Worker to Worker communication for Rust components
See the Worker to Worker communication page for a general overview of how workers can invoke each other in Golem.
Setting up
For doing worker to worker communication between Golem components implemented in Rust, in addition to the Rust compiler and the cargo-component
tool described in the setup page, we need two additional dependencies:
- The
golem-cli
tool which automates some steps of setting up the communication between workers. - The
cargo-make
crate task runner used to further automate the build steps.
Install cargo-make
by running:
$ cargo install cargo-make
Creating a project with two components
A project building two Golem components where one can call the other can be set up with the following primary steps:
Create the two components
First create two Rust components using the techniques described in the defining components page. The two components should be in two separate directories, enclosed by a root directory for the whole project.
It is important to use a different package name for each component, otherwise the stub generator will fail.
This can be done by using the --package-name
parameter of golem-cli new
for example --package-name rpcdemo:component1
and --package-name rpcdemo:component2
.
Create a cargo workspace
Make the root directory a cargo workspace. This is done by creating a Cargo.toml
file in the root directory with the following content:
[workspace]
resolver = "2"
# list the two components here
members = [
"component1",
"component2"
]
# common release profile for all components
[profile.release]
opt-level = "s"
lto = true
strip = true
To learn more about cargo workspaces, check the official documentation (opens in a new tab).
Check it works by running cargo component build
in the root directory of the project, which should build both components in parallel:
$ cargo component build
...
Creating component target/wasm32-wasi/debug/component2.wasm
Creating component target/wasm32-wasi/debug/component1.wasm
Set up the Golem WASM-RPC stub generator
In the next step we use golem-cli
to modify our cargo workspace and add a cargo-make
task description file to automate the build steps.
This can be done by running:
$ golem-cli stubgen initialize-workspace --targets component1 --callers component2
...
Writing updated Cargo.toml to "/Users/vigoo/tmp/doc-temp/rpc/Cargo.toml"
Done
There can be multiple --targets
and --callers
parameters. Every component listed as a caller
will be able to call every component listed as a target.
Only add links between components where there really is a need for RPC calls, because every link increases the size of the generated component, and also the build fails on unused links.
Test the setup by running cargo make build-flow
in the root directory of the project:
$ cargo make build-flow
...
Error: no dependencies of component `target/wasm32-wasi/debug/component2.wasm` were found
[cargo-make] ERROR - Error while executing command, exit code: 1
[cargo-make] WARN - Build Failed.
This error is expected, as we haven't added any remote procedure calls yet, but we have a new component in our workspace, called component1-stub
and it has successfully compiled.
Import the generated stub in the caller component
The next step is to import the generated stub component's interface into the caller component (component2
in the above example) WIT
world.
This is done by modifying component2/wit/component2.wit
in the following way:
// ...
world component2 {
import demo:component1-stub/stub-component1; // new line added
export api;
}
The exact name to be imported depends on what package and component names were used when creating the target component.
In this example the package name was demo:component1
and the component name was component1
. This gives the generated stub the component name demo:component1-stub
and the generated stub interface within it is named stub-component1
(here component1
coming from the target component's name).
Start implementing the components
With this the workspace is set up, and the generated stubs can be used from the caller component to call the target component, as it is described in the next section.
Once the code is written, use:
cargo make build-flow
to build all the components quickly in debugcargo make release-build-flow
to create a release build of all the components
Writing blocking remote calls
For each exported function in the target component, the generated stub contains two exported variants:
- A blocking variant, prefixed with
blocking-
, which does not return until the remote worker finished processing the call. - A non-blocking variant, which returns a pollable result immediately (unless for remote functions without any return value, in which case it just triggers the invocation but does not return anything)
To use the blocking variants we need to construct a resource from the generated stub, passing the remote Golem URI of the target worker to the constructor.
Taking the rust-minimal-actor
example as the target component, which has the following exported interface:
package demo:component1;
interface api {
add: func(value: u64);
get: func() -> u64;
}
world component1 {
export api;
}
The generated stub will contain the following functions:
package demo:component1-stub;
interface stub-component1 {
use golem:rpc/types@0.1.0.{uri};
use wasi:io/poll@0.2.0.{pollable};
resource future-get-result {
subscribe: func() -> pollable;
get: func() -> option<u64>;
}
resource api {
constructor(location: uri);
blocking-add: func(value: u64);
add: func(value: u64);
blocking-get: func() -> u64;
get: func() -> future-get-result;
}
}
world wasm-rpc-stub-component1 {
export stub-component1;
}
To use this generated api
resource, the following steps are needed in the caller component (component2
in the example):
Import the generated types
The generated bindings for the generated stubs need to be imported in the module where the remote procedure calls are made.
use crate::bindings::golem::rpc::types::Uri;
use crate::bindings::demo::component1_stub::stub_component1::*;
Construct the URI for the target worker
The worker to be invoked is identified by an URI which consists of the component ID and the worker name.
