feat: support additional metadata for rule deprecations #116
Conversation
bmish
force-pushed
the
deprecated-rule-metadata
branch
from
bb7efa2
to
28bb87f
Compare
bmish
force-pushed
the
deprecated-rule-metadata
branch
from
28bb87f
to
6d6cc2f
Compare
| type RuleMeta = { | ||
| deprecated: | ||
| | boolean // Existing boolean option, backwards compatible. | ||
| | string // Shorthand property for general deprecation message, such as why the deprecation occurred. Any truthy value implies deprecated. |
There was a problem hiding this comment.
should an empty string be prohibited here, so that “not deprecated” is unambiguously false always?
| | string // Shorthand property for general deprecation message, such as why the deprecation occurred. Any truthy value implies deprecated. | ||
| | { | ||
| message?: string; // General deprecation message, such as why the deprecation occurred. | ||
| url?: string; // URL to more information about this deprecation in general. |
There was a problem hiding this comment.
one of these properties should be required - so like { message } | { url } | { message, url }
| }; | ||
| meta?: { | ||
| message?: string; // Message about this specific replacement, such as how to use/configure the replacement rule to achieve the same results as the rule being replaced. | ||
| url?: string; // URL to more information about this specific deprecation/replacement. |
There was a problem hiding this comment.
LGTM, with the strong caveat that I've never reviewed an RFC as a member of the committer team before and thus should not be taken very seriously. 😄
Very excited about this one though! I think it'll be great for docs generation & helping folks migrate to new rules. 🚀
| * Omit if the replacement is in the same plugin. | ||
| */ | ||
| plugin?: | ||
| | string // Shorthand property for the plugin name i.e. "eslint-plugin-example" that contains the replacement rule. |
There was a problem hiding this comment.
[Clarification] Are both "eslint-plugin-example" and "example" allowed?
There was a problem hiding this comment.
By "shorthand property", I just mean that a string can be provided instead of the more detailed object format.
My intention was for the full plugin name to be provided here eslint-plugin-foo or @foo/eslint-plugin or whatever so there wouldn't be any ambiguity. Simply providing the plugin prefix foo doesn't necessarily tell us the full plugin name as there can be various formats...
|
|
||
| > semi (deprecated) \ | ||
| > Replaced by [semi](https://eslint.style/rules/js/semi) from [@stylistic/js](https://eslint.style/). \ | ||
| > Stylistic rules are being moved out of ESLint core. [Read more](https://eslint.org/blog/2023/10/deprecating-formatting-rules/). |
There was a problem hiding this comment.
[Praise] This is a really good demonstration of the proposed change and its benefits. (and also why eslint-doc-generator is great!)
| 2. Consolidate `meta.replacedBy` into `meta.deprecated.replacedBy`. This alternative was strongly considered. If we were to design things from scratch, we might choose this approach. However, there's not much tangible benefit to it besides organizational tidiness, and the downsides include: | ||
| - It creates yet another migration burden for the plugin ecosystem (potentially hundreds of plugins) to migrate from the old property to the new property | ||
| - There's overhead for us to go through the process of deprecating the old property and encouraging or helping plugins to move to the new one | ||
| - Tooling would need to support both properties, perhaps indefinitely due to the long tail of rules that may never be updated |
There was a problem hiding this comment.
[Question] Are there existing examples in the wild of rules that have stayed deprecated for a long time? I'm wondering if there really would be much pain from, say:
- Adding
meta.deprecated.replacedByand markingmeta.replacedByas a@deprecatedalias in ESLint v9 - Removing
meta.replacedByin ESLint v10
The rest of the rule metadata I'd be much more hesitant to change. But this Sourcegraph search for replacedBy in lint/rule files shows only a little under 300 code results, with the occasional script.
There was a problem hiding this comment.
Given that i almost never do major bumps, all of the import, react, and jsx-a11y plugins have examples - but many are unlikely to be using replacedBy, since i only learned about it within the last year.
There was a problem hiding this comment.
It sounds like @JoshuaKGoldberg has a preference for consolidating the properties, whereas @ljharb is neutral or slightly opposed? I'm open to hearing more opinions and still open to changing the proposal on this based on reviewer feedback.
It is true that a migration from meta.replacedBy to meta.deprecated.replacedBy would be a relatively lightweight and inconsequential migration, compared to more painful ESLint migrations or breaking changes we have performed in the past.
I fleshed out this section with some more pros and cons.
Interested to hear more feedback on this decision.
There was a problem hiding this comment.
I mean in general I prefer to avoid breaking changes - in particular because then, how will an eslint plugin support both eslint 9 and 10, given that RuleTester fails with unknown properties?
There was a problem hiding this comment.
support both eslint 9 and 10
Another strategy could be to leave it @deprecated for two major versions instead of one. Just to give plguins more time to adopt. I'm personally not in favor of this as major versions don't drop very frequently.
given that RuleTester fails with unknown properties
You could write some quick code around the RuleTester usage that adds/deletes properties as needed. Given how rarely it is for plugins to support quite so long a list of backwards compatibility versions as yours, I imagine the total amount of community work there isn't very high.
