Currently, I'm working on new product and making REST API for both - public and internal needs. I started with {json:api} specification and I was pretty happy with it until I faced some questions I cannot find answers to.
According to JSON API specification, every resource MUST contain id.
http://jsonapi.org/format/
Every resource object MUST contain an id member and a type member. The values of the id and type members MUST be strings.
And that's fine in many cases but not all.
Most of our endpoints are about "resources"
If I ask for a "things" collection (http://example.com/things)
{
"data": [{
"type": "things",
"id": "1",
"attributes": {
"title": "first"
},
"links": {
"self": "http://example.com/things/1"
}
}, {
"type": "things",
"id": "1",
"attributes": {
"title": "second"
},
"links": {
"self": "http://example.com/things/2"
}
}]
}
If I ask for a single "things" resource (http://example.com/things/1)
{
"data": {
"type": "things",
"id": "1",
"attributes": {
"title": "first"
},
"links": {
"self": "http://example.com/things/1"
}
}
}
But what to do with endpoints which are not about resources and does not have ID?
For example, in our application, there is an endpoint http://example.com/stats which should return stats of current logged in user. Like
{
"active_things": 23,
"last_login": "2017"
}
There is no id for this "resource" (it's not actually a resource, is it?). Backend just collects some "stats" for logged in user and returns an object of stats. There many endpoints like this in this application, for example, we have Notification center page where the user can change email addresses for different notifications.
So frontend app (single-page-app) first has to get current values and it sends the request to GET http://example.com/notification-settings.
{
"notifications_about_new_thing": "arunas#example.com",
"notification_about_other_thing": "arunas#example.com"
}
And there are many more endpoints like this. The problem is - how to return these responses in JSONAPI format? There is no ID in these endpoints.
And the biggest question is - why nobody else is facing this issue (at least I cannot find any discussion about this)? :D All APIs I ever made has some endpoints which don't have "id".
I have two ideas, first is to fake id, like "id": "doesnt_matter", the second - do not use json-api for these endpoints. But I don't like both of them.
Think RESTfully and everything can (must) be a resource. There is no "logged in" user as there are no sessions in RESTful APIs as they are stateless. There's no session state maintained between REST API invocations, so you have to be explicit about who the user is.
In this case, the resource is the user who has some stats attributes (in the simple case) or perhaps a relationship to a separate stats relationship (more complicated, not shown):
GET /users/1234
{
"data": {
"type": "users",
"id": "1234",
"attributes": {
"name": "etc.",
"active_things": 23,
"last_login": "2017"
}
}
}
I'm no JSON API expert- but it's worth noting that while JSON API is a concrete specification, it is not the same thing as JSON, nor as a REST API. If you don't like its semantics, I agree with commenters who argue, "Don't use it." If you are going to use JSON API, do so in a compliant way, where every response is a resource; every resource has an ID and a type; and additional information is supplied as attributes of the resource.
Toward your question, I'm thinking about something similar where my application returns computation results. Now on the one hand, these are not strictly "resources" and so I've been toying with the idea of returning the raw result as an array (which I believe would be valid JSON, with a caveat), e.g:
[ 47 ]
On the other hand, there is the idea that the results are the results of a computation that the client specified RESTfully, in which case one of the following two cases is likely true:
The same request submitted later is likely to have the same result. This suggests that in fact the result really is a resource.
The same request submitted later is likely to have a different result. This suggests that the client may want to track how results change for various queries, and so at least the query parameters should be part of the response.
In both cases, the response really is a 'result' object, and even though it doesn't have an ID per se, it does have an identity. If nothing else fits, the ID could be the query that generated the response.
This seems RESTful to me. User #n2ygk suggests that this is not correct as regards the JSON API spec, that an ID should simply be a unique ID and not have another semantic interpretation.
I'd love to hear other perspectives.
Related
I am creating a JSON API that implements the JSON:API specification. I have a question about it however, and this question applies more generally to RESTful design in general: What is the recommended way to handle the "creation" of a resource, where one of the attributes is "calculated" by the server?
