:warning: Although I find the type-safety of this framework appealing, the Swift compiler currently has enough trouble with it that it can become difficult to reason about errors produced by small typos. Similarly, auto-complete fails to provide reasonable suggestions much of the time. If you get the code right, everything compiles, otherwise it can suck to figure out what is wrong. This is mostly a concern when creating resource objects in-code (servers and test suites must do this). Writing a client that uses this framework to ingest JSON API Compliant API responses is much less painful. :warning:
1. Allow creation of Swift types that are easy to use in code but also can be encoded to- or decoded from **JSON API v1.0 Spec** compliant payloads without lots of boilerplate code.
The big caveat is that, although the aim is to support the JSON API spec, this framework ends up being _naturally_ opinionated about certain things that the API Spec does not specify. These caveats are largely a side effect of attempting to write the library in a "Swifty" way.
If you find something wrong with this library and it isn't already mentioned under **Project Status**, let me know! I want to keep working towards a library implementation that is useful in any application.
To run the included Playground files, create an Xcode project using Swift Package Manager, then create an Xcode Workspace in the root of the repository and add both the generated Xcode project and the playground to the Workspace.
Note that Playground support for importing non-system Frameworks is still a bit touchy as of Swift 4.2. Sometimes building, cleaning and building, or commenting out and then uncommenting import statements (especially in the Entities.swift Playground Source file) can get things working for me when I am getting an error about `JSONAPI` not being found.
- [ ] Support sparse fieldsets. At the moment, not sure what this support will look like. A client can likely just define a new model to represent a sparse population of another model in a very specific use case. On the server side, it becomes much more appealing to be able to support arbitrary combinations of omitted fields.
- [x] Only allow `TransformedAttribute` and its derivatives as stored properties within `Attributes` struct. Computed properties can still be any type because they do not get encoded or decoded.
- [ ] (Maybe) Use `KeyPath` to specify `Includes` thus creating type safety around the relationship between a primary resource type and the types of included resources.
- [ ] (Maybe) Replace `SingleResourceBody` and `ManyResourceBody` with support at the `Document` level to just interpret `PrimaryResource`, `PrimaryResource?`, or `[PrimaryResource]` as the same decoding/encoding strategies.
- [ ] Support sideposting. JSONAPI spec might become opinionated in the future (https://github.com/json-api/json-api/pull/1197, https://github.com/json-api/json-api/issues/1215, https://github.com/json-api/json-api/issues/1216) but there is also an existing implementation to consider (https://jsonapi-suite.github.io/jsonapi_suite/ruby/writes/nested-writes). At this time, any sidepost implementation would be an awesome tertiary library to be used alongside the primary JSONAPI library. Maybe `JSONAPISideloading`.
- [ ] Error or warning if an included resource object is not related to a primary resource object or another included resource object (Turned off or at least not throwing by default).
In this documentation, in order to draw attention to the difference between the `JSONAPI` framework (this Swift library) and the **JSON API Spec** (the specification this library helps you follow), the specification will consistently be referred to below as simply the **SPEC**.
A `ResourceObjectDescription` is the `JSONAPI` framework's representation of what the **SPEC** calls a *Resource Object*. You might create the following `ResourceObjectDescription` to represent a person in a network of friends:
Note that an `enum` type is used here for the `ResourceObjectDescription`; it could have been a `struct`, but `ResourceObjectDescription`s do not ever need to be created so an `enum` with no `case`s is a nice fit for the job.
Once you have a `ResourceObjectDescription`, you _create_, _encode_, and _decode_`ResourceObjects` that "fit the description". If you have a `CreatableRawIdType` (see the section on `RawIdType`s below) then you can create new `ResourceObjects` that will automatically be given unique Ids, but even without a `CreatableRawIdType` you can encode, decode and work with resource objects.
The `ResourceObject` and `ResourceObjectDescription` together with a `JSONAPI.Meta` type and a `JSONAPI.Links` type embody the rules and properties of a JSON API *Resource Object*.
