Crates.io | sails-rs |
lib.rs | sails-rs |
version | 0.6.3 |
source | src |
created_at | 2024-07-18 09:14:38.520378 |
updated_at | 2024-11-07 10:29:56.616088 |
description | Main abstractions for the Sails framework |
homepage | |
repository | https://github.com/gear-tech/sails |
max_upload_size | |
id | 1307194 |
size | 73,959 |
Sails
is a library for bringing your experience of writing applications utilizing
Gear Protocol to the next level of simplicity and
clarity. It deals with things like:
NOTE
The
Sails
library is published under the namesails-rs
oncrates-io
.
Either use Sails
CLI:
cargo install sails-cli
cargo sails new-program my-ping
Or add the following to your Cargo.toml
[dependencies]
sails-rs = { version = "*", features = ["debug"] }
And then in your lib.rs
:
#![no_std]
use sails_rs::{gstd::debug, prelude::*};
struct MyPing;
#[service]
impl MyPing {
pub const fn new() -> Self {
Self
}
pub async fn ping(&mut self) -> bool {
debug!("Ping called");
true
}
}
#[derive(Default)]
struct MyProgram;
#[program]
impl MyProgram {
#[route("ping")]
pub fn ping_svc(&self) -> MyPing {
MyPing::new()
}
}
The entire idea of the Gear Protocol is based on the asynchronous version of the Request-Response Pattern. On-chain applications loaded onto Gear-based network receive and handle messages from the other on-chain or off-chain applications. Both can be treated as external consumers of services provided by your application, and the latter can represent ordinary people interacting with the network.
Sails
architecture for applications is based on a few key concepts.
The first one is service which is represented by an impl of some Rust struct
marked with the #[service]
attribute. The service main responsibility is
implementing some aspect of application business logic.
A set of service's public methods defined by the impl is essentially a set of
remote calls the service exposes to external consumers. Each such method working
over a &mut self
is treated as a command changing some state, whereas each method
working over a &self
is treated as a query keeping everything unchanged and
returning some data. Both types of methods can accept some parameters passed by a
client and can be synchronous or asynchronous. All the other service's methods and
associated functions are treated as implementation details and ignored. The code
generated behind the service by the #[service]
attribute decodes an incoming
request message and dispatches it to the appropriate method based on the method's
name. On the method's completion, its result is encoded and returned as a response
to a caller.
NOTE
In some cases, a command might need to return a certain amount of tokens (value) from the application's balance to the caller's one. This can be done via using a dedicated type,
CommandReply<T>
.
#[service]
impl MyService {
// This is a command
pub fn do_something(&mut self, p1: u32, p2: String) -> &'static [u8] {
...
}
// This is a command returning value along with the result
pub fn withdraw(&mut self, amount: u64) -> CommandReply<()> {
CommandReply::new(()).with_value(amount)
}
// This is a query
pub fn something(&self, p1: Option<bool>) -> String {
...
}
}
The second key concept is program which is similarly to the service represented
by an impl of some Rust struct marked with the #[program]
attribute. The program
main responsibility is hosting one or more services and exposing them to the external
consumers.
A set of its associated public functions returning Self
are treated as application
constructors. These functions can accept some parameters passed by a client and can be
synchronous or asynchronous. One of them will be called once at the very beginning of
the application lifetime, i.e. when the application is loaded onto the network. The
returned program instance will live until the application stays on the network. If there
are no such methods discovered, a default one with the following signature will be generated:
pub fn default() -> Self {
Self
}
A set of program's public methods working over &self
and having no other parameters
are treated as exposed service constructors and are called each time when an incoming
request message needs be dispatched to a selected service. All the other methods and
associated functions are treated as implementation details and ignored. The code
generated behind the program by the #[program]
attribute receives an incoming request
message from the network, decodes it and dispatches it to a matching service for actual
processing. After that, the result is encoded and returned as a response to a caller.
Only one program is allowed per application.
#[program]
impl MyProgram {
// Application constructor
pub fn new() -> Self {
...
}
// Yet another application constructor
pub fn from_u32(p1: u32) -> Self {
...
}
// Service constructor
pub fn ping_svc(&self) -> MyPing {
...
}
}
And the final key concept is message routing. This concept doesn't have a
mandatory representation in code, but can be altered by using the #[route]
attribute applied to those public methods and associated functions described above.
