Please remove onError parameter from format/formatToParts

[FrankYFTang]

Currently, the format/formatToParts function
1.3.3 Intl.MessageFormat.prototype.format ( [ values [ , onError ] ] )
and
1.3.4 Intl.MessageFormat.prototype.formatToParts ( [ values [ , onError ] ] )

has a new onError parameters and introduce a new requirement for format function in ECMA402.

https://tc39.es/proposal-intl-messageformat/#sec-intl.messageformat.prototype.format

I am very concern about this and I do not think this functionality is useful. Could we remove this functionality from this proposal and just let the format/formatToParts throw ?

How would this functionality be useful in real life programming? How would that user code looks like? Why this functionality (handling the missing data) cannot be done in the user land?

[FrankYFTang]

Currently, it is used in HandleMessageFormatError(ctx, error).
1.5.5 ResolveExpression ( ctx, expression )
in https://tc39.es/proposal-intl-messageformat/#sec-resolveexpression

in 1.5.5.1 ResolveFunction ( ctx, source, resArg, annotation )
https://tc39.es/proposal-intl-messageformat/#sec-resolvefunction

1.5.6 FormatMarkupPart ( ctx, markup )
https://tc39.es/proposal-intl-messageformat/#sec-formatmarkuppart

1.5.10 ResolveValue ( ctx, value )
https://tc39.es/proposal-intl-messageformat/#sec-resolvevalue

1.3.3 Intl.MessageFormat.prototype.format ( [ values [ , onError ] ] )
https://tc39.es/proposal-intl-messageformat/#sec-intl.messageformat.prototype.format

1.3.4 Intl.MessageFormat.prototype.formatToParts ( [ values [ , onError ] ] )
https://tc39.es/proposal-intl-messageformat/#sec-intl.messageformat.prototype.formatToParts

1.5.4.1 SelectPattern ( ctx )
https://tc39.es/proposal-intl-messageformat/#sec-selectpattern

Why don't we just throw some kind of Error in these cases instead?

[eemeli]

The immediate reason for the onError argument is that it's required by MF2:

In all cases, when encountering a runtime error, a message formatter MUST provide some representation of the message. An informative error or errors MUST also be separately provided.

At a slightly deeper level, the reason for this approach is to ensure that localization will always provide some result to a user, rather than just failing. MessageFormat is different from the other Intl formatters in that it relies not only on testable CLDR data, but also on user-provided data, i.e. the message. For locales beyond the source locale used by the original developer, that data is the result of translation work, and translation may introduce bugs that can cause a runtime error. Sure, validation can catch many such bugs, but no system is perfect.

In other words, an error thrown by e.g. NumberFormat or DateTimeFormat is actionable by a developer, and rarely seen in production use, but an error thrown by MessageFormat is often not visible to a developer or even caught by CI systems, and much more likely to end up happening at production runtime when it's not actionable.

Throwing from MessageFormat would effectively introduce an expectation that all of its users have a full test suite that they run for all locales when any of its contents change. I do not think this is a reasonable expectation, as I'm not sure that anything like that is currently done by anyone.

[FrankYFTang]

I do not understand about this, you argue MessageFormat is different because " is often not visible to a developer or even caught by CI systems, and much more likely to end up happening at production runtime when it's not actionable."

If that is truely not actionable, then why do you need an onError parameter at all? In that case, the developer cannot take any action anyway so the onError could only be a noop, right?

Currently the HandleMessageFormatError will pass completionRecord.[[Value]] to onError if one is provided. So from my understand, there are two kind of conditions

Case A: The error is not actionable
Case B: The error is actionable

Therefore, for Case A, the completionRecord.[[Value]] is not needed to pass to onError and an onError is not need to be registered, because it is not actionable even if the onError receive the completionRecord.[[Value]] and can do nothing (since it is not actionable, mean no action can be taken to make thing better)

Now, for Case B, I think that is the part that the onError and completionRecord.[[Value]] is useful, right?

so in that case, your current design requires

fmt = new Intl.MessageFormat(...);
let myError = function(errorData) { /* do something */ };
let str = fmt.format(obj, myError);
/* do something with str */

Why it cannot be

fmt = new Intl.MessageFormat(...);
let myError = function(errorData) { /* do something */ };
try {
   str = fmt.format(obj);
} catch(e) {
   myError(e.errorData) 
}

I understand, sometime MessageFormat.format() may encounter error internally and need to still return a string but with informaiton about what were fallbacked during the format() operation. I wonder, could that be done via

fmt = new Intl.MessageFormat(...);
let myLogFallbackInfo = function(fallbackReasons) { /* do something */ };
let result = fmt.formatWithFallbackInfo(obj);
let str = result.formatted;
myLogFallbackInfo(result.fallbackReasons);

@sffc
The difference is w/ the onError handler, you can only see the partial information about the fallback one at a time, during the format but the formatWithFallbackInfo approach (or. we can use a different name) allow the myLogFallbackInfo to see all the fallback information at once (it could be an array of all the error data of 5 fallbacks during the format () call)

[sffc]

I don't have an official position yet, but it seems that Intl.MessageFormat.prototype.format could throw an exception of type Intl.MessageFormatError which could have a field such as fallbackMessage that contains the message with fallback placeholders as required by the specification.

[eemeli]

Replying to @FrankYFTang:

I do not understand about this, you argue MessageFormat is different because " is often not visible to a developer or even caught by CI systems, and much more likely to end up happening at production runtime when it's not actionable."

