[RFC] Wrap interface for `tower::Service<Request>`

[nmoutschen]

Remark: This is a follow-up to #374 as I'm currently implementing it. See #401 for progress.

Background

Right now, the current implementation in lambda_http::handler takes a handler function that takes a Request and returns an IntoResponse implementation. This allows to pass the wrapped handler function into lambda_runtime::run directly. However, it has a few limitations.

First, Request and IntoResponse aren't Deserialize and Serialize, but lambda_runtime::run require those traits. It means we have to pass through an intermediate Adapter to transform the requests and responses. This is currently handled by lambda_http::handler. As a consequence, the current implementation doesn't allow taking a lambda_runtime::Handler implementation directly. It offers a lambda_http::Handler, which implements it.

This causes some switching away from custom Handler traits towards tower::Service. Ideally, we should allow users to provide tower::Service<Request> implementations, and handle the transformation under the hood on their behalf. However, due to lambda_runtime::run's signature, we cannot do that as-is. We also cannot implement Deserialize and Serialize for Request and Response, as those are all foreign traits and types.

Proposition 1: extension trait for Service<Request>

use tower::Service;
use lambda_http::{Request, ServiceExt};

#[derive(Default)]
struct MyHandler;

impl Service<Request> for MyHandler {
    // skipped
}

#[tokio::main]
async fn main() -> Result<(), Error> {
    // Use .into_runtime() to transform the Service
    lambda_runtime::run(MyHandler::default().into_runtime()).await
}

Proposition 2: create a lambda_http::run wrapper

use tower::Service;
use lambda_http::Request;

#[derive(Default)]
struct MyHandler;

impl Service<Request> for MyHandler {
    // skipped
}

#[tokio::main]
async fn main() -> Result<(), Error> {
    // Use lambda_http::run to wrap the Service
    lambda_http::run(MyHandler::default()).await
}

Proposition 3: your idea here

If you have an idea that gives a cleaner interface, let me know!

[brainstorm]

I do find Proposition 2 more clean than prop 1, as an end user that is.

[rimutaka]

What are the limitations of Prop 1 vs Prop 2? If the only difference is having .into_runtime() then Prop 2 seems nicer.

[calavera]

Option 2 looks more clear to me. I link that I don't have to remember a different crate name to run the service, it's all in lambda_http.

[greenwoodcm]

I agree that Prop 2 seems cleaner as well

[nmoutschen]

Looks like everyone's on agreement on this, and this will simplify the external interface quite a lot. 😄 Thanks, folks!

[LucioFranco]

I think prop 2 looks good. I would probably use service_fn over a manual service implementation. In addition, this supports people to use things like warp or axum which is nice. I think building around tower is a great initiative (I think ive wanted this in here for a while).

cc @davidpdrsn may be interested

[davidpdrsn]

I think building around tower is a great initiative (I think ive wanted this in here for a while).

I agree! This looks great 😃

[coltonweaver]

+1 to seeing if we can use service_fn, but overall Prop 2 is great. Thanks for putting this together @nmoutschen!

[nmoutschen]

We'll be able to support service_fn here, for lambda_http, lambda_runtime, and lambda_extension.

For lambda_runtime, the requirement will be to pass a lambda_runtime::LambdaEvent<T>. I've kept the handler_fn so people can still provide a function with 2 arguments (one for the event, one for the context).

For lambda_http, I'd like to use ServiceExt to hold the context, so people could pass a tower::Service<http::Request>, and thus use service_fn with a function that takes a Request.

[nmoutschen]

Closing the ticket as this is now implemented in 84d41c9