The concept itself is about rules for dispatching an incoming request message to
a specific service's method using service and method names. By default, every
service exposed via program is exposed using the name of the service constructor
method converted into PascalCase. For example:
#[program]
impl MyProgram {
// The `MyPing` service is exposed as `PingSvc`
pub fn ping_svc(&self) -> MyPing {
...
}
}
This behavior can be changed by applying the #[route]
attribute:
#[program]
impl MyProgram {
// The `MyPing` service is exposed as `Ping`
#[route("ping")] // The specified name will be converted into PascalCase
pub fn ping_svc(&self) -> MyPing {
...
}
}
The same rules are applicable to service method names:
#[service]
impl MyPing {
// The `do_ping` method is exposed as `Ping`
#[route("ping")]
pub fn do_ping(&mut self) {
...
}
// The `ping_count` method is exposed as `PingCount`
pub fn ping_count(&self) -> u64 {
...
}
}
Sails
offers a mechanism to emit events from your service while processing commands.
These events serve as a means to notify off-chain subscribers about changes in
the application state. In Sails
, events are configured and emitted on a per-service
basis through the events
argument of the #[service]
attribute. They are defined
by a Rust enum, with each variant representing a separate event and its optional data.
Once a service declares that it emits events, the #[service]
attribute automatically
generates the notify_on
service method. This method can be called by the service
to emit an event. For example:
fn counter_mut() -> &'static mut u32 {
static mut COUNTER: u32 = 0;
unsafe { &mut COUNTER }
}
struct MyCounter;
#[derive(Encode, TypeInfo)]
enum MyCounterEvent {
Incremented(u32),
}
#[service(events = MyCounterEvent)]
impl MyCounter {
pub fn new() -> Self {
Self
}
pub fn increment(&mut self) {
*counter_mut() += 1;
self.notify_on(MyCounterEvent::Incremented(*counter_mut())).unwrap();
}
// This method is generated by the `#[service]` attribute
fn notify_on(&mut self, event: MyCounterEvent) -> Result<()> {
...
}
}
It's important to note that, internally, events use the same mechanism as any other message transmission in the Gear Protocol. This means an event is only published upon the successful completion of the command that emitted it.
A standout feature of Sails
is its capability to extend (or mix in) existing services.
This is facilitated through the use of the extends
argument in the #[service]
attribute. Consider you have Service A
and Service B
, possibly sourced from
external crates, and you aim to integrate their functionalities into a new
Service C
. This integration would result in methods and events from Services A
and B
being seamlessly incorporated into Service C
, as if they were originally
part of it. In such a case, the methods available in Service C
represent a combination
of those from Services A
and B
. Should a method name conflict arise, where both
Services A
and B
contain a method with the same name, the method from the service
specified first in the extends
argument takes precedence. This strategy not only
facilitates the blending of functionalities but also permits the overriding of specific
methods from the original services by defining a method with the same name in the
new service. With event names, conflicts are not allowed. Unfortunately, the IDL
generation process is the earliest when this can be reported as an error. For example:
struct MyServiceA;
#[service]
impl MyServiceA {
pub fn do_a(&mut self) {
...
}
}
struct MyServiceB;
#[service]
impl MyServiceB {
pub fn do_b(&mut self) {
...
}
}
struct MyServiceC;
#[service(extends = [MyServiceA, MyServiceB])]
impl MyServiceC {
// New method
pub fn do_c(&mut self) {
...
}
// Overridden method from MyServiceA
pub fn do_a(&mut self) {
...
}
// do_b from MyServiceB will exposed due to the extends argument
}
An application written with Sails
uses SCALE Codec to encode/decode data
at its base.
Every incoming request message is expected to have the following format:
| SCALE encoded service name | SCALE encoded method name | SCALE encoded parameters |
Every outgoing response message has the following format:
| SCALE encoded service name | SCALE encoded method name | SCALE encoded result |
Every outgoing event message has the following format:
| SCALE encoded service name | SCALE encoded event name | SCALE encoded event data |
Having robust interaction capabilities with applications is crucial. Sails
offers
several options for interaction.
