@ember-data packages

[runspired]

rendered

[lupestro]

Are the classes for which there will be no imports ever returned by public methods? If so, is there any scenario where not having an export for the classes returned by public methods prevents doing something we'd reasonably want to do? I can think of an ill-advised testing hack that somebody might use this information to do (shudder) but I'm asking if there's a legitimate case for needing those classes.

[chriskrycho]

Thinking along the lines @lupestro raises, we'll want a story for the TypeScript side of things for types we need to name but which are not importable. I'll have more later, as I'm excited about this.

[runspired]

Are the classes for which there will be no imports ever returned by public methods?

@lupestro I think you are asking " are there public classes for which there will be no public imports", to which the answer is YES.

is there any scenario where not having an export for the classes returned by public methods prevents doing something we'd reasonably want to do?

Generally speaking, public classes that don't have public import paths are ones which require extensive setup and lifecycle management by the store and for which it would be infeasible for end users to generate one except via the store.

For the most part, public store interfaces are available for accessing instances of these things, where public methods are not available these gaps should be addressed. The common case folks point out is snapshots, which is something we have in our sights for addressing soon.

[mehulkar]

This is great, we're already using some mixture of direct imports and DS, and it was never quite clear to me what the difference was (or why the guides don't show import Model from 'ember-data/model, for example)

[igorT]

Need to add RecordData in

[igorT]

Should add a discussion of moving the package into a separate github org

[tomdale]

I don't think the naming rules for Ember Data should deviate from the rules used for Ember.js in RFC #176.

[chriskrycho]

FWIW, I strongly agree with @tomdale here and I think if we have some set of limitations imposed by Node then we should work around Node's limitations.

However, it's worth note that it's certainly possible to support non-Node-friendly lookup rules (even with the "moduleResolution": "node" value set in tsconfig.json) in an application – we're doing that in our app today with my True Myth library, which does have nested module exports, not just at the root level:

importMaybe,{map}from'true-myth/maybe';

☝️ is not a legal Node import (and accordingly there are workarounds in the library and notes in its docs for Node-only consumers), but it is a resolvable import in contexts which understand ES modules, and for all the reasons that RFC #176 hashed out, it works extremely well to support the design of module APIs that express the actual intent and design of the library.

[chriskrycho]

An addendum to the above after I chatted with @runspired elsewhere for a while: one of the goals here is to make it so that the custom shenanigans I do to make this "just work" in the current Ember CLI pipeline is not necessary. And the base Node algorithm, as noted in that RFC, doesn't support it. That doesn't mean we shouldn't do the workarounds necessary, but it would be a lot nicer not to have to.

The net of that conversation was me feeling 🤔 🧐 about the right way forward here.

[runspired]

cc @chriskrycho

one of the goals here is to make it so that the custom shenanigans I do to make this "just work" in the current Ember CLI pipeline is not necessary.

Precisely. A key constraint is that both maintainers and consumers of ember-data should not need to implement work arounds to have first-class typescript and editor support. There are tradeoffs, and these concerns are being raised just to make sure we properly take them into account.

---

cc @tomdale

I don't think the naming rules for Ember Data should deviate from the rules used for Ember.js in RFC #176.

Putting aside the conversation surrounding editor / typescript support, (e.g. in a perfect world in which those constraints were fully solved), ember-data still faces unique constraints that Ember did not have to consider when implementing the Ember Modules RFC. Here, I believe we can escape with merely following the spirit of the law, and not implementing it to the rule.

The Spirit

Guiding Learners
We can group framework classes and utilities by functionality, making it clear what things are related > and how they should work together. People can feel confident they are getting only what they need
at that moment, not an entire framework that they're not sure they're benefiting from.

For that reason, package names should assist the developer in understanding what capabilities are added by bringing in that new package.

(source)

Furthermore, the original RFC acknowledges that it may not account for all the necessary constraints

And keep in mind, this RFC specifies a baseline module API. Nothing here precludes adding additional models in the future, as we discover missing pieces.

Most importantly, the mental model was the driving principle behind the organization proposed by the RFC

This proposal attempts to re-align module naming with that simplified mental model, placing everything into packages based on the chunk of functionality they provide:

This organization was also driven by two rules:

• Classes are Default Exports
• One Level Deep

If we organize by mental model, these rules conflict when put into practice in ember-data.

Specifically, AdapterError and it's subclasses are classes meant for use with Adapters. Transform and it's subclasses are classes meant for use with the reference Serializer implementations.

