Thoughts on RESTful API design

Release 1.0

Geert Jansen

July 27, 2016

Contents

Lessons learnt from designing the Red Hat Enterprise Virtualization API
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 The Job of the API Designer . . . . . . . . . . . . . . . . . . . . . . .
1.3 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 URLs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7 Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.8 Miscellaneous Topics . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

1
1
1
3
8
10
13
14
19

ii

CHAPTER 1

Lessons learnt from designing the Red Hat Enterprise Virtualization

API

Author: Geert Jansen <[email protected]>

Date: November 15th, 2012
This work is licensed under a Creative Commons Attribution 3.0 Unported License.
The source of this document can be found on github.

1.1 Introduction
This essay is an attempt to put down my thoughts on how to design a real-world yet beautiful RESTful API. It
draws from the experience I have gained being involved in the design of the RESTful API for Red Hats Enterprise
Virtualization product, twice. During the design phase of the API we had to solve many of the real-world problems
described above, but we werent willing to add non-RESTful or RPC-like interfaces to our API too easily.
In my definition, a real-world RESTful API is an API that provides answers to questions that you wont find in
introductory texts, but that inevitably surface in the real world, such as whether or not resources should be described
formally, how to create useful and automatic command-line interfaces, how to do polling, asynchronous and other
non-standard types of requests, and how to deal with operations that have no good RESTful mapping.
A beautiful RESTful API on the other hand is one that does not deviate from the principles of RESTful architecture
style too easily. One important design element for example that is not always addressed is the possibility for complete
auto-discovery by the API user, so that the API can be used by a human with a web browser, without any reference to
external documentation. I will discuss this issue in detail in Forms.
This essay does not attempt to explain the basics about REST, as there are many other texts about that. For a good
description of REST, I would recommend reading at least the Atom Publishing Protocol and this blog post by Roy
Fielding.
While the views exposed in this essay are my own, they are heavily based on the public discussion on the rhevmapi mailing list. My thanks goes to all people who have contributed and are still contributing to it, including Mark
McLoughlin, Michael Pasternak, Ori Liel, Sander Hoentjen, Ewoud Kohl van Wijngaarden, Tomas V.V. Cox and
everybody else I forgot.

1.2 The Job of the API Designer

Before we dive fully into RESTful API design, it makes sense to look in a little more detail what the job of a RESTful
API designer is.

Thoughts on RESTful API design, Release 1.0

APIs dont exist in isolation. Instead, APIs expose functionality of an application or service that exists independently
of the API. In my view, the job of the RESTful API designer is two-fold.
1. Understanding enough of the important details of the application for which an API is to be created, so that an
informed decision can be made as to what functionality needs to be exposed, how it needs to be exposed, and
what functionality can instead be left out.
2. Modeling this functionality in an API that addresses all use cases that come up in the real world, following the
RESTful principles as closely as possible.
There are three distinct components involved in RESTful API design: the application, the API code, and the client.
The image above illustrates how these three components interact.

1.2.1 The Application

The application for which an API is to be provided exists independently of that API. Maybe the application has a GUI,
and you have a requirement to add a programmatic interface to it. Or maybe the application was designed under the
assumption that it would be accessed only via the API you are designing.
Like any application, the application for which the API is to be created contains state. That state is dynamic, and will
change due to various operations that are executed on it. This state, and the operations on it, need to be modeled and
exposed, and will form the API you are designing.
The easiest way to think about the application state, is to assume that it is is described by an application data model,
which can be described by an Entity-Relationship diagram. The ER-Diagram lists the details of the entities (I will call
them objects) in the application state, and the relationships between them.
In some cases, it is very easy to create the ER diagram. For example, in case of a web application that stores all its
state in a database, it can be trivially created from the schema of the database. In other cases where there is not such
a precise definition, the job of the API designer becomes a little more difficult. In such a case it would actually make
sense to create an ER diagram for the application in question. That is a useful exercise in its own right, as it will
make you better understand the application youre designing the API for. But more importantly, it will also help you
in designing a good RESTful API. We will talk more about that in a minute. Going forward, I will assume that an ER
diagram is available.
In addition to understanding the application data model, and the operations on it, you of course need an entry point into
the application that allows you to access and change the application state. This way in is fully application dependent
and could take many forms. We will call this way in the application interface. Formally, this application interface
could be considered an API as well. The difference, though, is that this interface is usually not intended for external

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

consumption or even fully documented. In order not to introduce any confusion, we will not refer to this interface as
an API: that term will be reserved for the RESTful API that is being designed.

1.2.2 The API Code

The job of the API code is to access the application state, as well as the operations on that state, via the application
interface, and expose it as a RESTful API. In between the application interface and the RESTful API, there is a
transformation step that adapts the application data model and makes it comply with the RESTful architecture style.
The result of this transformation would be RESTful resources, operations on those resources, and relationships between the resources. All of these are described by what we call the RESTful resource model.
Resources are the foundation behind any RESTful API, and we will go into a lot of detail on them in Resources. For
now, just think of resources as being very similar to the entities from the ER diagram (which is why I encouraged you
to create an ER diagram for your application in case it didnt exist).
Relationships between resources are expressed as hyperlinks in the representation of the resource. This is one of the
fundamental principles of RESTful API design. Resources also respond to a very limited set of operations (usually
just 4), which is a second principle of the RESTful architectural style.
When transforming objects from the application data model to RESTful resources, you may find it useful to define
two utility functions:
to_resource()
This function is assumed to take an object from the application data model, and convert it into a resource.
from_resource()
This function is assumed to take a resource, and translate it into an object in the application data model.
We dont discuss these methods further, other than mentioning that these can be rather simple functions if the application data model is similar to the resource model youre exposing, and can be quite complicated if they are very
different.