Firstly, it supports manual interaction using the Gear Protocol. You can use:
msg::send
functions from the gstd
crate to interact between applications.gclient
crate to interact from off-chain code with an on-chain application.@gear-js/api
library to interact with your program from JavaScript.All you need to do is compose a byte payload according to the layout outlined in the Payload Encoding section and send it to the application.
Thanks to the generated IDL, Sails
provides a way to interact with your application
using generated clients with an interface similar to the one exposed by latter in
a clearer way. Currently, Sails
can generate client code for Rust and TypeScript.
When it comes to Rust, there are two options:
For TypeScript see generated clients documentation.
Say you have an application that exposes a service MyService
with a command do_something
:
struct Output {
m1: u32,
m2: String,
}
#[service]
impl MyService {
pub fn do_something(&mut self, p1: u32, p2: String) -> Output {
...
}
}
#[program]
impl MyProgram {
pub fn my_service(&self) -> MyService {
MyService::new()
}
}
Then, in a client application, provided the code generation happens in a Rust build script, you can use the generated code like this (option 1):
include!(concat!(env!("OUT_DIR"), "/my_service.rs"));
fn some_client_code() {
let call_payload = my_service::io::DoSomething::encode_call(42, "Hello".to_string());
let reply_bytes = gstd::msg::send_bytes_for_reply(target_app_id, call_payload, 0, 0).await.unwrap();
let reply = my_service::io::DoSomething::decode_reply(&reply_bytes).unwrap();
let m1 = reply.m1;
let m2 = reply.m2;
}
Or like this (option 2):
include!(concat!(env!("OUT_DIR"), "/my_service.rs"));
fn some_client_code() {
let mut my_service = MyService::new(remoting); // remoting is an abstraction provided by Sails
let reply_ticket = client.do_something(42, "Hello".to_string())
.with_reply_deposit(42)
.publish(target_app_id)
.await.unwrap();
let reply = reply_ticket.reply().await.unwrap();
let m1 = reply.m1;
let m2 = reply.m2;
}
The second option provides you with an option to have your code testable, as the generated code depends on the trait which can be easily mocked.
As you may have noticed, the option 2 uses the concept of a remoting
object, which needs
to be passed to the client instantiation code. This object should implement the Remoting
trait from the sails-rs
crate. It abstracts the low-level communication details
between client and the application. The sails-rs
crate provides three implementations of this
trait:
sails_rs::gstd::calls::GStdRemoting
should be used when the client code is executed
as a part of another on-chain application.sails_rs::gclient::calls::GClientRemoting
should be used when the client code is executed
as a part of an off-chain application.sails_rs::gstd::calls::GTestRemoting
should be used when the client code is executed
as a part of a tests utilizing the gtest
crate.When it comes to TypeScript, sails-js
library can be used to interact with the program. Check out sails-js
documentation for more details.
Occasionally, you may need to design a system where a business transaction spans multiple
applications. In other words, a single business transaction may involve several local
transactions within different applications. This challenge is typically addressed using
a pattern called Saga
. You can find detailed documentation about this pattern here
and implementation guidelines here.
The process of addressing this issue in applications built with Sails
is similar, but only
the orchestration approach can be used, as Sails
applications cannot catch events from each
other. Additionally, it is important to handle infrastructure errors that can arise,
particularly those related to the Gear Protocol and its concept of gas.
Let's refer to the guidelines and explore the nuances involved.
First, it is important to note that all the errors mentioned in the guidelines are business errors — only business errors can trigger compensation actions. In contrast, infrastructure errors are expected to be resolved through retries. This applies to both retriable and compensatable transactions. The key difference is that the former should never return a business error.
Normally, these retries are handled programmatically. However, when a RunOutOfGas
error is
received from another application, retrying is not an option. The only solution in this case
is to propagate the error to the top-level caller, who will then need to attach more funds
and attempt the entire business transaction again.
For retries to function correctly, local transactions must be idempotent (i.e., safe to retry
without causing side effects, such as sending duplicate events). This can be ensured by
assigning a unique identifier to each business transaction, which is passed to all actions
within the Saga
, whether it’s the initial attempt or a retry.
The Saga’s state must track this identifier to skip actions that were completed in previous attempts. Similarly, the actions themselves should recognize this identifier to prevent repeating changes that have already been made.