A `ResourceObject` needs to be specialized on four generic types. The first is the `ResourceObjectDescription` described above. The others are a `Meta`, `Links`, and `MaybeRawId`.
The second generic specialization on `ResourceObject` is `Meta`. This is described in its own section [below](#jsonapimeta). All `Meta` at any level of a JSON API Document follow the same rules. You can use `NoMetadata` if you do not need to package any metadata with the `ResourceObject`.
The third generic specialization on `ResourceObject` is `Links`. This is described in its own section [below](#jsonnapilinks). All `Links` at any level of a JSON API Document follow the same rules, although the **SPEC** makes different suggestions as to what types of links might live on which parts of the Document. You can use `NoLinks` if you do not need to package any links with the `ResourceObject`.
The last generic specialization on `ResourceObject` is `MaybeRawId`. This is either a `RawIdType` that can be used to uniquely identify `ResourceObjects` or it is `Unidentified` which is used to indicate a `ResourceObject` does not have an `Id` (which is useful when a client is requesting that the server create a `ResourceObject` and assign it a new `Id`).
The raw type of `Id` to use for the `ResourceObject`. The actual `Id` of the `ResourceObject` will not be a `RawIdType`, though. The `Id` will package a value of `RawIdType` with a specialized reference back to the `ResourceObject` type it identifies. This just looks like `Id<RawIdType, ResourceObject<ResourceObjectDescription, Meta, Links, RawIdType>>`.
Having the `ResourceObject` type associated with the `Id` makes it easy to store all of your resource objects in a hash broken out by `ResourceObject` type; You can pass `Ids` around and always know where to look for the `ResourceObject` to which the `Id` refers. This encapsulation provides some type safety because the Ids of two `ResourceObjects` with the "raw ID" of `"1"` but different types will not compare as equal.
Often you can use one `RawIdType` for many if not all of your `ResourceObjects`. That means you can save yourself some boilerplate by using `typealias`es like the following:
Note that I am assuming an unidentified person is a "new" person. I suspect that is generally an acceptable conflation because the only time the **SPEC** allows a *Resource Object* to be encoded without an `Id` is when a client is requesting the given *Resource Object* be created by the server and the client wants the server to create the `Id` for that object.
There are two types of `Relationships`: `ToOneRelationship` and `ToManyRelationship`. A `ResourceObjectDescription`'s `Relationships` type can contain any number of `Relationship` properties of either of these types. Do not store anything other than `Relationship` properties in the `Relationships` struct of a `ResourceObjectDescription`.
In addition to identifying resource objects by Id and type, `Relationships` can contain `Meta` or `Links` that follow the same rules as [`Meta`](#jsonapimeta) and [`Links`](#jsonapilinks) elsewhere in the JSON API Document.
To describe a relationship that may be omitted (i.e. the key is not even present in the JSON object), you make the entire `ToOneRelationship` or `ToManyRelationship` optional. However, this is not recommended because you can also represent optional relationships as nullable which means the key is always present. A `ToManyRelationship` can naturally represent the absence of related values with an empty array, so `ToManyRelationship` does not support nullability at all. A `ToOneRelationship` can be marked as nullable (i.e. the value could be either `null` or a resource identifier) like this:
`Relationship` values boil down to `Ids` of other resource objects. To access the `Id` of a related `ResourceObject`, you can use the custom `~>` operator with the `KeyPath` of the `Relationship` from which you want the `Id`. The friends of the above `Person``ResourceObject` can be accessed as follows (type annotations for clarity):
The `Attributes` of a `ResourceObjectDescription` can contain any JSON encodable/decodable types as long as they are wrapped in an `Attribute`, `ValidatedAttribute`, or `TransformedAttribute``struct`.
Sometimes you need to use a type that does not encode or decode itself in the way you need to represent it as a serialized JSON object. For example, the Swift `Foundation` type `Date` can encode/decode itself to `Double` out of the box, but you might want to represent dates as ISO 8601 compliant `String`s instead. The Foundation library `JSONDecoder` has a setting to make this adjustment, but for the sake of an example, you could create a `Transformer`.