1.2.3 The Client

The client consumes the RESTful API via the standard HTTP protocol. In theory, the service could be provided on
top of other protocols as well. Since HTTP is so ubiquitous, however, it is not certain how valuable such a mapping to
another protocol would be in the real world. Therefore, this essay limits itself to describing our RESTful protocol in
terms of HTTP.
Clients would typically use an HTTP library to access the RESTful API. HTTP has become a moderately complex
protocol, and very good libraries exist for many target platforms / languages. It therefore makes a lot of sense to use
one of those libraries.
In some cases, it may make sense to use a generic REST library on top of an HTTP library. However, since there are so
many different conventions in RESTful APIs, this library may actually be specific to the API that you are consuming.

1.3 Resources
The fundamental concept in any RESTful API is the resource. A resource is an object with a type, associated data,
relationships to other resources, and a set of methods that operate on it. It is similar to an object instance in an objectoriented programming language, with the important difference that only a few standard methods are defined for the
resource (corresponding to the standard HTTP GET, POST, PUT and DELETE methods), while an object instance
typically has many methods.

1.3. Resources

Thoughts on RESTful API design, Release 1.0

Resources can be grouped into collections. Each collection is homogeneous so that it contains only one type of
resource, and unordered. Resources can also exist outside any collection. In this case, we refer to these resources as
singleton resources. Collections are themselves resources as well.
Collections can exist globally, at the top level of an API, but can also be contained inside a single resource. In the
latter case, we refer to these collections as sub-collections. Sub-collections are usually used to express some kind of
contained in relationship. We go into more detail on this in Relationships.
The diagram below illustrates the key concepts in a RESTful API.

We call information that describes available resources types, their behavior, and their relationships the resource model
of an API. The resource model can be viewed as the RESTful mapping of the application data model.

1.3.1 Resource Data

Resources have data associated with them. The richness of data that can be associated with a resource is part of the
resource model for an API. It defines for example the available data types and their behavior.
Based on my experience, I have developed a strong conviction that the JSON data model has just the right richness
so that it is an ideal data model for RESTful resources. I would recommend that everybody use it.
In JSON, just three types of data exist:
scalar (number, string, boolean, null).
array
object
Scalar types have just a single value. Arrays contain an ordered list of values of arbitrary type. Objects consist of a
unordered set of key:value pairs (also called attributes, not to be confused with XML attributes), where the key is a
string and the value can have an arbitrary type. For more detailed information on JSON, see the JSON web site.
Why the strong preference for JSON? In my view, JSON has the right balance between expressiveness, and broad
availability. The three types of data (scalars, arrays and objects) are powerful enough to describe in a natural way
virtually all the data that you might want to expose as resource, and at the same time these types are minimal enough
so that almost any modern language has built-in support for them.

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

XML would be the other obvious contender. Actually, in the final incarnation of the RHEV-M API, XML is used to
describe resources, via an XMLSchema definition. With hindsight, I believe that the XML data model is a bad choice
for a RESTful API. On one side, it is too rich, and on the other side, it lacks features. XML, as an SGML off-shoot, is
in my view great for representing structured documents, but not for representing structured data.
Features of XML that are too rich include:
1. Attributes vs elements. An XML element can have both attributes as well as sub-elements. A data item associated with a resource could be encoded in either one, and it would not be clear beforehand which one a client or
a server should use.
2. Relevance of order. The order between child-elements is relevant in XML. It is not natural in my view for object
attributes to have ordering.
The limitations of the XML data model are:
1. Lack of types. Elements in XML documents do not have types, and in order to use types, one would have to use
e.g. XMLSchema. XMLSchema unfortunately is a strong contender for the most convoluted specification ever
written.
2. Lack of lists. XML cannot natively express lists. This can lead to issues whereby it is not clear whether a certain
element is supposed to be a list or an object, and where that element ends up being both.
Application Data
We define the data that can be associated with a resource in terms of the JSON data model, using the following mapping
rules:
1. Resources are modeled as a JSON object. The type of the resource is stored under the special key:value pair
_type.
2. Data associated with a resource is modeled as key:value pairs on the JSON object. To prevent naming conflicts
with internal key:value pairs, keys must not start with _.
3. The values of key:value pairs use any of the native JSON data types of string, number, boolean, null, or arrays
thereof. Values can also be objects, in which case they are modeling nested resources.
4. Collections are modeled as an array of objects.
We will also refer to key:value pairs as attributes of the JSON object, and we will be sloppy and use that same term
for data items associated with resources, too. This use of attributes is not to be confused with XML attributes.
REST Metadata
In addition to exposing application data, resources also include other information that is specific to the RESTful API.
Such information includes URLs and relationships.
The following table lists generic attributes that are defined and have a specific meaning on all resources. They should
not be used for mapping application model attributes.
Attribute
id
href
link

Type

Meaning

String
String
Object

Identifies the unique ID of a resource.