If that is truely not actionable, then why do you need an onError parameter at all? In that case, the developer cannot take any action anyway so the onError could only be a noop, right?

Continuing a bit from my previous comment, the two main use cases that the onError callback is designed to handle are:

1. Providing a way to report errors encountered in production to developers, while still providing the user with a fallback formatted result.
2. Throwing rather than warning about errors, e.g. when they are encountered during testing.

So while a thrown error isn't necessarily actionable to the user, reporting it to a developer could make it actionable by allowing them to fix it later.

Now, for Case B, I think that is the part that the onError and completionRecord.[[Value]] is useful, right?

In case it's not clear in the spec, in all the cases where onError is called, the completionRecord.[[Value]] is an Error, and its value is never used by anything within Intl.MessageFormat after onError is called with it as an argument.

So onError is not able to change anything about the fallback formatting of the message or its placeholders.

I understand, sometime MessageFormat.format() may encounter error internally and need to still return a string but with informaiton about what were fallbacked during the format() operation. I wonder, could that be done via

fmt = new Intl.MessageFormat(...);
let myLogFallbackInfo = function(fallbackReasons) { /* do something */ };
let result = fmt.formatWithFallbackInfo(obj);
let str = result.formatted;
myLogFallbackInfo(result.fallbackReasons);

In theory, that approach could work. An additional method formatToPartsWithFallbackInfo() would also be required, and we would need to decide what to do with errors when the plain format() and formatToParts() are called. So a user would need to pick between four different "format" methods. An additional object would also be created when formatting every message.

I don't think this would be better or clearer for users.

The difference is w/ the onError handler, you can only see the partial information about the fallback one at a time, during the format but the formatWithFallbackInfo approach (or. we can use a different name) allow the myLogFallbackInfo to see all the fallback information at once (it could be an array of all the error data of 5 fallbacks during the format () call)

Note that additional information like the message source data or an identifier for the message would also probably need to be included when reporting/logging errors.

---

Replying to @sffc

I don't have an official position yet, but it seems that Intl.MessageFormat.prototype.format could throw an exception of type Intl.MessageFormatError which could have a field such as fallbackMessage that contains the message with fallback placeholders as required by the specification.

In theory, this approach could also work. We might need two different errors, to cover errors encountered during both format() and formatToParts(), as their fallback messages have different shapes. This approach would also require error objects to carry complex user-defined values within them, which would be novel for the spec, and might be quite challenging to account for. As it's possible for multiple errors to be encountered while formatting a message, this would also swallow the errors beyond the first (while still formatting their placeholders with fallback values), or require the thrown error to include something like an additionalErrors array.

[eemeli]

This topic was also covered during the 2023-10-25 incubator call where @sffc also participated, and I recall @ljharb having some thoughts on attaching user-defined data to errors, though that may have been limited to values that are later read by spec internals.

[ryzokuken]

Given the context provided by @eemeli above, I too strongly feel that the onError pattern is the best one for us. Specifically, I strongly disagree with the idea put forward in #58 (comment) and I can't comprehend why stretching the problem space to the two extremes is helpful.

If we go the "always throw" route then the obvious downside in my opinion would be that due to the nature of MessageFormat as @eemeli mentioned, a developer would practically speaking be forced to wrap all usage of Intl.MessageFormat in try-catch blocks, significantly affecting the ergonomics of the proposal: I believe this is absolutely the wrong pattern in JavaScript, where try-catch is supposed to be used lightly. Secondly, by itself this approach would only give the user the first error and fallback so in the best case scenario we'd probably want to go through the whole message "collecting" errors and then throw a sort of AggregateError with the full fallback at which point we're back to square one.

If we go the "never throw" route it will be really difficult for developers to be able to debug or report errors. Furthermore it'll be a waste of effort since the parser and formatter already knew about the errors which instead of getting reported got discarded and then the developer had to put in substantial effort to go through the result and find those out again, at times losing context that would be vital to be able to debug it.

Because of both of these outcomes, I don't understand the benefit of trying to create a dichotomy here, since we need to come to a single solution that caters to the most common use-cases. This is addressed really well by the onError pattern which is not only popular in JavaScript but by far the most appropriate way to do non-blocking error handling as you may notice from these examples.

[ljharb]

What are the pros/cons of an alternative where instead of returning a bare string, we always return an object with the bare string as a "value" property (or something), and then that object gives us a place to put an errors array?

[eemeli]

What are the pros/cons of an alternative where instead of returning a bare string, we always return an object with the bare string as a "value" property (or something), and then that object gives us a place to put an errors array?

The main cost is that we would need to create that object for every formatting call, and that the result of a format() call wouldn't be a string, which is the case for all other Intl formatters.

So it would add a cost to every successful formatting call in order to deal with exceptional cases, and make this API different from all the other ones. To some extent it'd be a bit of a return to the kind of API this proposal had before #22, where the "formatted" output still needs a bit of processing to be used, albeit of course less than with the resolveMessage() approach.

[sffc]

ICU APIs have been moving in the direction of returning an upgraded type from .format that can be converted to a string but also has extra metadata. We use .formatToString when the function directly returns a string.

I'd be okay with .format returning an upgraded type, maybe even one with a valueOf that returns a string, or introducing a new function like formatFull that returns the upgraded object.

[sffc]

TG2 discussion: https://github.com/tc39/ecma402/blob/main/meetings/notes-2024-05-23.md#please-remove-onerror-parameter-from-formatformattoparts

We want to seek feedback from TG1 about error handling code style.