Another issue to consider is the possibility of receiving a Timeout
error while waiting for
a response from another application. In some systems, timeouts can be treated as business errors,
triggering compensation. However, if the timeout is caused by infrastructure issues (e.g.,
network congestion), it can be handled similarly to the RunOutOfGas
error.
There are optimization opportunities in such cases. Since the Gear Protocol guarantees that the caller will eventually receive a response (successful or not), one simple optimization is to increase the number of blocks allowed for waiting on a response. On the Vara network, this value is set to 100 blocks by default. Increasing it to 10,000 blocks would mitigate most network load issues without significantly raising transaction costs. This adjustment can be made using the with_wait_up_to method.
Another option is to use the with_reply_hook method, which involves
additional logic to manage the Saga
’s state. The reply hook can be triggered while the main
code handles the Timeout
error, allowing the caller to initiate another attempt. During this
process, the timed-out action can be marked as completed in the Saga
’s state, preventing it
from being re-executed. However, it's worth considering whether this added complexity is
necessary, as retrying an already completed idempotent action is harmless and only slightly
increases the overall cost.
To summarize:
Saga
(orchestrator application) by maintaining its state.Saga
’s state, along
with the status of each action (e.g., not executed, succeeded, failed).RunOutOfGas
and Timeout
. The simplest approach
is to propagate these errors to the top-level caller for retries.
Timeout
errors, optimize by increasing the number of blocks allowed for waiting on a response.You can find all examples here along with some descriptions provided at the folder level. You can also find some explanatory comments in the code. Here is a brief overview of features mentioned above and showcased by the examples:
The examples are composed on a principle of a few programs exposing several services.
See DemoProgram which demonstrates this, including
the use of program's multiple constructors and the #[route]
attribute for one of
the exposed services. The example also includes Rust build script
building the program as a WASM app ready for loading onto Gear network.
There are a couple of services which demonstrate basic service structure exposing some primitive methods operating based on input parameters and returning some results. They serve as an excellent starting point for developing your services. See Ping and ThisThat services. The latter, in addition to the basics, showcases the variety of types which can be used as parameters and return values in service methods.
In the real world, almost all apps work with some form of data, and apps developed
using Sails
are no exception. As discussed in the Application section,
services are instantiated for every incoming request message, indicating that these
services are stateless. However, there are a few ways to enable your services to
maintain some state. In this case, the state will be treated as external to the service.
The most recommended way is demonstrated in the Counter
service, where the data is stored as part of the program and passed to the service
via RefCell
. The service module merely defines the shape of the data but requires
the data itself to be passed from the outside. This option provides you with full
flexibility and allows you to unit test your services in a multi-threaded environment,
ensuring the tests do not affect each other.
Another method is illustrated in the RmrkCatalog
and RmrkResource services, where the data
is stored in static variables within the service module. This strategy ensures that the
state remains completely hidden from the outside, making the service entirely self-contained.
However, this approach is not ideal for unit testing in a multi-threaded environment
because each test can potentially influence others. Additionally, it's important
not to overlook calling the service's seed
method before its first use.
You can also explore other approaches, such as making a service require &'a mut
for its
data (which makes the service non-clonable), or using Cell
(which requires data copying,
incurring additional costs).
In all scenarios, except when using Cell
, it's crucial to consider the static nature
of data, especially during asynchronous calls within service methods. This implies that
data accessed before initiating an asynchronous call might change by the time the call
completes. See the RmrkResource service's
add_part_to_resource
method for more details.
You can find an example of how to emit events from your service in the Counter and RmrkResource services.
An example of service extension is demonstrated with the Dog
service, which extends the Mammal service from
the same crate and the Walker service from a different crate.
The service being extended must implement the Clone
trait, while the extending
service must implement the AsRef
trait for the service being extended.
The Demo Client crate showcases how to generate client code from an IDL file as a separate Rust crate. Alternatively, you can use the same approach directly in your application crate. See Rmrk Resource.
You can find various examples of how to interact with the application using the generated client code in Demo Tests. Check the comments in the code for more details.
Since the generated code is the same for all environments, whether it is an interaction
from tests or from another application, the techniques for these interactions are the same.
You can find an example of the interaction from an application in the
Rmrk Resource service's add_part_to_resource
method.
Bear in mind that working with the generated client requires the sails_rs
crate to
be in dependencies.