Identifies the URL of the current resource.
Identifies a relationship for a resource. This attribute is itself an object and has rel href
attributes.

1.3. Resources

Thoughts on RESTful API design, Release 1.0

Other Data
Apart from application data, and REST metadata, sometimes other data is required as well. This is usually RPC like
data where a setting is needed for an operation, and where that setting will not end up being part of the resource itself.
One example that I can give here is where a resource creation needs a reference to another resource that is used during
the creation, but where that other resource will not become part of the resource itself.
It is the responsibility of the API code to merge the application data together with the REST metadata and the other
data into a single resource, resolving possible naming conflicts that may arise.

1.3.2 Representations
We have defined resources, and defined the data associated with them in terms of the JSON data model. However,
these resources are still abstract entities. Before they can be communicated to a client over an HTTP connection, they
need to be serialized to a textual representation. This representation can then be included as an entity in an HTTP
message body.
The following representations are common for resources. The table also lists the appropriate content-type to use:
Type
JSON

Content-Type
application/x-resource+json
application/x-collection+json

YAML
application/x-resource+yaml
application/x-collection+yaml
XML
application/x-resource+xml
application/x-collection+xml
HTML

text/html

Note: all these content types use the x- experimental prefix that is allowed by RFC2046.
JSON Format
Formatting a resource to JSON is trivial because the data model of a resource is defined in terms of the JSON model.
Below we give an example of a JSON serialization of a virtual machine:
{
"_type": "vm",
"name": "A virtual machine",
"memory": 1024,
"cpu": {
"cores": 4,
"speed": 3600
},
"boot": {
"devices": ["cdrom", "harddisk"]
}
}

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

YAML Format
Formatting to YAML is only slightly different than representing a resource in JSON. The resource type that is stored
under the _type key/value pair is serialized as a YAML !type annotation instead. The same virtual machine as
above, now in YAML format:
!vm
name: A virtual machine
memory: 1024
cpu:
cores: 4
speed: 3600
boot:
devices:
- cdrom
- harddisk

XML Format
XML is the most complex representation format due to both its complexity as well as its limitations. I recommend the
following rules:
Resources are mapped to XML elements with a tag name equal to the resource type.
Attributes of resources are mapped to XML child elements with the tag name equal to the attribute name.
Scalar values are stored as text nodes. A special type attribute on the containing element should be used to
refer to an XML Schema Part 2 type definition.
Lists should be stored as a single container element with child elements for each list item. The tag of the
container element should be the English plural of the attribute name. The item tag should be the English singular
of the attribute name. Lists should have the xd:list type annotation.
The same VM again, now in XML:
<vm xmlns:xs="http://www.w3.org/2001/XMLSchema">
<name type="xs:string">My VM</name>
<memory type="xs:int">1024</memory>
<cpu>
<cores type="xs:int">4</cores>
<speed type="xs:int">3600</speed>
</cpu>
<boot>
<devices type="xs:list">
<device type="xs:string">cdrom</device>
<device type="xs:string">harddisk</device>
</devices>
</boot>
</vm>

HTML Format
The exact format of a HTML response can be API dependent. HTML is for human consumption, and the only requirement is therefore that it be easy to understand. A simple implementation may choose the following representation:
For a collection, a <table> with a column per attribute where each object is listed in a separate row.

1.3. Resources

Thoughts on RESTful API design, Release 1.0

For a resource, a <table> with two columns, one with all the attribute names, one with the corresponding attribute
value.

1.3.3 Content-Types
As can be seen above, I am advocating the use of a generic content types application/x-resource+format and
application/x-collection+format. In my view this represents the right middle ground between two extremes that
are commonly found in RESTful APIs:
Some RESTful APIs only use the bare XML, JSON or YAML content types. An example of such as API is the Red
Hat Enterprise Virtualization API. In this case, the content type expresses nothing but the fact that an entity is in XML,
JSON or YAML format. In my view, this is not sufficient. Resources and collections have some specific semantics
around for example the use of href attributes, link attributes, and types. Therefore, these are a specialization of
XML, JSON and YAML and should be defined as such.
Other RESTful APIs define a content-type for every resource type that exists in the resource model. Examples of such
APIs include for example VMwares vSphere Director API. In my view, this is not proper either. Specifying detailed
content types invites both the API implementer, as well as a client implementer to think about these types as having
specific interfaces. In my view though, all resources should share the same basic interface, which is defined by the
RESTful design principles and expressed by the application/x-resource content type.
One reason that is sometimes given in favor of defining detailed content types is that this way, the content type can
be associated with a specific definition in some type definition language (such as XMLSchema). This, supposedly,
facilitates client auto-discovery because a client can know available attributes for a certain type. I go into a lot of detail
on this topic in Forms but the summary is that I do not agree with this argument.
Selecting a Representation Format
Clients can express their preference for a certain representation format using the HTTP Accept header. The HTTP
RFC defines an elaborate set of rules in which multiple formats can be requested, each with its own priority. In the
following example, the client tells the API that it accepts only YAML input:
GET /api/collection
Accept: application/x-collection+yaml

1.4 URLs
1.4.1 Entry Point
A RESTful API needs to have one and exactly one entry point. The URL of the entry point needs to be communicated
to API clients so that they can find the API.
Technically speaking, the entry point can be seen as a singleton resource that exists outside any collection. It is
common for the entry point to contain some or all of the following information:
Information on API version, supported features, etc.
A list of top-level collections.
A list of singleton resources.
Any other information that the API designer deemed useful, for example a small summary of operating status,
statistics, etc.