In my example, I have a POST /auth/tokens endpoint that accepts a user's credentials and returns a JWT. I've used a POST endpoint, because it seems to me that we are creating a token resource, even if that token is not saved to a database. However, according to JSON:API, what would the proper request/response look like? This?:
POST /auth/tokens
{
"data": {
"type": "tokens",
"attributes": {
"email": "...",
"password": "..."
}
}
}
However, does it even make sense to create a token with an email and password? It seems that it would be creating a token for an email/password. Is there a difference?
More importantly, what would the response look like? It seems like it would look something like:
{
"data": {
"type": "tokens",
"attributes": {
"token": "..."
}
}
}
But the specification states:
Every resource object MUST contain an id member and a type member. The values of the id and type members MUST be strings.
Since the tokens aren't saved to the database, I don't really have an ID for them. What should I do?
JSON:API specification doesn't say anything about such case but if you look on some live implementations like https://docs.unit.co/customer-api-tokens#customers-create-customer-bearer-token they use similar approach as you described. It seems to be just ok in this case to omit ID for a resource which is short lived.
I have the following data structure that contains an array of sectionIds. They are stored in the order in which they were completed:
applicationProgress: ["sectionG", "sectionZ", "sectionA"]
I’d like to be able to do something like:
GET /application-progress - expected: sectionG, sectionZ, sectionA
GET /application-progress?filter[first]=true - expected: sectionG
GET /application-progress?filter[current]=true - expected: sectionA
GET /application-progress?filter[previous]=sectionZ - expected: sectionG
I appreciated the above URLs are incorrect, but I’m not sure how to name/structure them to get the expected data e.g. Are the resources here "sectionids"?
I'd like to adhere to the JSON:API specification.
UPDATE
I'm looking to adhere to JSON:API v1.0
In terms of resources I believe I have "Section" and "ProgressEntry". Each ProgressEntry will have a one-to-one relationship with a Section.
I'd like to be able to query within the collection e.g.
Get the first item in the collection:
GET /progress-entries?filter[first]
Returns:
{
"data": {
"type": "progress-entries",
"id": "progressL",
"attributes": {
"sectionId": "sectionG"
},
"relationships": {
"section": {
"links": {
"related": "http://example.com/sections/sectionG"
}
}
}
},
"included": [
{
"links": {
"self": "http://example.com/sections/sectionG"
},
"type": "sections",
"id": "sectionG",
"attributes": {
"id": "sectionG",
"title": "Some title"
}
}
]
}
Get the previous ProgressEntry given a relative ProgressEntry. So in the following example find a ProgressEntry whose sectionId attribute equals "sectionZ" and then get the previous entry (sectionG). I wasn't clear before that the filtering of this is based on the ProgressEntry's attributes:
GET /progress-entries?filter[attributes][sectionId]=sectionZ&filterAction=getPreviousEntry
Returns:
{
"data": {
"type": "progress-entries",
"id": "progressL",
"attributes": {
"sectionId": "sectionG"
},
"relationships": {
"section": {
"links": {
"related": "http://example.com/sections/sectionG"
}
}
}
},
"included": [
{
"links": {
"self": "http://example.com/sections/sectionG"
},
"type": "sections",
"id": "sectionG",
"attributes": {
"id": "sectionG",
"title": "Some title"
}
}
]
}
I started to comment on jelhan's reply though my answer was just to long for a reasonable comment on his objection, hence I include it here as it more or less provides a good introduction into the answer anyways.
A resource is identified by a unique identifier (URI). A URI is in general independent from any representation format else content-type negotiation would be useless. json-api is a media-type that defines the structure and semantics of representations exchanged for a specific resource. A media-type SHOULD NOT force any constraints on the URI structure of a resource as it is independent from it. One can't deduce the media-type to use based on a given URI even if the URI contains something like vnd.api+json as this might just be a Web page talking about json:api. A client may as well request application/hal+json instead of application/vnd.api+json on the same URI and receive the same state information just packaged in a different representation syntax, if the server supports both representation formats.
