[2.0.0 REVIEW] Thoughts on a selection expression for models

[jasonmac01]

In regards to the oneOf support for a message within the Operation object:

Having the concept of "multiple message shapes can be received over this channel operation" is very useful - Perfect!

However, what indicates to the recipient which message should be chosen at runtime?

Within AWS API Gateway, we support a concept of a modelSelectionExpression on our Routes - I think this roughly aligns with your Runtime Expression concept at a high level. Do you think this sort of runtime expression should be added within the Operation?

-Jason
AWS API Gateway

[fmvilas]

The meaning of oneOf is that your message should match one (and only one) of those at any time, meaning that there's no way to understand which one is going to be matching. At runtime you should evaluate your message against all of them to 1) Make sure it only matches against one, and 2) understand which one is "the matching one". Does it make sense?

[jasonmac01]

Sure, I suppose iterating through the collection is a possibility, but perhaps a bit inefficient. It may actually be error-prone as well when shapes have overlapping attribute names.

What would be your honest opinion on having an expression that allowed the actor (message broker, producer, or consumer) to extract a "key" that was then used to lookup the appropriate model?

[fmvilas]

Sure I would like to have something like this but don't forget the constraint that it should only validate against one of them. How are you going to guarantee that constraint if you don't evaluate all of them? And it's an honest question, I don't have an answer for this.

[jasonmac01]

The selection expression is what solves this guarantee problem.

As it currently stands, just having oneOf seems ambiguous. What if the payload could in fact be serialized to multiple models? Do you just pick the first, or maybe pick the one with the most matching properties? Either way, that's expensive to accomplish and difficult to predict.

With a selection expression, the developer is able to model how the actor receiving the message should choose the appropriate model.

This would guarantee that only one model is chosen because the models would be stored in a map, whereby the keys are inherently unique. Here's an example:

subscribe:
  messageSelectionExpression: "$message.header#Content-Type"
  message:
    selection:
      application/json: 
        $ref: '#/components/messages/UserJsonMsg'
      application/xml: 
        $ref: '#/components/messages/UserXmlMsg'

I.e. Choose the message/model based on the header's Content-Type value.

[fmvilas]

Oh ok! So you're advocating for a totally different way without using oneOf. That is a different beast and looks like it makes sense to have something like this.

But let me clarify something, when you say "What if the payload could in fact be serialized to multiple models?" that's exactly what I meant before. It cannot happen. If that happens then it's an error. If you make use of oneOf, the payload MUST be defined by only one of the definitions.

This is how you'd do this right now:

subscribe:
  oneOf:
    - $ref: '#/components/messages/UserJsonMsg'
    - $ref: '#/components/messages/UserXmlMsg'

components:
  messages:
    UserJsonMsg:
      payload:
        ...
      bindings:
        http:
          headers:
            Content-Type:
              type: string
              enum: ['application/json']
    UserXmlMsg:
      payload:
        ...
      bindings:
        http:
          headers:
            Content-Type:
              type: string
              enum: ['application/xml']

But I get your point. It'd be great to have a routing mechanism at operation level. I'll label this as "Under Consideration" for now. Thanks!

[jasonmac01]

I think I see your point - you're suggesting that the spec is mandating that all messages be mutually exclusive. I suppose that works, but it puts a lot of work on any implementors trying to validate a definition. It may also be constraining in some cases.

Let me just run through one more example as documentation. I'm not expecting a response.

---

Rather than giving an example of XML versus JSON, let's use this example which is more agnostic of serialization method:

  oneOf:
    - $ref: '#/components/messages/Human'
    - $ref: '#/components/messages/SuperHero'

In this example, a SuperHero is a superset of Human. Same properties plus a few more.

If this is invalid because of the specification enforcement, then that could be too constraining, right?

If this is valid because they are in fact different models, then it presents a problem related to deserialization. Think of this issue from the perspective of automated code generation. The underlying deserializer would be able to deserialize an inbound Human message to either a Human or a SuperHero model. Which would be chosen?

As an alternative, if the entire response had some indicator to suggest the model to be chosen, then it enables a simplified code generation and validation process (lookup the key somewhere in the entire payload, map it to a unique model).

[fmvilas]

Agree there should be another mechanism for message selection, aside from oneOf. Actually, the purpose of oneOf wasn't about message selection but to act as a restrictive constraint. Usually, there's a field like type inside your message that would indicate what type of message it is. It should serve for both, uniqueness and message routing/selection.

[ciaranodonnell]

This last idea with the type field is how OneOf works in protocol buffers.
When you specify a OneOf, the code generation (for c# at least) makes an enum field that specifies which one is present.
Heres a link to the protobuf.net docs to explain that approach.
https://developers.google.com/protocol-buffers/docs/reference/csharp-generated#oneof

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity 😴
It will be closed in 30 days if no further activity occurs. To unstale this issue, add a comment with detailed explanation.
Thank you for your contributions ❤️

[smoya]

I just added the Strawman and Needs Champion label because I consider this a nice feature someone could be willing to champion.
I didn't read any request to this besides this issue anyway but let's keep it opened for some more time.