Add design doc for expression attributes

[eemeli]

As discussed on the past two calls. A preview of the doc is available here.

See also #450 for a possible implementation of the proposed design.

[aphillips]

As mentioned elsewhere, I would prefer a prefix to name content.

Thought: will there be classes or packages of attributes that can be plugged in, e.g.:

{$foo @its:translate=no @Xliff:canCopy=yes}

[aphillips]

I would tighten this up. I would imagine that we'll add to the spec text similar to:

An implementation MUST provide functions the ability to query the set of attributes that apply to the calling expression.

(I didn't say "an API" on purpose)

[eemeli]

What you're proposing is actually the opposite of what I'm proposing. The formatting context as currently specified is an explicit set of fields, extensible by an implementation but not custom functions:

Implementations MAY include additional fields in their formatting context.

So the proposal here would allow an implementation to modify the formatting context as it's applied to an expression, but it would not allow for a custom function definition "to query the set of attributes".

If it makes sense for an attribute to influence the behaviour of a custom function, and that attribute isn't an explicit part of the implementation's formatting context, why should its value be communicated via an expression attribute rather than a function option?

[aphillips]

A formatting context would be "a way to query the set of attributes"? I like your wording:

Implementations MAY include additional fields in their formatting context.

Although my question would be: how do I ensure that my custom annotation reaches my function's runtime via the context?

[eemeli]

That wording is from our current formatting spec, actually.

how do I ensure that my custom annotation reaches my function's runtime via the context?

By using a function option to convey that information, rather than an expression attribute.

[aphillips]

My thinking here goes:

• We are defining a feature "expression attributes"
• We allow implementation or user defined attributes
• How do these attributes reach the runtime?

The answer cannot be function options or I would have used a function option and not an expression attribute.

Apparently the answer is: the attributes appear in the formatting context and we can just say that?

[eemeli]

In that situation, why not use a function option? If it's something affecting the behaviour of a single function and its value can be defined in the expression, why not use a function option for it?

Apparently the answer is: the attributes appear in the formatting context and we can just say that?

We have not defined the shape of the formatting context with any specificity, and I don't think we should -- it's really an implementation detail. If an implementation wants to make its formatting context extensible in some way and for a custom expression attribute to affect it, I think that's very much an implementation detail and outside the scope of this spec.

[aphillips]

We don't have to define the shape of the formatting context with any specificity in order to say that implementations must provide access to any attribute values in a message's context. We already have at least one contextual value (the locale) that we require the implementation to ensure.

So I agree that we don't want to define exactly how a function gets access to the context. We might not even require that there be a way for it to be obtained. But we can still require that any contextual mechanism (using whatever shape) provides access to the attribute.

If we don't follow my thinking through, wouldn't we be introducing a feature that doesn't work in at least some implementations?

[eemeli]

To rephrase and reiterate my own position on this and the currently proposed phrase, I do not think that functions should have access to the expression attributes or their values.

I do not think it would be a good idea for implementations to be required to provide access to any attribute values in a message's context. As we already have function options, I do not think that expression attributes should be duplicating their functionality. Allowing custom functions to access expression attributes would do exactly that.

Message formatting happens in some context that has properties such as the current locale, timezone, etc. that have an effect on the results. In the spec, we refer to this as the "formatting context", but do not exhaustively define its shape or properties. This is implementation-defined, and implementations may internally allow for its extensibility.

The general use case I have in mind for expression attributes that have a formatting impact is to provide a way to locally override formatting context values or error handling behaviour.

What is the use case for allowing custom functions to access custom attributes?

[aphillips]

What is the use case for allowing custom functions to access custom attributes?

What is the use case for allowing custom attributes? If we allow them, don't we need to provision them? If we don't allow them, won't implementations re-invent them to suit their own needs?

On some level, all functions are custom functions. They are all described through the function registry and made available to message writers.

[aphillips]

Actually, it's easy--but it isn't machine readable. In addition, there is no guarantee that resource-level commentary will be preserved during the translation process or be operable to MT or CAT tools. This is particularly true if the translation process segments inside the MF pattern.

[eemeli]

It's hard if there is more than one expression with the exact same syntax within a message.

For example, consider this message:

let $count = {$count :number}
match {$count}
when one {{$count} thing}
when * {{$count} things}

With that, how can I assign the equivalent of @example=2 to the {$count} in the * variant, but not the match selector or the placeholder in the one variant?

[aphillips]

With that, how can I assign the equivalent of @example=2 to the {$count} in the * variant, but not the match selector or the placeholder in the one variant?

Actually you just did an admirable--but not machine readable--job of it:

@comment="@example=2 to the {$count} in the * variant, but not the match selector or the placeholder in the one variant"

My first sentence is more of a bit of humor, really. We're in violent agreement here. The problem I wanted to highlight was the separation represents a probably loss of functionality as soon as you parse the message.

[eemeli]

Rebased & squashed changes to get rid of the irrelevant Prettier fixes.