Profiles, as mentioned by jelhan, are just extension mechanisms to the actual media-type that allow a general media-type to specialize through adding further constraints, conventions or extensions. Such profiles use URIs similar to XML namespaces, and those URIs NEED NOT but SHOULD BE de-referencable to allow access to further documentation. There is no talk about the URI of the actual resource other than given by Web Linking that URIs may hint a client on the media-type to use, which I would not recommend as this requires a client to have certain knowledge about that hint.
As mentioned in my initial comments, URIs shouldn't convey semantics as link relations are there for!
Link-relations
By that, your outlined resource seems to be a collection of some further resources, sections by your domain language. While pagination as defined in json:api does not directly map here perfectly, unless you have so many sections that you want to split these into multiple pages, the same concept can be used using standardized link relations defined by IANA.
Here, at one point a server may provide you a link to the collection resource which may look like this:
{
"links": {
"self": "https://api.acme.org/section-queue",
"collection": "https://api.acme.org/app-progression",
...
},
...
}
Due to the collection link relation standardized by IANA you know that this resource may hold a collection of entries which upon invoking may return a json:api representation such as:
{
"links": {
"self": "https://api.acme.org/app-progression",
"first": "https://api.acme.org/app-progression/sectionG",
"last": "https://api/acme.org/app-progression/sectionA",
"current": "https://api.acme.org/app-progression",
"up": "https://api.acme.org/section-queue",
"https://api/acme.org/rel/section": "https://api.acme.org/app-progression/sectionG",
"https://api/acme.org/rel/section": "https://api.acme.org/app-progression/sectionZ",
"https://api/acme.org/rel/section": "https://api.acme.org/app-progression/sectionA",
...
},
...
}
where you have further links to go up or down the hierarchy or select the first or last section that finished. Note the last 3 sample URIs that leverages the extension relation types mechanism defined by RFC 5988 (Web Linking).
On drilling down the hierarchy further you might find links such as
{
"links": {
"self": "https://api.acme.org/app-progression/sectionZ",
"first": "https://api.acme.org/app-progression/sectionG",
"prev": "https://api.acme.org/app-progression/sectionG",
"next": "https://api.acme.org/app-progression/sectionA",
"last": "https://api.acme.org/app-progression/sectionA",
"current": "https://api.acme.org/app-progression/sectionA",
"up": "https://api.acme.org/app-progression",
...
},
...
}
This example should just showcase how a server is providing you with all the options a client may need to progress through its task. It will simply follow the links it is interested in. Based on the link relation names provided a client can make informed choices on whether the provided link is of interest or not. If it i.e. knows that a resource is a collection it might to traverse through all the elements and processes them one by one (or by multiple threads concurrently).
This approach is quite common on the Internet and allows the server to easily change its URI scheme over time as clients will only act upon the link relation name and only invoke the URI without attempting to deduce any logic from it. This technique is also easily usable for other media-types such as application/hal+json or the like and allows each of the respective resources to be cached and reused by default, which might take away load from your server, depending on the amount of queries it has to deal with.
Note that no word on the actual content of that section was yet said. It might be a complex summary of things typical to sections or it might just be a word. We could classify it and give it a name, as such even a simple word is a valid target for a resource. Further, as Jim Webber mentioned, your resources you expose via HTTP (REST) and your domain model are not identical and usually do not map one-to-one.
Filtering
json:api allows to group parameters together semantically by defining a customized x-www-form-urlencoded parsing. If content-type negotiation is used to agree on json:api as representation format, the likelihood of interoperability issues is rather low, though if such a representation is sent unsolicitedly parsing of such query parameters might fail.
It is important to mention that in a REST architecture clients should only use links provided by the server and not generate one on their own. A client usually is not interested in the URI but in the content of the URI, hence the server needs to know how to act upon the URI.
