Use endpoint instead of model in the inference API

[davidkyle]

Update the inference APIs to refer to Inference endpoints rather than models. This involves renaming classes and updating comments but also includes the potentially breaking changes:

• inference.delete_model renamed to inference.delete
• inference.get_model renamed to inference.get
• inference.put_model renamed to inference.put

The format of the get response has also changed, the models field is renamed to endpoints:

#previously
{
    "models": [...]  
}

#with this change
{
    "endpoints": [...]  
}

In GetResponse.ts I've used an @alias tag to retain the models option

Corresponding Elasticsearch change elastic/elasticsearch#107704

[pquentin]

Thanks! Since this is for 8.15, I suggest waiting for elastic/elasticsearch#107704 so that we make sure 1/ the JSON spec is exactly the same and 2/ allow clients-flight-recorder to keep up with the change and validate this properly.

I left a few comments there too.

[davidkyle]

Thanks! Since this is for 8.15, I suggest waiting for elastic/elasticsearch#107704

That PR is green now. I hesitated before pressing the merge button as after merging we must have a compatible client ready for the next serverless release and when it comes to the breaking change I am hoping there is someway to soften the blow.

Taking Python as an example, existing code will use client.inference.put_model(...) and after this change the Python function is client.inference.put(...). Is there anyway of having inference.put_model resolve to inference.put so that we don't break existing code samples? Perhaps by having spec for both inference.put and inference.put_model but the latter is deprecated and not documented?

[pquentin]

Taking Python as an example, existing code will use client.inference.put_model(...) and after this change the Python function is client.inference.put(...). Is there anyway of having inference.put_model resolve to inference.put so that we don't break existing code samples? Perhaps by having spec for both inference.put and inference.put_model but the latter is deprecated and not documented?

I'm not sure we can deprecate an endpoint today if the endpoint itself is gone from Elasticsearch. (The only existing example is _knn_search.) Also, the clients are one thing, but on Serverless right now the usage of Kibana is quite high. And it's not like deprecating a function has any practical effect: things will break eventually. I suppose that's why tech preview is for.

That PR is green now. I hesitated before pressing the merge button as after merging we must have a compatible client ready for the next serverless release

We don't have great automation here, most serverless clients don't get frequent updates yet. We can coordinate to get the serverless Python client out when you want, if that helps.

[pquentin]

Thanks! Requested a few cosmetic fixes and asked about the models alias.

[pquentin]

Since we decided to break compatibility, should we keep models as an alias? For what it's worth, the current Python client does the wrong thing and would send models as is to the server. That could be fixed in time for 8.15 though.

[davidkyle]

I agree and have removed the alias.

The change breaks compatibility, we can do that because the API is neither GA or heavily used therefore it should be a clean break.

[pquentin]

Thanks! LGTM.

[github-actions]

Following you can find the validation results for the APIs you have changed.

API	Status	Request	Response
inference.delete	⚪	Missing test	Missing test
inference.get	🟠	Missing recording	Missing recording
inference.inference	⚪	Missing test	Missing test
inference.put	⚪	Missing test	Missing test

You can validate these APIs yourself by using the make validate target.

[davidkyle]

Thanks for the reviews