A minimalist AWS Lambda runtime for Rust.
fn main() -> ! {
minlambda::run_ok(|_: serde::de::IgnoredAny| "Hello, world!")
}
minlambda implements the AWS Lambda runtime interface, deserializing events and serializing responses with Serde JSON.
To communicate with the runtime API over HTTP, minlambda uses a purpose-built HTTP client.
minlambda doesn't parse response headers in the invocation event (other than the request ID). This includes the function deadline, function ARN, AWS X-Ray tracing header, or additional AWS Mobile SDK data. The crate author has never needed these and, well, this is a minimal runtime.
minlambda doesn't run your handler in an async runtime. If you're using async code, you can
create a runtime outside of lambda::run
and call its blocking function (e.g. Tokio's
Runtime::block_on
). An example for Tokio is available.
Probably most of the time.
If you're using Lambda to interact with other AWS services, which is very likely, you are probably using an SDK (such as Rusoto) that probably relies on hyper and Tokio, and you're not really reducing your total dependency closure compared to the AWS Labs runtime.
The HTTP client was built to work with Lambda's runtime API, and not to be a generic RFC-compliant HTTP client; if Lambda's underlying interface subtly changes, this runtime could break unexpectedly. (This probably won't happen: we believe that the subset of the HTTP spec we implement is by the book.)
You like simple things, or your code already has minimal dependencies.
Some lovely examples are available in our repository.
Building binaries that actually work in the Lambda execution environment is a bit of an art, as it contains stable (old) versions of glibc and the like. Your compiler is probably targeting a system with newer shared libraries and symbol versions than what the execution environment has available, resulting in cryptic dynamic linker errors at runtime.
If you find the musl libc toolchain reasonable to work with, building a fully static binary is probably the way to go. If you find containers reasonable to work with, using softprops/lambda-rust is probably the way to go.
The author of this crate works at AWS, but this is not an official AWS project, nor does it necessarily represent opinions of or recommended best-practices on AWS.