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

1.4.2 URL Structure

Each collection and resource in the API has its own URL. URLs should never be constructed by an API client. Instead,
the client should only follow links that are generated by the API itself.
The recommended convention for URLs is to use alternate collection / resource path segments, relative to the API
entry point. This is best described by example. The table below uses the :name URL variable style from Rails
Routes implementation.
URL
/api
/api/:coll
/api/:coll/:id
/api/:coll/:id/:subcoll
/api/:coll/:id/:subcoll/:subid

Description
The API entry point
A top-level collection named coll
The resource id inside collection coll
Sub-collection subcoll under resource id
The resource subid inside subcoll

Even though sub-collections may be arbitrarily nested, in my experience, you want to keep the depth limited to 2, if
possible. Longer URLs are more difficult to work with when using simple command-line tools like curl.

1.4.3 Relative vs Absolute

It is strongly recommended that the URLs generated by the API should be absolute URLs.
The reason is mostly for ease of use by the client, so that it never has to find out the correct base URI for a resource,
which is needed to interpret a relative URL. The URL RFC specifies the algorithm for determining a base URL, which
is rather complex. One of the options for finding the base URL is to use the URL of the resource that was requested.
Since a resource may appear under multiple URLs (for example, as part of a collection, or stand-alone), it would be a
significant overhead to a client to remember where he retrieved a representation from. By using absolute URLs, this
problem doesnt present itself.

1.4.4 URL Templates

A draft standard for URL templates exists. URL templates can be useful in link attributes where the target URL takes
query arguments. It is recommended though to make only conservative use of these. So far the only good use case I
have come across is when searching in a collection. In that case, it seems fair that the search criteria can be specified
as GET style query arguments appended to the collection URL.
I would recommend to only use the literal expansion part of the URL templates spec, because the expression
expansion part adds too much complexity to a client in my view, for very little benefit.

1.4.5 Variants
Sometimes you need to work on a variant of a resource. In our RHEV-M API for example, some attributes on a virtual
machine can be updated while a virtual machine is running. This amounts to a hot plug/unplug of the resource, which
is a different operation altogether from changing the saved representation. A nice way to implement this is using
;variant identifier in a URL. For example:
URL
/api/:coll/:id;saved
/api/:coll/:id;current

Description
Identifies the saved variant of a resource.
Identifies the current variant of a resource.

The use of the semicolon to provide options that are specific to a path segment is explicitly alowed in RFC3986 The
advantage over using a ?variant query argument is that this format is specific to a path segment only, and would
allow you to e.g. work on the saved variant of a sub-resource in a sub-collection of the current variant of another
resource.
1.4. URLs

Thoughts on RESTful API design, Release 1.0

1.5 Methods
1.5.1 Standard Methods
We already discussed that resources are the fundamental concept in a RESTful API, and that each resource has its own
unique URL. Methods can be executed on resources via their URL.
The table below lists the standard methods that have a well-defined meaning for all resources and collections.
Method
GET
GET
HEAD
HEAD
POST
PUT
PATCH
DELETE
OPTIONS

Scope
collection
resource
collection
resource
collection
resource
resource
resource
any

Semantics
Retrieve all resources in a collection
Retrieve a single resource
Retrieve all resources in a collection (header only)
Retrieve a single resource (header only)
Create a new resource in a collection
Update a resource
Update a resource
Delete a resource
Return available HTTP methods and other options

Normally, not all resources and collections implement all methods. There are two ways to find out which methods are
accepted by a resource or collection.
1. Use the OPTIONS method on the URL, and look at the Allow header that is returned. This header contains a
comma-separated list of methods are are supported for the resource or collection.
2. Just issue the method you want to issue, but be prepared for a 405 Method Not Allowed response that indicates
the method is not accepted for this resource or collection.

1.5.2 Actions
Sometimes, it is required to expose an operation in the API that inherently is non RESTful. One example of such an
operation is where you want to introduce a state change for a resource, but there are multiple ways in which the same
final state can be achieved, and those ways actually differ in a significant but non-observable side-effect. Some may
say such transitions are bad API design, but not having to model all state can greatly simplify an API. A great example
of this is the difference between a power off and a shutdown of a virtual machine. Both will lead to a vm resource
in the DOWN state. However, these operations are quite different.
As a solution to such non-RESTful operations, an actions sub-collection can be used on a resource. Actions are
basically RPC-like messages to a resource to perform a certain operation. The actions sub-collection can be seen as
a command queue to which new action can be POSTed, that are then executed by the API. Each action resource that is
POSTed, should have a type attribute that indicates the type of action to be performed, and can have arbitrary other
attributes that parameterize the operation.
It should be noted that actions should only be used as an exception, when theres a good reason that an operation
cannot be mapped to one of the standard RESTful methods. If an API has too many actions, then thats an indication
that either it was designed with an RPC viewpoint rather than using RESTful principles, or that the API in question is
naturally a better fit for an RPC type model.

1.5.3 PATCH vs PUT

The HTTP RFC specifies that PUT must take a full new resource representation as the request entity. This means that
if for example only certain attributes are provided, those should be removed (i.e. set to null).