A `Transformer` just provides one static function that transforms one type to another. You might define one for an ISO 8601 compliant `Date` like this:
Note that the first generic parameter of `TransformAttribute` is the type you expect to decode from JSON, not the type you want to end up with after transformation.
If you make your `Transformer` a `ReversibleTransformer` then your life will be a bit easier when you construct `TransformedAttributes` because you have access to initializers for both the pre- and post-transformed value types. Continuing with the above example of a `ISODateTransformer`:
You can also creator `Validators` and `ValidatedAttribute`s. A `Validator` is just a `Transformer` that by convention does not perform a transformation. It simply `throws` if an attribute value is invalid.
You can add computed properties to your `ResourceObjectDescription.Attributes` struct if you would like to expose attributes that are not explicitly represented by the JSON. These computed properties do not have to be wrapped in `Attribute`, `ValidatedAttribute`, or `TransformedAttribute`. This allows computed attributes to be of types that are not `Codable`. Here's an example of how you might take the `person[\.name]` attribute from the example above and create a `fullName` computed property.
If your computed property is wrapped in a `AttributeType` then you can still use the default subscript operator to access it (as would be the case with the `person[\.fullName]` example above). However, if you add a property to the `Attributes``struct` that is not wrapped in an `AttributeType`, you must either access it from its full path (`person.attributes.newThing`) or with the "direct" subscript accessor (`person[direct: \.newThing]`). This keeps the subscript access unambiguous enough for the compiler to be helpful prior to explicitly casting, comparing, or storing the result.
`ResourceObject` is a value type, so copying is its default behavior. There are two common mutations you might want to make when copying a `ResourceObject`:
The entirety of a JSON API request or response is encoded or decoded from- or to a `Document`. As an example, a JSON API response containing one `Person` and no included resource objects could be decoded as follows:
A JSON API Document is guaranteed by the **SPEC** to be "data", "metadata", or "errors." If it is "data", it may also contain "metadata" and/or other "included" resources. If it is "errors," it may also contain "metadata."
The first generic type of a `JSONAPIDocument` is a `ResourceBody`. This can either be a `SingleResourceBody<PrimaryResource>` or a `ManyResourceBody<PrimaryResource>`. You will find zero or one `PrimaryResource` values in a JSON API document that has a `SingleResourceBody` and you will find zero or more `PrimaryResource` values in a JSON API document that has a `ManyResourceBody`. You can use the `Poly` types (`Poly1` through `Poly6`) to specify that a `ResourceBody` will be one of a few different types of `ResourceObject`. These `Poly` types work in the same way as the `Include` types described below.
If you expect a `SingleResourceBody` to sometimes come back `null`, you should make your `PrimaryResource` optional. If you do not make your `PrimaryResource` optional then a `null` primary resource will be considered an error when parsing the JSON.
You cannot, however, use an optional `PrimaryResource` with a `ManyResourceBody` because the **SPEC** requires that an empty document in that case be represented by an empty array rather than `null`.
The second generic type of a `JSONAPIDocument` is a `Meta`. This `Meta` follows the same rules as `Meta` at any other part of a JSON API Document. It is described below in its own section, but as an example, the JSON API document could contain the following pagination info in its meta entry:
The fourth generic type of a `JSONAPIDocument` is an `Include`. This type controls which types of `ResourceObject` are looked for when decoding the "included" part of the JSON API document. If you do not expect any included resource objects to be in the document, `NoIncludes` is the way to go. The `JSONAPI` framework provides `Include`s for up to six types of included resource objects. These are named `Include1`, `Include2`, `Include3`, and so on.
**IMPORTANT**: The number trailing "Include" in these type names does not indicate a number of included resource objects, it indicates a number of _types_ of included resource objects. `Include1` can be used to decode any number of included resource objects as long as all the resource objects are of the same _type_.
To specify that we expect friends of a person to be included in the above example `JSONAPIDocument`, we would use `Include1<Person>` instead of `NoIncludes`.