I think this point also emphasizes how not-very-breaking this is compared to potential other schema changes. RuleTester failing with unknown properties only impacts the plugin being tested, right? So even if a few plugins such as eslint-plugin-import need to do a bit of hackery in tests, I think that'd be justified for the user benefit of a more understandable+unified property.
There was a problem hiding this comment.
@nzakas @mdjermanovic I'm interested to hear if either of you have a preference between keeping the two existing properties or consolidating meta.replacedBy into meta.deprecated.replacedBy?
| you can remove this section. | ||
| --> | ||
|
|
||
| 1. Is there additional deprecation information we'd like to represent? Note that additional information can always be added later, but it's good to consider any possible needs now. |
There was a problem hiding this comment.
[Content] From eslint/eslint#18061 (comment):
The only other point that comes to mind for me is being able to specify options in the new rule. https://github.com/typescript-eslint/tslint-to-eslint-config -> https://github.com/typescript-eslint/tslint-to-eslint-config/blob/470d44de20beb7c7366de993edb8898d0766b8aa/docs/Architecture/Linters.md might be a good reference. Deprecated rules could map to multiple new ones, including specifying options and changed or missing functionality.
The "specifying options and changed or missing functionality" point might be relevant. How common is it that a rule deprecation points to a new rule with a particular option?
The only example that comes to mind for me is https://typescript-eslint.io/rules/no-type-alias. It's replaced by two rules:
- Behavior change, but recommended: https://typescript-eslint.io/rules/consistent-type-definitions with the default
"interface"option - Preserving the original behavior: https://eslint.org/docs/latest/rules/no-restricted-syntax, with an option like
["TypeAliasDeclaration"]
I personally don't think there's a need for to add functionality to the RFC for options (yet?). There are a lot of edge cases to account for and the message string is probably more than sufficient for most of them. But I think it'd be good to mention it as out of scope.
Another alternative that could be mentioned is storing a conversion function, the way tslint-to-eslint-config does. That in theory could be useful for users who want to get exact conversions... but again, I think is unnecessary work and best to just leave as out of scope.
There was a problem hiding this comment.
These are great callouts. I added a note about these. I do agree we can probably skip these for now in favor of just using message, but open to adding them now or later if someone wants to champion them.
There was a problem hiding this comment.
I'd prefer to explicitly not encourage discussion of replacements options as part of a rule's meta data. These are subject to change, and keeping that information in sync with a rule in a different repo is a maintenance headache. The source of truth should be in the replacement rule's docs -- the only job of the deprecated rule meta information is to successfully get users from the deprecated rule to the new rule's docs. Whether and which options to use should be found there.
Co-authored-by: Josh Goldberg ✨ <[email protected]>
Co-authored-by: Josh Goldberg ✨ <[email protected]>
| Pros: | ||
|
|
||
| - Organizational tidiness | ||
| - Avoids the inconsistent situation where `meta.deprecated` is `false` or omitted but `meta.replacedBy` has a value, resulting in ambiguity about whether the rule is deprecated or not |
There was a problem hiding this comment.
This is an open design question. If we stick with two separate properties, what does it mean for meta.deprecated to be false or omitted but meta.replacedBy to contain a value? It seems implied that a rule that has replacements would be deprecated, but it might not be explicitly marked as so which is concerning.
There was a problem hiding this comment.
Perhaps this is an opportunity for a RuleTester check?
There was a problem hiding this comment.
Thinking from the perspective of types: this is tough to represent precisely in TypeScript. Either...
- ... the types are straightforward (
{ deprecated?: boolean; replacedBy?: ReplaceBy }) and allow not-good situations ({ replacedBy: [...] }) - ... the types are something more complex such as a union of objects, and are more difficult to understand
Complex types are generally a symptom of complex code ideas. Which I think is true here. It's not intuitive to answer these questions.
Not saying that TypeScript specifically is a reason to do anything 😄 - just that its type system is a good way to sniff out conceptual complexity.
| name?: string; // Replacement rule name (without plugin prefix). | ||
| url?: string; // URL to rule documentation. | ||
| }; | ||
| meta?: { |
There was a problem hiding this comment.
I think meta.replacedBy[].meta is a bit confusing. It seems like the purpose of this meta is really to provide additional info about the replacement. Perhaps note or info would be a more appropriate name?
| Pros: | ||
|
|
||
| - Organizational tidiness | ||
| - Avoids the inconsistent situation where `meta.deprecated` is `false` or omitted but `meta.replacedBy` has a value, resulting in ambiguity about whether the rule is deprecated or not |
There was a problem hiding this comment.
Perhaps this is an opportunity for a RuleTester check?
| you can remove this section. | ||
| --> | ||
|
|
||
| 1. Is there additional deprecation information we'd like to represent? Note that additional information can always be added later, but it's good to consider any possible needs now. |
There was a problem hiding this comment.
I'd prefer to explicitly not encourage discussion of replacements options as part of a rule's meta data. These are subject to change, and keeping that information in sync with a rule in a different repo is a maintenance headache. The source of truth should be in the replacement rule's docs -- the only job of the deprecated rule meta information is to successfully get users from the deprecated rule to the new rule's docs. Whether and which options to use should be found there.
Summary
This RFC suggests a format for storing additional information in rule metadata about rule deprecations and replacement rules, allowing automated documentation and website tooling to generate more informative deprecation notices.
Related Issues