10

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

An additional method called PATCH has been proposed recently. The semantics of this call are like PUT inthat it
updates a resource, but unlike PUT, it applies a delta rather than replacing the entire resource. At the time of writing,
PATCH was still a proposed standard waiting final approval.
For simple resource representations, the difference is often not important, and many APIs simply implement PUT as
a synonym for PATCH. This usually doesnt give any problems because it is not very common that you need to set an
attribute to null, and if you need to, you can always explicitly include it.
However for more complex representations, especially including lists, it becomes very important to be able to express
accurately the changes you want to make. Therefore, it is my recommendation now to both provide PATCH and PUT,
and make PATCH do an relative update and have PUT replace the entire resource.
It is important to realize that the request entity to PATCH is of a different content-type than the entity that it is
modifying. Instead of being a full resource, it is a resource that describes modifications to be made to a resource. For a
JSON data model, which is what this essay is advocating, I believe that there are two sensible ways to define the patch
format.
1. An informal approach where you accept a dict with a partial representation of the object. Only attributes that
are present are updated. Attributes that are not present are left alone. This approach is simple, but it has the
drawback that if the resource has a complex internal structure e.g. containing a big list of dicts, then that entire
list of dicts need to be given in the entity. Effectively PATCH becomes similar to PUT again.
2. A more formal approach would be to accept a list of modifications. Each modification can be a dict specifying
the JSON path of the node to modify, the modification (add, remove, change) and the new value.

1.5.4 Link Headers

An Internet draft exists defining the Link: header type. This header takes the link attributes of the response entity,
and formats them as an HTTP header. It is argued that the Link header is useful because it allows a client to quickly
get links from a response without having to parse the response entity, or even without retrieving the response entity at
all using the HEAD HTTP method.
In my view, the usefulness of this feature is dubious. First of all, it can increase response sizes quite significantly.
Second, it can only be used when a resource is being returned; it does not make sense to be used with collections.
Because I have not yet seen any good use of this header in the context of a RESTful API, I recommend not to implement
Link headers.

1.5.5 Asynchronous Requests

Sometimes an action takes too long to be completed in the context of a single HTTP request. In that case, a 202
Accepted status can be returned to the client. Such a response should only be returned for POST, PUT, PATCH or
DELETE.
The response entity of a 202 Accepted response should be a regular resource with only the information filled in that
was available at the time the request was accepted. The resource should contain a link attribute that points to a status
monitor that can be polled to get updated status information.
When polling the status monitor, it should return a response object with information on the current status of the
asynchronous request. If the request is still in progress, such a response could look like this (in YAML):
!response
status: 202 Accepted
progress: 50%

If the call has finished, the response should include the same headers and response body had the request been fulfilled
synchronously:

1.5. Methods

11

Thoughts on RESTful API design, Release 1.0

!response
status: 201 Created
headers:
- name: content-type
value: applicaton/x-resource+yaml
response: !!str
Response goes here

After the response has been retrieved once with a status that is not equal to 202 Accepted, the API code may garbage
collect it and therefore clients should not assume it will continue to be available.
A client may request the server to modify its asynchronous behavior with the following Expect headers:
Expect: 200-ok/201-created/204-no-content disables all asynchronous functionality. The server may return a
417 Expectation Failed if it is not willing to wait for an operation to complete.
Expect: 202-accepted explicitly request an asynchronous response. The server may return a 417 Expectation
Failed if it is not willing to perform the request asynchronously.
If no expectation is provided, client must be prepared to accept a 202 Accepted status for any request other than GET.

1.5.6 Ranges / Pagination

When collections contain many resources, it is quite a common requirement for a client to retrieve only a subset of the
available resources. This can be implemented using the Range header with a resource range unit:
GET /api/collection
Range: resources=100-199

The above example would return resources 100 through 199 (inclusive).
Note that it is the responsibility of the API implementer to ensure a proper and preferably meaningful ordering can be
guaranteed for the resources.
Servers should provide an Accept-Ranges: resource header to indicate to a client that they support resource-based
range queries. This header should be provided in an OPTIONS response:
OPTIONS /api/collection HTTP/1.1
HTTP/1.1 200 OK
Accept-Ranges: resources

1.5.7 Notifications
Another common requirement is where a client wants to be notified immediately when some kind of event happens.
Ideally, such a notification would be implemented using a call-out from the server to the client. However, there is no
good portable standard to do this over HTTP, and it also breaks with network address translation and HTTP proxies.
A second approach called busy-loop polling is horribly inefficient.
In my view, the best approach is what is is called long polling. In long polling, the client will retrieve a URL but the
server will not generate a response yet. The client will wait for a configurable amount of time, until it will close the
connection and reconnect. If the server becomes aware of an event that requires notification of clients, it can provide
that event immediately to clients that are currently waiting.
Long polling should be disabled by default, and can be enabled by a client using an Expect header. For example, a
client could long poll for new resources in a collection using a combination of long-polling and a resource-based range
query:

12

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

GET /api/collection
Range: 100Expect: nonempty-response

In this case, resource 100 would be the last resource that was read, and the call is requesting the API to return at
least one resource with an ID > 100.
Server implementers need to decide whether they want to implement long polling using one thread per waiting client,
or one thread that uses multiplexed IO to wait for all clients. This is a trade-off to be made between ease of implementation and scalability (that said, threads are pretty cheap on modern operating systems).