The fifth generic type of a `JSONAPIDocument` is an `APIDescription`. The type represents the "JSON:API Object" described by the **SPEC**. This type describes the highest version of the **SPEC** supported and can carry additional metadata to describe the API.
You can specify this is not part of the document by using the `NoAPIDescription` type.
You can describe the API by a version with no metadata by using `APIDescription<NoMetadata>`.
You can supply any `JSONAPI.Meta` type as the metadata type of the API description.
The final generic type of a `JSONAPIDocument` is the `Error`. You should create an error type that can decode all the errors you expect your `JSONAPIDocument` to be able to decode. As prescribed by the **SPEC**, these errors will be found in the root document member `errors`.
A `Meta` struct is totally open-ended. It is described by the **SPEC** as a place to put any information that does not fit into the standard JSON API Document structure anywhere else.
You can specify `NoMetadata` if the part of the document being described should not contain any `Meta`.
A `Links` struct must contain only `Link` properties. Each `Link` property can either be a `URL` or a `URL` and some `Meta`. Each part of the document has some suggested common `Links` to include but generally any link can be included.
You can specify `NoLinks` if the part of the document being described should not contain any `Links`.
If you want to create new `JSONAPI.ResourceObject` values and assign them Ids then you will need to conform at least one type to `CreatableRawIdType`. Doing so is easy; here are two example conformances for `UUID` and `String` (via `UUID`):
There is not anything special going on at the `JSONAPI.Attributes` and `JSONAPI.Relationships` levels, so you can easily provide custom key mappings by taking advantage of `Codable`'s `CodingKeys` pattern. Here are two models that will encode/decode equivalently but offer different naming in your codebase:
You can safely provide your own encoding or decoding functions for your Attributes struct if you need to as long as you are careful that your encode operation correctly reverses your decode operation. Although this is generally not necessary, `AttributeType` provides a convenience method to make your decoding a bit less boilerplate ridden. This is what it looks like:
This advanced feature may not ever be useful, but if you find yourself in the situation of dealing with an API that does not 100% follow the **SPEC** then you might find meta-attributes are just the thing to make your resource objects more natural to work with.
Suppose, for example, you are presented with the unfortunate situation where a piece of information you need is only available as part of the `Id` of a resource object. Perhaps a user's `Id` is formatted "{integer}-{createdAt}" where "createdAt" is the unix timestamp when the user account was created. The following `UserDescription` will expose what you need as an attribute. Realistically, the following example code is still terrible for its error handling. Using a `Result` type and/or invariants would clean things up substantially.
This works because `createdAt` is defined in the form: `var {name}: ({ResourceObject}) -> {Value}` where `{ResourceObject}` is the `JSONAPI.ResourceObject` described by the `ResourceObjectDescription` containing the meta-attribute.
This advanced feature may not ever be useful, but if you find yourself in the situation of dealing with an API that does not 100% follow the **SPEC** then you might find meta-relationships are just the thing to make your resource objects more natural to work with.
Similarly to Meta-Attributes, Meta-Relationships allow you to represent non-compliant relationships as computed relationship properties. In the following example, a relationship is created from some attributes on the JSON model.
This works because `friend` is defined in the form: `var {name}: ({ResourceObject}) -> {Identifier}` where `{ResourceObject}` is the `JSONAPI.ResourceObject` described by the `ResourceObjectDescription` containing the meta-relationship.
The following serves as a sort of pseudo-example. It skips server/client implementation details not related to JSON:API but still gives a more complete picture of what an implementation using this framework might look like. You can play with this example code in the Playground provided with this repo.
The `JSONAPI` framework is packaged with a test library to help you test your `JSONAPI` integration. The test library is called `JSONAPITesting`. It provides literal expressibility for `Attribute`, `ToOneRelationship`, and `Id` in many situations so that you can easily write test `ResourceObject` values into your unit tests. It also provides a `check()` function for each `ResourceObject` type that can be used to catch problems with your `JSONAPI` structures that are not caught by Swift's type system. You can see the `JSONAPITesting` in action in the Playground included with the `JSONAPI` repository.