The outlined suggestions can be used but also URIs of the form
.../application-progress?filter=first
.../application-progress?filter=current
.../application-progress?filter=previous&on=sectionZ
can be used instead. This approach should in addition to that also work on almost all clients without the need to change their url-encoded parsing mechanism. In addition to that he management overhead to return URIs for other media-types generated may be minimized as well. Note that each of the URIs in the example above represent their own resource and a cache will store responses to such resources based on the URI used to retrieve such results. Queries like .../application-progress?filter=next&on=sectionG and .../application-progress?filter=previous&on=sectionA which retrieve basically the same representations are two distinctive resources which will be processed two times by your API as the response of the first query can't be reused as the cache key (URI) is different. According to Fielding caching is one of the few constraints REST has which has to be respected otherwise you are violating it.
How you design such URIs is completely up to you here. The important thing is, how you teach a client when to invoke such URIs and when it should not. Here, again, link-relations can and should be used.
Summary
In summary, which approach you prefer is up to you as well as which URI style you choose. Clients, especially in a REST environment, do not care about the structure of the URI. They operate on link-relations and use the URI just for invoking it to progress on with their task. As such, a server API should help a client by teaching it what it needs to know like in a text-based computer game in the 70/80's as mentioned by Jim Webber. It is helpful to think of the interaction model to design as affordances and state machine as explained by Asbjørn Ulsberg .
While you could apply filtering on grouped parameters provided by json:api such links may only be usable within the `json:api´ representation. If you copy & paste such a link to a browser or to some other channel, it might not be processable by that client. Therefore this would not be my first choice, TBH. Whether or not you design sections to be their own resource or just properties you want to retrieve is your choice here as well. We don't know really what sections are in your domain model, IMO it sounds like a valid resource though that may or may not have further properties.
I came across HATEOAS on my researches and was thinking : doesn't HATEOAS multiplicate HTTP requests ?
Let's take the basic customer and order example.
Let's say you want to retrieve an order, the endpoint would be /orders/2
with the following JSON response :
{
"id": 2,
"total": 50.00,
"links": [{
"rel": "customer",
"href": "http://api.domain.com/customer/1
}]
}
Now what if I also need the customer ? Do I have to make another request to /customer/1 ? Doesn't this overload the HTTP traffic ?
Couldn't I get the couple customer + order with a single endpoint like /customers/1/orders/2 ?
Or just send the customer in the /orders/2 JSON response ?
{
"id": 2,
"total": 50.00,
"customer": {
"id": 1,
"name": "Dylan Gauthier"
}
}
What's the benefit(s) of one solution or another ? When do I need one or the other ?
Thanks ! :-)
If the server only supplies the customer and order separately, then you have to make two requests regardless of whether they are following REST or not.
Nothing about REST or its HATEOAS constraint prevents the server from providing both customer and order in the same resource, exactly as you have suggested:
GET /orders/2
{
"id": 2,
"total": 50.00,
"customer": {
"name": "Dylan Gauthier"
}
}
But the customer in that response has no connection to the identifier /customers/1 — the server could combine the two ideas:
{
"id": 2,
"total": 50.00,
"links": [{
"rel": "customer",
"href": "http://api.domain.com/customer/1
}],
"resources": {
"http://api.domain.com/customer/1": {
"name": "Dylan Gauthier"
}
}
}
or better yet, group the links by their relation to the requested resource:
{
"id": 2,
"total": 50.00,
"links": {
"customer": [{
"href": "http://api.domain.com/customer/1"
}]
},
"resources": {
"http://api.domain.com/customer/1": {
"name": "Dylan Gauthier"
}
}
}
Whilst this would make it a bit more work for the client to print the name of the customer (nothing at all taxing, mind), it allows the client to fetch more information about the customer if they want to!
Just to add to Nicholas' answer:
Embedding related resources
Pros: saves you a trip to the server
Cons: While it saves you a trip the first time and may be a few lines of code, you are giving up on caching: if something changes in a related resource (that you embedded) client cache is no more valid, so the client has to make the request again. Of course, assuming you leverage HTTP caching. Which you should...