1.6 Relationships
As we have seen in Resources, the resource is the fundamental unit in RESTful API design. Resources model objects
from the application data model.
Resources do not exist in isolation, but have relationships to other other resources. Sometimes these relationships
exist between the mapped objects in the application data model as well, sometimes they are specific to the RESTful
resources.
One of the principles of the RESTful architecture style is that these relationships are expressed by hyperlinks to the
representation of a resource.
In our resource model, we interpret any object with an href attribute as a hyperlink. The value of the href attribute
contains an absolute URL that can be retrieved with GET. Using GET on a such a URL is guaranteed to be side-effect
free.
Two types of hyperlinks are normally used:
1. Using link objects. This is a special object of type link with href and a rel attributes. The rel attribute
value identifies the semantics of the relationship. The link object is borrowed from HTML.
2. Using any object with a href attribute. The object type defines the semantics of the relationship. Ill call these
object links, not to be confused with the link objects above.
Below is an example of a virtual machine representation in YAML that illustrates the two different ways in which a
relationship can be expressed:
!vm
id: 123
href: /api/vms/123
link:
rel: collection/nics
href: /api/vms/123/nics
cluster:
href: /api/clusters/456

The link object is used with rel=collection/nics to refer to a sub-collection to the VM holding the virtual network
interfaces. The cluster link instead points to the cluster containing this VM by means of a cluster object.
Note that the VM itself has a href attribute as well. This is called a self link, and it is a convention for a resource to
provide such a link. That way, the object can later be easily re-retrieved or updated without keeping external state.
There are no absolute rules when to use link objects, and when not. That said, we have been using the rules below
in the rhevm-api project with good success. They seem to be a good trade-off between compactness (favoring object
links) and the ability to understand links without understanding the specific resource model (favoring link objects).
Link objects are used to express structural relationships in the API. So for example, the top-level collections,
singleton resources and sub-collections (including actions) are all referenced using link objects.

1.6. Relationships

13

Thoughts on RESTful API design, Release 1.0

Object links are used to express semantic relationships from the application data model. In the example above,
the vm to cluster link comes directly from the application data model and is therefore modeled as a link.
Note however that sub-collections can express both a semantic relationship, as well as a structural relationship. See
for example the collection/nics sub-collection to a VM in the example above. In such a case, our convention has
been to use link objects.

1.6.1 Standard Structural Relationships

It makes sense to define a few standard structural relationship names that can be relevant to many different APIs. In
this way, client authors know what to expect and can write to some extent portable clients that work with multiple
APIs. The table below contains a few of these relationship names.
Relationship
collection/{name}
resource/{name}
form/{name}

Semantics
Link to a related collection {name}.
Link to a related resource {name}.
Link to a related form {name}.

Here, {name} is replaced with a term that has meaning for the specific API. For example, a collection of virtual
machines could be linked from the entry point of a virtualization API using rel=collection/vms.

1.6.2 Modeling Semantic Relationships

Semantic relationships can be modeled either by an object link, or by a sub-collection. I believe that choosing the right
way to represent a sub-collection is important to get a consistent API experience for the client. I would advocate using
the following rules:
1. In case of a 1:N relationship, where the target object is existentially dependent on the source object, Id recommend to use a sub-collection. By existentially dependent I mean that a target object cannot exist without its
source. In database parlance, this would be a FOREIGN KEY relationship with an ON DELETE CASCADE.
An example of such a relationship are the NICs that are associated with a VM.
2. In case of a 1:N relationship, where there is data that is associated with the link, Id recommend to use a subcollection. Note that we are talking about data here that is neither part of the source object, nor the target object.
The resources in the sub-collection can hold the extra data. In this case, the data in the sub-resource would
therefore be a merge of the data from the mapped object from the application data model, and link data.
3. In case of any other 1:N relationship, Id recommend to use an object link.
4. In case of a N:M relationship, Id recommend to use a sub-collection. The sub-collection can be defined initially
to support the lookup direction that you most often need to follow. If you need to query both sides of the
relationship often, then two sub-collections can be defined, one on each linked object.

1.7 Forms
In this section, a solution is proposed to a usability issue that is present in a lot of RESTful APIs today. This issue is
that, without referring to external documentation, an API user does not know what data to provide to operations that
take input. For example, the POST call is used to create a new resource in a collection, and an API user can find out
with the OPTIONS call that a certain collection supports POST. However, he does not know what data is required to
create the new resource. So far, most RESTful APIs require the user to refer to the API documentation to get that
information. This violates an important RESTful principle that APIs should be self-descriptive.
As far as I can see, this usability issue impacts two important use cases:

14

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

1. A person is browsing the API, trying to use and learn it as he goes. The person can either be using a web
browser, or a command-line tool like curl.
2. The development of a command-line or graphical interface to a RESTful API. Without a proper description of
required input types, knowledge about these has to be encoded explicitly in the client. This has the disadvantage that new features are not automatically exposed, and that there are strict requirements around not making
backwards incompatible changes.
When we look at the web, this issue doesnt exist. Millions of users interact with all kinds of IT systems over the web
at any given moment, and none of them has to resort to API documentation to understand how to do that. Interactions
with IT systems on the web are all self-explanatory using hypertext, and input is guided by forms.
In fact, forms are what I see as the solution here. But before I go into that, lets look into two classes of solutions
that have been proposed to address this issue, and why I think they actually do not solve it. The first is using a type
definition language like XMLSchema to describe the input types, the second is using some kind of service description
language like WADL.