Workers can get their own URI by using the get_self_uri
function from the Golem Rust SDK, but it can also be constructed manually.
If the worker pointed by the URI does not exists, it is going to be created automatically and it inherits the environment variables of the caller.
A common pattern is to use environment variables to configure the deployed component's component ID and use that to construct the URI. See the defining components page for more information about passing configuration values to workers.
The following code snippet gets the target component's id from an environment variable, and constructs the URI for the target worker with a fixed worker name target-worker
:
let component_id =
std::env::var("TARGET_COMPONENT_ID").expect("TARGET_COMPONENT_ID not set");
let target_uri = Uri {
value: format!("worker://{component_id}/target-worker"),
};
Construct and use the API resource
With target_uri
the Api
resource can be constructed and used to call the remote functions.
let remote_api = Api::new(&target_uri);
remote_api.blocking_get()
Writing non-blocking remote calls
Using the non-blocking variants of the generated stub functions requires the same steps as the blocking variants, but the returned value is a special future which needs to be subscribed to and polled using the low-level WASI poll
interface.
The following snippet demonstrates how to get the value of three counters simultaneously, assuming that counter1
, counter2
and counter3
are all insteances of the generated stub's Api
resource:
// Making the non-blocking calls
let future_value1 = counter1.get_value();
let future_value2 = counter2.get_value();
let future_value3 = counter3.get_value();
let futures = &[&future_value1, &future_value2, &future_value3];
// Subscribing to get the results
let poll_value1 = future_value1.subscribe();
let poll_value2 = future_value2.subscribe();
let poll_value3 = future_value3.subscribe();
let mut values = [0u64; 3];
let mut remaining = vec![&poll_value1, &poll_value2, &poll_value3];
let mut mapping = vec![0, 1, 2];
// Repeatedly poll the futures until all of them are ready
while !remaining.is_empty() {
let poll_result = bindings::wasi::io::poll::poll(&remaining);
// poll_result is a list of indexes of the futures that are ready
for idx in &poll_result {
let counter_idx = mapping[*idx as usize];
let future = futures[counter_idx];
let value = future
.get()
.expect("future did not return a value because after marked as completed");
values[counter_idx] = value;
}
// Removing the completed futures from the list
remaining = remaining
.into_iter()
.enumerate()
.filter_map(|(idx, item)| {
if poll_result.contains(&(idx as u32)) {
None
} else {
Some(item)
}
})
.collect();
// Updating the index mapping
mapping = mapping
.into_iter()
.enumerate()
.filter_map(|(idx, item)| {
if poll_result.contains(&(idx as u32)) {
None
} else {
Some(item)
}
})
.collect();
}
// values contains the results of the three calls
Deploying the resulting components
To build deployable WASM files of the involved components, use
$ cargo make release-build-flow
...
Creating component target/wasm32-wasi/release/component1.wasm
Creating component target/wasm32-wasi/release/component2.wasm
Creating component target/wasm32-wasi/release/component1_stub.wasm
...
Writing composed component to "target/wasm32-wasi/release/component2_composed.wasm"
Done
In this example, where component1
was the target, and component2
was the caller, the following artifacts are to be deployed:
target/wasm32-wasi/release/component1.wasm
is the target component which itself does not perform any remote procedure callstarget/wasm32-wasi/release/component2_composed.wasm
is the version ofcomponent2
composed with the RPC stubs, ready to be deployed to Golem
If you try to upload the non-composed build output, component2.wasm
to Golem, it will fail to
create the worker because not all the imports of the WebAssembly component are satisfied.
Adding new connections in a workspace
When during development new components are added to the workspace, and/or new connections need to be estabilished between them, running the golem-cli stubgen initialize-workspace
command again will regenerate the build steps in the cargo makefile. Make sure to list all the targets and callers, not just the new ones.
There is currently no automated way to remove connections from the workspace. If this is needed, the generated stub subprojects and their WIT definitions from the caller component's wit/deps
must be removed manually.