Composer: Add API virtual package and require an implementation

[joshdifabio]

Resolves #129.

There's a potential issue here: Travis is now going to require a compatible event-loop-implementation in order to run the tests for event-loop. To solve this, perhaps we could create an event-loop-dummy-implementation repo in async-interop, not add it to Packagist so it's not pulled in elsewhere, move tests\DummyDriver.php to that new repo, and add the following to the event-loop composer.json file:

{
    "repositories": [ { "type": "git", "url": "https://github.com/async-interop/event-loop-dummy-implementation.git" } ]
}

Thoughts?

[kelunik]

We could also move async-interop/event-loop-implementation into suggest.

[joshdifabio]

We could also move async-interop/event-loop-implementation into suggest.

We'd then have the issue of people having to specify dual requirements (api and implementation) in their libs.

[kelunik]

I don't think people really need to specify the event-loop-implementation requirement.

[joshdifabio]

I don't think people really need to specify the event-loop-implementation requirement.

Personally, I agree. I think the implementation requirement will only be needed in certain packages. I think if we're all happy with that then it would make sense to move event-loop-implementation to suggest.

@bwoebi @trowski

[kelunik]

will only be needed in certain packages

Which ones?

[joshdifabio]

Which ones?

Ones like Guzzle which might use a loop but hide that fact from applications.

[kelunik]

We don't care about such packages here. Every package that hides the loop isn't interoperable.

[joshdifabio]

When I say hide the loop, I mean by implementing functionality which implicitly runs the loop and awaits promise resolution instead of requiring clients to bootstrap and run the loop themselves, such as how Guzzle does with its awaitable promises. This is useful in synchronous applications.

[joshdifabio]

Anyway, I'm happy to make this a suggest if everyone else is.

[kelunik]

@joshdifabio I know what you mean and IMO this is not implementable in a interop way, because it blocks then.

[trowski]

Moving event-loop-implementation to suggest is an acceptable solution IMO. I would hope most people would quickly figure out their problem if the suggestion popped up.

[joshdifabio]

Okay, I'm gonna move it to suggest.

[kelunik]

Suggestion: "A loop implementation is required in order to run a loop."

[joshdifabio]

I should squash commits before this gets merged.

[kelunik]

@joshdifabio What's still missing is a note in the README about libraries only depending on event-loop-api and only driver implementations depending on event-loop directly.

[joshdifabio]

What's still missing is a note in the README about libraries only depending on event-loop-api and only driver implementations depending on event-loop directly.

Right. Is someone working on a proper readme & metadoc?

[kelunik]

@bwoebi Wanted to do that, don't know whether he started yet. But in the meantime, a note in the README is fine I guess.

[kelunik]

I just noticed that the suggest doesn't make any sense if packages only depend on the event-loop-api, because event-loop will only be pulled as dependency of the driver implementation.

[kelunik]

While we might want to add a new method somewhere in the future, I think the proposed ways are too complicated and aren't worth the trouble, we should just increase the major version in case we need to add a method. I think it's unlikely to happen after 1.0 is out.

[kelunik]

I think we should just go with #151 and then we don't need anything like this.

[joshdifabio]

I think we should just go with #151 and then we don't need anything like this.

As long as we allow others to define their own drivers #151 doesn't actually change anything. We would still need to provide some kind of semantic version guarantee to both loop users and loop implementors. The question is still whether we would want to bump the major version for loop users when we change the driver in a way which only breaks implementations.

I don't think this PR represents any real complexity, but it gives loop users the power to opt-in to non-breaking API changes to the driver interface. I think that's worth the single provide in composer.json.

[kelunik]

I don't think this PR represents any real complexity, but it gives loop users the power to opt-in to non-breaking API changes to the driver interface. I think that's worth the single provide in composer.json.

What's the proposal for libraries to depend on after #151 is merged?

[joshdifabio]

What's the proposal for libraries to depend on after #151 is merged?

If I understand the question, I see three options:

1. We have no virtual packages and only maintain a single version number. The version number covers the SPI and so any changes to Driver would be a major version bump for Loop clients as well as Driver implementors.
2. We have a virtual package async-interop/event-loop-api which covers the client API, while the main version number covers the SPI. Libraries which don't implement or extend Driver would depend on async-interop/event-loop-api. Libraries which implement or extend Driver would depend on async-interop/event-loop.
3. We have a virtual package async-interop/event-loop-spi which covers the SPI, while the main version number covers the client API only. Libraries which don't implement or extend Driver would depend on async-interop/event-loop. Libraries which implement or extend Driver would depend on async-interop/event-loop-spi.

I think 3 would be my preferred approach.

[kelunik]

Closing, as project discontinued for now.