Using the default export rule and completely ignoring the one level deep requirement, organizing by the mental model could result in the following:

*note: mixins are treated as classes below, with the -mixin addition to the name to avoid an extra /

importAdapterfrom'@ember-data/adapter';importBuildURLfrom'@ember-data/adapter/build-url-mixin';importAdapterfrom'@ember-data/adapter/rest';importAdapterfrom'@ember-data/adapter/json-api';importAdapterError,{errorsHashToArray,errorsArrayToHash}from'@ember-data/adapter/error';importInvalidErrorfrom'@ember-data/adapter/error/invalid';importTimeoutErrorfrom'@ember-data/adapter/error/timeout';importAbortErrorfrom'@ember-data/adapter/error/abort';importUnauthorizedErrorfrom'@ember-data/adapter/error/unauthorized';importForbiddenErrorfrom'@ember-data/adapter/error/forbidden';importNotFoundErrorfrom'@ember-data/adapter/error/not-found';importConflictErrorfrom'@ember-data/adapter/error/conflict';importServerErrorfrom'@ember-data/adapter/error/server';//...importSerializerfrom'@ember-data/serializer';importSerializerfrom'@ember-data/serializer/json';importSerializerfrom'@ember-data/serializer/rest';importEmbeddedRecordsfrom'@ember-data/serializer/rest/embedded-records-mixin';importSerializerfrom'@ember-data/serializer/json-api';importTransformfrom'@ember-data/serializer/transform';importDateTransformfrom'@ember-data/serializer/transform/date';importBooleanTransformfrom'@ember-data/serializer/transform/boolean';importNumberTransformfrom'@ember-data/serializer/transform/number';importStringTransformfrom'@ember-data/serializer/transform/string';

What this shows is that a strict interpretation of the classes are default exports condition is incompatible with the one level deep condition when organizing according to the mental model.

However

There's also the following considerations here that Ember does not really have:

• sub-classes (multiple) of primary classes
• sub-classes that will almost never be directly imported (transforms) due to being automatically registered in a manner such that they may be "resolved"

With this in mind, we could potentially make the argument that sub-classes could be "named exports". In the case of Errors this follows the intuitive naming of these classes anyway (InvalidErrorTimeoutError etc.), but it does mean we end up with Transform and Adapter in the names of the reference implementations of adapters, serializers, and transforms. This is a tradeoff we may just want to accept: there is an outlier in the case of embedded-records though, which results in a potential move it to it's own package.

importAdapter,{RESTAdapter,JSONApiAdapter,BuildURLMixin}from'@ember-data/adapter';importAdapterError,{errorsHashToArray,errorsArrayToHash,InvalidError,TimeoutError,AbortError,UnauthorizedError,ForbiddenError,NotFoundError,ConflictError,ServerError}from'@ember-data/adapter/error';//...importSerializer,{JSONSerializer,RESTSerializer,JSONAPISerializer}from'@ember-data/serializer';importEmbeddedRecordsfrom'@ember-data/embedded-records-mixin';importTransform,{DateTransform,BooleanTransform,NumberTransform,StringTransform}from'@ember-data/serializer/transform';

In some cases, this is a nice cleanup, but in others it produces awkwardness.

There is still a third option, a middle ground, in which we recognize the increased importance of Adapter and Serializer and their relative Prominence in the mental model by keeping their subclasses as default exports. Personally, I find this the most intuitive, but it must be conceded that the "mixed" nature of this approach risks confusion... but likely no more confusion than "sub classes not being default exports" already risks.

importAdapter,{BuildURL}from'@ember-data/adapter';importAdapterfrom'@ember-data/adapter/rest';importAdapterfrom'@ember-data/adapter/json-api';importAdapterError,{errorsHashToArray,errorsArrayToHash,InvalidError,TimeoutError,AbortError,UnauthorizedError,ForbiddenError,NotFoundError,ConflictError,ServerError}from'@ember-data/adapter/error';//...importSerializerfrom'@ember-data/serializer';importSerializerfrom'@ember-data/serializer/json';importSerializer,{EmbeddedRecords}from'@ember-data/serializer/rest';importSerializerfrom'@ember-data/serializer/json-api';importTransform,{DateTransform,BooleanTransform,NumberTransform,StringTransform}from'@ember-data/serializer/transform';

It's been quite some time since I came to the realization that regardless of approach, we're weighing inescapable tradeoffs and making compromises when it comes to the adapter and serializer package organization. The main positive is that this split into packages is the first step towards providing a world in which these packages aren't necessary at all, and whatever compromises made here might end here too.

[tomdale]

I don't understand the concern about IDE/editor integration. Can someone unpack "custom build shenanigans"?

The Node module resolution algorithm absolutely supports nested modules in npm packages, in the form @scope/pkg/module, natively, without any additional configuration needed. You can ship packages like this as static files to npm and they will work just fine with TypeScript, Node, webpack, Rollup, or anything that uses the Node module resolution algorithm under the hood.

See https://docs.google.com/document/d/1CZC2rcpxffTDfRDs6p1cfbmKNLA6x5O-NtkJglDaBVs/preview#heading=h.k0mh3o8u5hx for a description of how Angular handles structuring their packages to accomplish this natively.

[runspired]

After further discussion, we've decided to accept this RFC.