Support for `runtime: provided` without requiring `useDocker`

[cnuss]

Description

This PR allows usage of runtime: provided in serverless.yml and useDocker is unset or false in custom.serverless-offline:

GET /dev/hello (λ: hello)
× Unsupported runtime
× Uncaught exception

It now uses the execa library to do a local execution of the bootstrap script:

GET /dev/hello (λ: hello)

(λ: hello) RequestId: b440ae80-b0ca-4d0d-ab6f-ea3226cb9d1a  Duration: 55.80 ms  Billed Duration: 56 ms

Motivation and Context

• I have a project where I compile a Go application into a binary.
• The compiled binary is so simple that it doesn't need Docker as an abstraction layer for execution
• This allows for direct execution of the bootstrap script without launching it inside a docker container
• This makes for a faster local execution of binaries
• This allows people to connect to debug ports for their compiled binaries (such as launching with delve)

How this works:

• When Lambdas run, the end up starting an embedded server called the Runtime server
• Each invocation gets its own short-lived Runtime Server on a random port
• It has a couple basic endpoints:
  • https://docs.aws.amazon.com/lambda/latest/dg/runtimes-api.html#runtimes-api-next
  • https://docs.aws.amazon.com/lambda/latest/dg/runtimes-api.html#runtimes-api-response
• Each invocation when runtime: provided and useDocker: false|undefined, will create a RuntimeServer on a random port, and set the AWS_LAMBDA_RUNTIME_API environment variable so that the bootstrap script can interact with it.
  • Then as soon as /response is called, the event is cleared from the RuntimeServer so subsequent invocations of /next indefinitely block.

Notes

• Design choice: one RuntimeServer per invocation:
  • We could, possibly, create a single runtime server and share it between multiple invocations. I thought this to be a premature optimization, so I opted to make a brand new RuntimeServer for each invocation, for a more lambda-like experience.
  • A single RuntimeServer could be shared by setting the AWS_LAMBDA_RUNTIME_API environment variable to be: http://localhost:${some-non-random-port}/function-name instead of http://localhost:${some-random-port}
  • I'd love your feedback on if you want this or not...
• This doesn't support Lambda Layers, yet. Although I presume it could
  • For now it will throw an error if useDocker: false or undefined and there are layers on the function

How Has This Been Tested?

• I copied ./tests/integration/docker/provided to ./tests/lambda-run-mode/provided
  • I removed useDocker: true and began development until it worked

[cnuss]

Hi @dherault and @DorianMazur this PR might stir up some debate, so let me know what you think!

Long story short, I never liked the requirement to have Docker to run serverless applications that have a compiled binary. It also caused long-ish first-runs of a lambda function as it downloads and runs the base container and layers.

I found myself wishing Docker wasn't involved in the execution of the bootstrap script, so I decided to make this PR.

[cnuss]

hi @dherault and @DorianMazur it's been about a week, have you been able to take a look at this?

[DorianMazur]

Thanks for the PR @cnuss, but I'd stick with docker in this case. Let's wait for @dherault's response.