1.7.1 Type Definition

As alluded to in Resources, some RESTful APIs use a type definition language to provide information about how to
construct request entities. In my view this is a bad approach, and has the following issues:
1. It creates a tight coupling between servers and clients.
2. It still does not allow a plain web browser as an HTTP client.
3. It does not help in the automatic construction of CLIs.
Often, XMLSchema is used as the type definition language. This brings in a few more issues:
1. XMLSchema is a horribly complicated specification on one side, but on the other side it cant properly solve
the very common requirement of an unordered set of key:value pairs (Note that <xs:all> does not solve this. Its
limitations make it almost useless.).
2. Type definitions in XMLSchema are context-free. This means that you cannot just define one type, you need
to define one type for PUT, one for POST, etc. And since one XML element can only have one type, this
means youre PUTing a different element that youd POST or GET. This breaks that homogeneous collection
principle of REST.

1.7.2 Service Definition

WADL is an approach to define a service description language for RESTful APIs. It tries to solve the problem by
meticulously defining entry points and parameters to those. The problem I have with WADL, is that it feels very
non-RESTful. I cant see a lot of difference between a method on a WADL resource, and an RPC entry point.
It would be possible to construct a more RESTful service description language. It would focus on mapping the
RESTful concepts of resource, collection, relationships, and links. It would probably need a type definition language
as well to define the constraints on types for each method (again, PUT may have different constraints than POST).
There have been discussions in the RHEV-M project on this.
In the end though, this approach also feels wrong. What I think is needed is something that works like the web, where
everything is self descriptive, and guided only by the actual URL flow the user is going through, not by reference to
some external description.

1.7. Forms

15

Thoughts on RESTful API design, Release 1.0

1.7.3 Using Forms to Guide Input

The right solution to guide client input in my view is to use forms. Note that Im using the word forms here in a
generic sense, as an entity that guides user input, and not necessarily equal to an HTML form.
In my view, forms need to specify three pieces of information:
1. How to contact the target and format the input.
2. A list of all available input fields.
3. A list of constraints which which the input fields must comply.
HTML forms provide for #1 and #2, but not for #3 above. #3 is typically achieved using client-side Javascript. Because
Javascript is not normally available to clients that use the API, we need to define another way to do validation. The
approach I propose here is to define a form definition language that captures the information #1 through #3. A form
expressed in this language can be sent over to the client as-is, in case the client understands the form language. It can
also be translated into an HTML form with Javascript in case the client is a web browser.
Form Definition Language
Our forms are defined as a special RESTful resource with type form. This means they use the JSON data model,
can be represented in JSON, YAML and XML according to the rules in Resources.
Forms have their own content type. This is listed in the table below:
Type
Form

Content-Type
application/x-form+json
application/x-form+yaml
application/x-form+xml

Rather than providing a formal definition, I will introduce the form definition language by using an example. Below
is an example of a form that guides the the creation of a virtual machine. In YAML format:
!form
method: POST
action: {url}
type: vm
fields:
- name: name
type: string
regex: [a-zA-Z0-9]{5,32}
- name: description
type: string
maxlen: 128
- name: memory
type: number
min: 512
max: 8192
- name: restart
type: boolean
- name: priority
type: number
min: 0
max: 100
constraints:

16

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

- sense: mandatory
field: name
- sense: optional
field: description
- sense: optional
field: cpu.cores
- sense: optional
field: cpu.sockets
- sense: optional
exclusive: true
constraints:
- sense: mandatory
field: highlyavailable
- sense: optional
field: priority

As can be seen the form consists of 3 parts: form metadata, field definitions, and constraints.
Form Metadata
The form metadata is very simple. The following attributes are defined:
Attribute
method
url
type

Description
The HTTP method to use. Can be GET, POST, PUT or DELETE.
The URL to submit this form to.
The type of the resource to submit.

Fields
The list of available fields are specified using the fields attribute. This should be a list of field definitions. Each field
definition has the following attributes:
Attribute
name
type
min
max
minlen
maxlen
regex
multiple

Description
The field name. Should be in dotted-name notation.
One of string, number or boolean
Field value must be greater than or equal to this (numbers)
Field value must be less than or equal to this (numbers)
Minimum field length (strings)
Maximum field length (strings)
Field value needs to match this regular expression (strings)
Boolean that indicates if multiple values are accepted (array).

Constraints
First we need to answer the question of what kind of constraints we want to express in our form definition language. I
will start by mentioning that in my view, it is impossible to express each and every constraint on the client side. Some
constraints for example require access to other data (e.g. when creating relationships), are computationally intensive,
or even unknown to the API designer because they are undocumented for the application the API is written for. So in
my view we need to find a good subset that is useful, without making it too complicated and without worrying about
the fact that some constraints possibly cannot not be expressed.
This leads me to define the following two kinds of constraints, which in my view are both useful, as well as sufficient
for our purposes:

1.7. Forms

17

Thoughts on RESTful API design, Release 1.0

1. Constraints on individual data values.

2. Presence constraints on fields, i.e. whether a field is allowed or not allowed, and if allowed, whether it is
mandatory.
The constraints on data values are useful because CLIs and GUIs could use this information to help a user input data.
For example, depending on the type, a GUI could render a certain field a checkbox, a text box, or a dropdown list.
For brevity, constraints on individual data values are defined as part of the field definition, and were discussed in the
previous section.
Presence constraints are also useful, as they allow an API user to generate a concise synopsis on how to call a certain
operation to be used e.g. in CLIs. Presence constraints specify which (combination of) input fields can and cannot
exist. Each presence constraint has the following attributes:
Attribute
sense
field
constraints
exclusive

Description
One of mandatory or optional
This constraint refers to a field.
This constraint is a group with nested constraints.
This is an exclusive group (groups only).

