the originating system. It also keeps track of pending RPC requests, to match request to response upon reply. There are special extension points in ractor
which are added to specifically support RemoteActor
s that aren't generally meant to be used outside of the standard
Actor::spawn(Some("name".to_string()), MyActor).await
pattern.
Quick-start
The basics of setting up a networked cluster of actors lives in the NodeServer
struct. This structure handles the low-level network ownership over a server port along with all of the lifecycle of cluster inter-connections. By spawning this single struct, you're able to accept incoming connections between hosts!
Nodes in the network are authenticated to each other with a "magic cookie" following the Erlang specification in Erlang's distribution protocol. If you want to connect to another host, you need to
- Initialize your own
NodeServer
- Execute a "client-connection" to the remote
NodeServer
you're trying to connect to like
let host = "1.2.3.4";
let port = "4697";
ractor_cluster::client_connect(
&actor,
format!("{host}:{port}"),
)
Similarly there is a client_connect_enc
to connect to a NodeServer
which is utilizing encrypted communication. That's it! If your nodes are sharing a proper magic cookie value, they should authenticate to each other and you'll see remote actors spawned on your local system which you can communciate with through the various pg
or pid
-based registries.
Designing remote-supported actors
Note not all actors are created equal. Actors need to support having their message types sent over the network link. This is done by overriding specific methods of the ractor::Message
trait all messages need to support. Due to the lack of specialization support in Rust, if you choose to use ractor_cluster
you'll need to derive the ractor::Message
trait for all message types in your crate. However to support this, we have a few procedural macros to make this a more painless process
Deriving the basic Message trait for in-process only actors
Many actors are going to be local-only and have no need sending messages over the network link. This is the most basic scenario and in this case the default ractor::Message
trait implementation is fine. You can derive it quickly with:
use ractor_cluster::RactorMessage;
use ractor::RpcReplyPort;
#[derive(RactorMessage)]
enum MyBasicMessageType {
Cast1(String, u64),
Call1(u8, i64, RpcReplyPort<Vec<String>>),
}
The will implement the default ractor::Message
trait for you without you having to write it out by hand.
Deriving the network serializable message trait for remote actors
If you want your actor to support remoting, then you should use a different derive statement, namely:
use ractor_cluster::RactorClusterMessage;
use ractor::RpcReplyPort;
#[derive(RactorClusterMessage)]
enum MyBasicMessageType {
Cast1(String, u64),
#[rpc]
Call1(u8, i64, RpcReplyPort<Vec<String>>),
}
which adds a significant amount of underlying boilerplate (take a look yourself with cargo expand
!) for the implementation. But the short answer is, each enum variant needs to serialize to a byte array of arguments, a variant name, and if it's an RPC give a port that receives a byte array and de-serialize the reply back. Each of the types inside of either the arguments or reply type need to implement the ractor_cluster::BytesConvertable
trait which just says this value can be written to a byte array and decoded from a byte array. If you're using prost
for your message type definitions (protobuf), we have a macro to auto-implement this for your types.
ractor_cluster::derive_serialization_for_prost_type! {MyProtobufType}
Besides that, just write your actor as you would. The actor itself will live where you define it and will be capable of receiving messages sent over the network link from other clusters!
Contributors
The original author of ractor
is Sean Lawlor (@slawlor). To learn more about contributing to ractor
please see CONTRIBUTING.md
License
This project is licensed under MIT.
Dependencies
~13–28MB
~519K SLoC