This guide describes a set of HTTP+JSON API design practices, originally extracted from work on the Heroku Platform API.
This guide informs additions to that API and also guides new internal APIs at Heroku. We hope it’s also of interest to API designers outside of Heroku.
Our goals here are consistency and focusing on business logic while avoiding design bikeshedding. We’re looking for a good, consistent, well-documented way to design APIs, not necessarily the only/ideal way.
We assume you’re familiar with the basics of HTTP+JSON APIs and won’t cover all of the fundamentals of those in this guide.
We welcome contributions to this guide.
Require TLS to access the API, without exception. It’s not worth trying to figure out or explain when it is OK to use TLS and when it’s not. Just require TLS for everything.
Respond to non-TLS requests with 403 Forbidden. Redirects are
discouraged since they allow sloppy/bad client behaviour without
providing any clear gain. Clients that rely on redirects double up on
server traffic and render TLS useless since sensitive data will already
have been exposed during the first call.
Version the API from the start. Use the Accepts header to communicate
the version, along with a custom content type, e.g.:
Accept: application/vnd.heroku+json; version=3
Prefer not to have a default version, instead requiring clients to explicitly peg their usage to a specific version.
Include an ETag header in all responses, identifying the specific
version of the returned resource. The user should be able to check for
staleness in their subsequent requests by supplying the value in the
If-None-Match header.
Include a Request-Id header in each API response, populated with a
UUID value. If both the server and client log these values, it will be
helpful for tracing and debugging requests.
Paginate any responses that are liable to produce large amounts of data.
Use Content-Range headers to convey pagination requests. Follow the
example of the Heroku Platform API on Ranges
for the details of request and response headers, status codes, limits,
ordering, and page-walking.
Return appropriate HTTP status codes with each response. Successful responses should be coded according to this guide:
200: Request succeeded for aGETcalls, and forDELETEorPATCHcalls that complete synchronously201: Request succeeded for aPOSTcall that completes synchronously202: Request accepted for aPOST,DELETE, orPATCHcall that will be processed asynchronously206: Request succeeded onGET, but only a partial response returned: see above on ranges
Pay attention to the use of authentication and authorization error codes:
401 Unauthorized: Request failed because user is not authenticated403 Forbidden: Request failed because user does not have authorization to access a specific resource
Return suitable codes to provide additional information when there are errors:
422 Unprocessable Entity: Your request was understood, but contained invalid parameters429 Too Many Requests: You have been rate-limited, retry later500 Internal Server Error: Something went wrong on the server, check status site and/or report the issue
Refer to the HTTP response code spec for guidance on status codes for user error and server error cases.
Provide the full resource representation (i.e. the object with all
attributes) whenever possible in the response. Always provide the full
resource on 200 and 201 responses, including PUT/PATCH and DELETE
requests, e.g.:
$ curl -X DELETE \
https://service.com/apps/1f9b/domains/0fd4
HTTP/1.1 200 OK
Content-Type: application/json;charset=utf-8
...
{
"created_at": "2012-01-01T12:00:00Z",
"hostname": "subdomain.example.com",
"id": "01234567-89ab-cdef-0123-456789abcdef",
"updated_at": "2012-01-01T12:00:00Z"
}
202 responses will not include the full resource representation, e.g.:
$ curl -X DELETE \
https://service.com/apps/1f9b/dynos/05bd
HTTP/1.1 202 Accepted
Content-Type: application/json;charset=utf-8
...
{}
Accept serialized JSON on PUT/PATCH/POST request bodies, either
instead of or in addition to form-encoded data. This creates symmetry
with JSON-serialized response bodies, e.g.:
$ curl -X POST https://service.com/apps \
-H "Content-Type: application/json" \
-d '{"name": "demoapp"}'
{
"id": "01234567-89ab-cdef-0123-456789abcdef",
"name": "demoapp",
"owner": {
"email": "[email protected]",
"id": "01234567-89ab-cdef-0123-456789abcdef"
},
...
}
Use the plural version of a resource name unless the resource in question is a singleton within the system (for example, in most systems a given user would only ever have one account). This keeps it consistent in the way you refer to particular resources.
Prefer endpoint layouts that don’t need any special actions for
individual resources. In cases where special actions are needed, place
them under a standard actions prefix, to clearly delineate them:
/resources/:resource/actions/:action
e.g.
/runs/{run_id}/actions/stop
Use downcased and dash-separated path names, for alignment with hostnames, e.g:
service-api.com/users
service-api.com/app-setups
Downcase attributes as well, but use underscore separators so that attribute names can be typed without quotes in JavaScript, e.g.:
service_class: "first"
In some cases it may be inconvenient for end-users to provide IDs to identify a resource. For example, a user may think in terms of a Heroku app name, but that app may be identified by a UUID. In these cases you may want to accept both an id or name, e.g.:
$ curl https://service.com/apps/{app_id_or_name}
$ curl https://service.com/apps/97addcf0-c182
$ curl https://service.com/apps/www-prod
Do not accept only names to the exclusion of IDs.