If you want to go this route, you are better off using something like GraphQL... but wait!
Going "pure" HATEOS
Pros: resources have independent life-cycles; easier to make each (type of) resource evolve without impacting the others. By fully leveraging the cache, overtime, the overall performance is far better.
Cons: more requests (at first access), this might be a little slower on first access; some more code to manage the HATEOS thing...
I personally tend to use the second approach whenever possible.
The classic web analogy:
If it can help, a classic website is just another api that serves html related resources, the client app being the browser itself. If you have ever done some html/css/js, you might want to approach it the same way:
For the given particular website, given its navigation architecture...etc would you rather inline all/part of the css/js (the related resources) in the html pages (the main resource) or not.
I understand HATEOAS represents the applications state by sending all actions that can be performed at that point in time within the application as it's response (HAL, JSON-LD, etc).
For example, viewing an account resource of a bank may allow you to deposit, withdraw or close the account (OPTIONS which may return UPDATE and DELETE verbs).
In terms of runtime discoverability of these links (by the consuming client), how might one go about this?
If the purpose of sending these links is the decouple the client from the server and drive the state by the hypermedia in the response, there must be an amount of knowledge the developer must hardcode in the application in order to make any sense of the responses being returned.
I understanding sending OPTIONS requests is the way to determine the current state of the resource and what you can do next, but in order to discover the actual URIs to use - would these simply be hardcoded as COOL URIs?
Like #VoicesOfUnreason said, in HATEOAS URIs are discoverable (and not documented) so that they can be changed. That is, unless they are the very entry points into your system (Cool URIs, the only ones that can be hard-coded by clients) - and you shouldn't have too many of those if you want the ability to evolve the rest of your system's URI structure in the future. This is in fact one of the most useful features of REST.
For the remaining non-Cool URIs, they can be changed over time, and your API documentation should spell out the fact that they should be discovered at runtime through hypermedia traversal.
Looking at the Richardson's Maturity Model (level 3), this would be where links come into play. For example, from the top level, say /api/version(/1), you would discover there's a link to the groups. Here's how this could look in a tool like HAL Browser:
Root:
{
"_links": {
"self": {
"href": "/api/root"
},
"api:group-add": {
"href": "http://apiname:port/api/group"
},
"api:group-search": {
"href": "http://apiname:port/api/group?pageNumber={pageNumber}&pageSize={pageSize}&sort={sort}"
},
"api:group-by-id": {
"href": "http://apiname:port/api/group/id" (OR "href": "http://apiname:port/api/group?id={id}")
}
}
}
The add would simply be a POST to that endpoint, and then you'd have 2 GET methods.
GET /api/group?pageNumber=0&pageSize=20&sort=asc
which could return something like this:
{
"groups": [
{
"id": 123,
"name": "Test Group"
},
{
"id": 134,
"name": "Tennis squad"
}
]
}
Then once you drill down to a particular group (say #123):
{
"Id" : 123,
"Name" : "test",
"_links": {
"self": {
"href": "/api/group/1" (OR "/api/group?id=1")
},
"edit": {
"href": "http://apiname:port/api/group/1"
},
"api:delete": {
"href": "http://apiname:port/api/group/1"
},
"api:items-query": {
"href": "http://apiname:port/api/bonus?groupId=1"
}
}
}
Here, the edit would simply be a PUT, and then you'll need a DELETE (see level 2 of REST in that same link), as for the items, you probably know best if they are just a property, or another endpoint; you could even embed them to be returned in the same call that's retrieving a group.
The advantage here would be that the client would only need to know the relationship (link) name (well obviously besides the resource structure/properties), while the server would be mostly free to alter the relationship (and resource) url.
There's a bunch of prior art around on trying to create expressive, discoverable hypermedia. You might want to review:
http://json-ld.org/
http://www.markus-lanthaler.com/hydra/
I am thinking maybe a series of if statement that checks for certain properties to determine the state or maybe even switch statements. Is this is correct path - or is there better means of hypermedia discovery?
