revisiting decorators for actions/properties/etc.

[rwb27]

I am sure we have discussed in the past the merits of defining a View subclass for each endpoint, e.g. in an Extension. I have a nasty feeling that I probably argued strongly against using decorators because they often make life quite confusing. However, I'm coming round to the idea. My use case is this:

I'm writing some code in an extension, which includes a function to do something. For argument's sake let's say that function should be an Action but I think the same consideration applies to properties. Ideally, I'd like it to be accessible either via HTTP or via Python. Currently, what I do is:

1. Write the Python function that does what I want (and give it a helpful docstring) as a method of my Extension
2. Define another class wrapping that Python function in a way that can be called nicely from the API
3. Reference said class from within the Extension class so it appears in the AP

There are several things that are less than ideal here:

1. You end up defining your arguments in multiple places - specifically, in both the method and the View class.
2. Docstrings don't propagate from the method to the View, so the HTTP API docs are not great

I can see a few ways to make it better:

1. Be more explicit about the fact that ActionView classes are more or less always wrapping a function to do an action. This would, for example, mean that instead of writing a post method, I would supply an action function. Doing that explicitly makes it way easier to wrap the action function nicely, e.g. propagate the docstrings.
2. Provide a decorator to put the definition of the ActionView right next to the function it's wrapping. This makes it harder to miss things being out of sync, and could also eliminate the need to register views explicitly.

I think if I argued against custom decorators in the past, it was because I felt they were non-standard ways of doing things, and wouldn't make sense to someone familiar with Flask. That's true, but I think there's enough that's non-standard about e.g. the way we generate our APISpec that this battle is possibly lost anyway. Also, if we define a more "helpful" ActionView subclass, it potentially provides a point of familiarity for Flask developers - so if they understand how @action maps onto ActionViewWrapper (which is potentially quite simple), it's easier to figure out what happens once you have a View.

Also, I think when we made this decision, it was more likely that you would want control over the get/post/delete/whatever methods. Now that the WoT specification is finalised, it's pretty well defined what each HTTP method should do, and I'm in favour of providing a mapping from HTTP methods to meaningful WoT interactions. I.e. a PropertyView always reads a property with GET and writes it with PUT and may define a way to update it with POST. So probably we should, by default, define a getter, setter, and update method - not all of which must exist. Similarly, an ActionView will always run actions with POST and return status with GET, so it makes sense to simply have a function responsible for running the action (which is what's currently implemented in ActionView.post. This also means that overriding ActionView.post would allow an action to do something custom before it's spawned in a background thread, such as validating arguments and returning an HTTP error code if they're wrong.

I'm raising this issue so @jtc42 can tell me why I was dead against this in the past, or if he remembers anything important that's not reflected in the discussion above.

[jtc42]

1. Be more explicit about the fact that ActionView classes are more or less always wrapping a function to do an action. This would, for example, mean that instead of writing a post method, I would supply an action function. Doing that explicitly makes it way easier to wrap the action function nicely, e.g. propagate the docstrings.

This is something I explored at one point though can't now remember if/where I ever implemented it. I basically changed the class so that the methods were named after the WoT standard, e.g. something like invokeaction, readproperty, writeproperty etc. They then got automatically mapped onto HTTP methods. It was pretty much the inverse of https://github.com/labthings/python-labthings/blob/master/src/labthings/views/op.py

2. Provide a decorator to put the definition of the ActionView right next to the function it's wrapping. This makes it harder to miss things being out of sync, and could also eliminate the need to register views explicitly.

It'd be good to have an example of how this would look, but I think I understand and agree.

I think if I argued against custom decorators in the past, it was because I felt they were non-standard ways of doing things, and wouldn't make sense to someone familiar with Flask. That's true, but I think there's enough that's non-standard about e.g. the way we generate our APISpec that this battle is possibly lost anyway.

Agreed. We're verging into "framework" rather than library territory here but I think we've been heading that way for a while now anyway.

Also, if we define a more "helpful" ActionView subclass, it potentially provides a point of familiarity for Flask developers - so if they understand how @action maps onto ActionViewWrapper (which is potentially quite simple), it's easier to figure out what happens once you have a View.

Also, I think when we made this decision, it was more likely that you would want control over the get/post/delete/whatever methods. Now that the WoT specification is finalised, it's pretty well defined what each HTTP method should do, and I'm in favour of providing a mapping from HTTP methods to meaningful WoT interactions. I.e. a PropertyView always reads a property with GET and writes it with PUT and may define a way to update it with POST. So probably we should, by default, define a getter, setter, and update method - not all of which must exist. Similarly, an ActionView will always run actions with POST and return status with GET, so it makes sense to simply have a function responsible for running the action (which is what's currently implemented in ActionView.post. This also means that overriding ActionView.post would allow an action to do something custom before it's spawned in a background thread, such as validating arguments and returning an HTTP error code if they're wrong.

See my above point. I think this is the best way forward.

I'm raising this issue so @jtc42 can tell me why I was dead against this in the past, or if he remembers anything important that's not reflected in the discussion above.

I'm having to remind myself to be honest... This all seems sensible though.

[jtc42]

I should add that a lot of the stuff you mentioned here is hanging over from before the WOT spec was standardised, and we were independently converging on more or less the same structure. We had some functionality that didn't map onto the draft spec, hence us defining everything in terms of HTTP methods still.

I really do think that making the HTTP method-based views a kind of behind-the-scenes thing you can override if needed, and focusing on WoT affordance-based views makes much more sense at this stage.

[rwb27]

ok, that sounds really sensible. I have a MWE of an action decorator for extensions working in the autofocus extension: https://gitlab.com/openflexure/openflexure-microscope-server/-/blob/1d1679268188f0bd96a4072cd397a5aa8964084e/openflexure_microscope/api/default_extensions/autofocus.py

Generalising this so it's ready for LabThings probably involves:

• Making it work outside extensions
• Splitting up the decorator into a factory function and a decorator version
• Repeating this for Properties and Events

That's enough work that it will take a while - and I might get some opinions on whether this verison of the autofocus plugin makes more sense before putting it in.