In data models with nested parent/child resource relationships, paths may become deeply nested, e.g.:
/orgs/{org_id}/apps/{app_id}/dynos/{dyno_id}
Limit nesting depth by preferring to locate resources at the root path. Use nesting to indicate scoped collections. For example, for the case above where a dyno belongs to an app belongs to an org:
/orgs/{org_id}
/orgs/{org_id}/apps
/apps/{app_id}
/apps/{app_id}/dynos
/dynos/{dyno_id}
Give each resource an id attribute by default. Use UUIDs unless you
have a very good reason not to. Don’t use IDs that won’t be globally
unique across instances of the service or other resources in the
service, especially auto-incrementing IDs.
Render UUIDs in downcased 8-4-4-4-12 format, e.g.:
"id": "01234567-89ab-cdef-0123-456789abcdef"
Provide created_at and updated_at timestamps for resources by default,
e.g:
{
...
"created_at": "2012-01-01T12:00:00Z",
"updated_at": "2012-01-01T13:00:00Z",
...
}These timestamps may not make sense for some resources, in which case they can be omitted.
Accept and return times in UTC only. Render times in ISO8601 format, e.g.:
"finished_at": "2012-01-01T12:00:00Z"
Serialize foreign key references with a nested object, e.g.:
{
"name": "service-production",
"owner": {
"id": "5d8201b0..."
},
...
}Instead of e.g:
{
"name": "service-production",
"owner_id": "5d8201b0...",
...
}This approach makes it possible to inline more information about the related resource without having to change the structure of the response or introduce more top-level response fields, e.g.:
{
"name": "service-production",
"owner": {
"id": "5d8201b0...",
"name": "Alice",
"email": "[email protected]"
},
...
}Generate consistent, structured response bodies on errors. Include a
machine-readable error id, a human-readable error message, and
optionally a url pointing the client to further information about the
error and how to resolve it, e.g.:
HTTP/1.1 429 Too Many Requests
{
"id": "rate_limit",
"message": "Account reached its API rate limit.",
"url": "https://docs.service.com/rate-limits"
}Document your error format and the possible error ids that clients may
encounter.
Rate limit requests from clients to protect the health of the service and maintain high service quality for other clients. You can use a token bucket algorithm to quantify request limits.
Return the remaining number of request tokens with each request in the
RateLimit-Remaining response header.
The first time a user sees your API is likely to be at the command line, using curl. It’s much easier to understand API responses at the command-line if they are pretty-printed. For the convenience of these developers, pretty-print JSON responses, e.g.:
{
"beta": false,
"email": "[email protected]",
"id": "01234567-89ab-cdef-0123-456789abcdef",
"last_login": "2012-01-01T12:00:00Z",
"created_at": "2012-01-01T12:00:00Z",
"updated_at": "2012-01-01T12:00:00Z"
}Instead of e.g.:
{"beta":false,"email":"[email protected]","id":"01234567-89ab-cdef-0123-456789abcdef","last_login":"2012-01-01T12:00:00Z", "created_at":"2012-01-01T12:00:00Z","updated_at":"2012-01-01T12:00:00Z"}Be sure to include a trailing newline so that the user’s terminal prompt isn’t obstructed.
For most APIs it will be fine performance-wise to pretty-print responses all the time. You may consider for performance-sensitive APIs not pretty-printing certain endpoints (e.g. very high traffic ones) or not doing it for certain clients (e.g. ones known to be used by headless programs).
Provide a machine-readable schema to exactly specify your API. Use
prmd to manage your schema, and ensure
it validates with prmd verify.
Provide human-readable documentation that client developers can use to understand your API.
If you create a schema with prmd as described above, you can easily
generate Markdown docs for all endpoints with with prmd doc.
In addition to endpoint details, provide an API overview with information about:
- Authentication, including acquiring and using authentication tokens.
- API stability and versioning, including how to select the desired API version.
- Common request and response headers.
- Error serialization format.
- Examples of using the API with clients in different languages.
Provide executable examples that users can type directly into their terminals to see working API calls. To the greatest extent possible, these examples should be usable verbatim, to minimize the amount of work a user needs to do to try the API, e.g.:
$ export TOKEN=... # acquire from dashboard
$ curl -is https://[email protected]/users
If you use prmd to generate Markdown docs, you will get examples for each endpoint for free.
Describe the stability of your API or its various endpoints according to its maturity and stability, e.g. with prototype/development/production flags.
See the Heroku API compatibility policy for a possible stability and change management approach.
Once your API is declared production-ready and stable, do not make backwards incompatible changes within that API version. If you need to make backwards-incompatible changes, create a new API with an incremented version number.