My current thinking is that you want to be shaping your ideas more along the lines of negotiating and following a protocol; so think state machine rather than if statements.
For starters, review How To GET a Cup of Coffee.
The hyperlinks in the documents served by RESTBucks are designed to assist the client in negotiating the RESTBucks protocol; the assumption that the client already understands that protocol is baked into the model. In other words, the client already understands that negotiating the protocol will allow it to reach it's goal.
Of course, there could be multiple protocols that serve the same goal. For instance RESTBucks could also support a "Give Away Day Old Coffee" protocol; announcing the presence of each, the client would be expected to choose which is the better expression of the goal, and follow that path.
If I wanted to create (POST) a new resource linking two independent resources, what is the most proper - with respect to HATEOAS and REST principles - way to structure the entity of the request?
Any references in RFCs, W3C documents, Fielding's thesis, etc., about the proper way for a client to request two independent resources be linked together would be most valuable. Or, if what I'm interested in is simply outside the scope of REST, HATEOAS, an explanation of why would also be great.
Hopefully my question above is clear. If not, here's a scenario and some background to ground the question.
Let's say I have two independent resources: /customer and /item, and a third resource /order intended to the two.
If I'm representing these resource to the client in a HATEOAS-like way (say with JSON-LD), a customer might (minimally) look like:
{
"#id": "http://api.example.com/customer/1"
}
and similarly an item like:
{
"#id": "http://api.example.com/item/1"
}
I'm more concerned about what scheme the entity of the POST request should have, rather than the URL I'm addressing the request to. Assuming I'm addressing the request to /order, would POSTing the following run afoul of HATEOAS and REST principles in any way?
{
"customer": {"#id": "http://api.example.com/customer/1"},
"item": {"#id": "http://api.example.com/item/1"}
}
To me, this seems intuitively OK. However, I can't find much or any discussion of the right way to link two independent resources with a POST. I discovered the LINK and UNLINK HTTP methods, but these seem inappropriate for a public API.
The client does not build URIs, so this is wrong unless these resource identifiers or at least their template came from the service. It is okay to use the id numbers instead of the URIs until you describe this in the response which contains the POST link.
An example from the hydra documentation:
{
"#context": "http://www.w3.org/ns/hydra/context.jsonld",
"#id": "http://api.example.com/doc/#comments",
"#type": "Link",
"title": "Comments",
"description": "A link to comments with an operation to create a new comment.",
"supportedOperation": [
{
"#type": "CreateResourceOperation",
"title": "Creates a new comment",
"method": "POST",
"expects": "http://api.example.com/doc/#Comment",
"returns": "http://api.example.com/doc/#Comment",
"possibleStatus": [
... Statuses that should be expected and handled properly ...
]
}
]
}
The "http://api.example.com/doc/#Comment" contains the property descriptions.
{
"#context": "http://www.w3.org/ns/hydra/context.jsonld",
"#id": "http://api.example.com/doc/#Comment",
"#type": "Class",
"title": "The name of the class",
"description": "A short description of the class.",
"supportedProperty": [
... Properties known to be supported by the class ...
{
"#type": "SupportedProperty",
"property": "#property", // The property
"required": true, // Is the property required in a request to be valid?
"readable": false, // Can the client retrieve the property's value?
"writeable": true // Can the client change the property's value?
}
]
}
A supported property can have an rdfs:range, which describes the value constraints. This is not yet (2015.10.22.) added to the hydra vocab as far as I can tell, but I don't have time to follow the project. I think you still can use the rdfs:range instead of waiting for a hydra range.
So in your case you could add an item property with a range of http://api.example.com/doc/#Item and so on. I assume you could add the links of the alternatives, something like http://api.example.com/items/, so you could generate a select input box. Be aware that this technology is not stable yet.
So you can send a simple JSON as POST body {item: {id:1}, customer: {id:1}} or something like that, which you generate based on the POST link. The RDF is for the client not for the server. The server can understand the data structure it requires, it does not need RDF. You don't need a dictionary to understand yourself...