Either field or constraints has to be specified, but not both. A constraint that has the field attribute set is called a
simple constraint. A constraint that has the constraints attribute set, is called a group.
Checking Constraints
Value constraints should be checked first, and should be checked only on non-null values.
After value constraints, the presence constraints should to be checked. This is a bit more complicated because the
constraints are not only used for making sure that all mandatory fields exist, but also that no non-optional fields are
present. The following algorithm should be used:
1. Start with an empty list called referenced that will collect all referenced fields.
2. Walk over all constraints in order. If a constraint is a group, you need to recurse into it, depth first, passing it the
referenced list.
3. For every simple constraint that validates, add the field name to referenced.
4. In exclusive groups, matching stops at the first matching sub-constraint, in which case the group matches. In
non-exclusive groups, matching stops at the first non-matching sub-constraint, in which case the group does not
match.
5. When matching a group, you need to backtrack to the previous value of referenced in case the group does not
match.
6. A constraint only fails if it is mandatory and it is a top-level constraint. If a constraint fails, processing may stop.
7. When youve walked through all constraints, it is an error if there are fields that have a non-null value but are
not in the referenced list.
Building the Request Entity
After all constraints have been satisfied, a client should build a request entity that it will pass in the body of the POST,
PUT or DELETE method. In case the form was requested in JSON, YAML or XML format, it is assumed that the
client is not a web browser, and the following applies:
First, a new resource is created of the type specified in the form metadata. Dotted field names should be interpreted
as follows. Each dot creates a new object, and stores it under the name immediately left of the dot in its parent object.
This means that the parent must be an object as well, which means it cannot correspond to a field definition with

18

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Thoughts on RESTful API design, Release 1.0

multiple set to true (which would make it a list). This resource is then represented in a format that the server
supports, using the rules described in Resources.
If a client requested a text/html representation of the form, it is assumed that the client is a web browser, and we
assume the form will be processed as a regular HTML form. In this case:
An HTML <form> should be generated, with an appropriate <input> element for each field.
The HTML forms method attribute should be set to POST, unconditionally. In case the RESTful forms
method is not POST, the server should include a hidden input element with the name _method to indicate to
the server the original method. (HTML does not support PUT or DELETE in a form).
The forms enctype should be set to multipart/form-data or application/x-www-form-urlencoded, as appropriate for the input elements.
A hidden field called _type is generated that contains the value of the type attribute in the form metadata.
The server may generate Javascript and include that in the HTML to check the value and presence constraints.

1.7.4 Linking to Forms

Forms are associated with resources and collections by link objects. The name of the link object defines the meaning
of the form. The following standard forms names are defined:
Name
form/search
form/create
form/update
form/delete

Scope
collection
collection
resource
resource

Description
Form to search for resources
Form to create a new resource
Form to update a resource
Form to delete a resource

1.8 Miscellaneous Topics

1.8.1 Offline Help
A requirement that sometimes comes up is that to have an offline description of the API that can be used to generate
offline help for a command-line interface (CLI) for that API. When connected to an API, all this information is of
course available because of the self-descriptive nature of a RESTful API. But the request is to have that information
offline as well, so that CLI users can get help without have to connect first.
In my view it is a bad idea to offer this offline help. It introduces a tight coupling between a client and a server, which
will break the RESTful model. A server cannot be independently upgraded anymore from a client, and clients become
tied to a specific server version. Also I doubt the usefulness of this feature as having to connect first to get help does
not seem like a big issue to me.
That said, it is relatively straightforward to define a way that such an offline description can be facilitated.
One server-side change is required for this. For each collection, the API should implement a placeholder resource with
a special ID (lets say _, but it doesnt matter as long as it cannot be a valid ID), and some fake data. When a special
HTTP Expect header is set, only this placeholder resource is returned when querying a collection.
Having the placeholder resources in place for every collection, the following procedure can be used by the client to
retrieve all relevant metadata:
1. Retrieve the entry point of the API and store it in memory as a resource.
2. For every hyperlink in the entry point, fetch the target, and store it under the target property under the link
object, recursively. When collections are retrieved, the special Expect header must be set.

1.8. Miscellaneous Topics

19

Thoughts on RESTful API design, Release 1.0

3. For every hyperlink in the entry point, execute an OPTIONS call on the target, and store the resulting headers
under the options property of the link object, recursively.
4. Serialize the resulting object to a file.
The result of this is basically a recursive dump of the GET-able part of an API with one placeholder resource in
every collection. It will contain all metadata including forms, options, available resource types and collections, in a
format that is very similar to the real on-line data. It can be used as an off-line store to generate any help that you
would be able to generate online as well.
An API can store a version number on its entry point, so that clients can see when their offline cached copy expired
and can therefore generate a new one.

20

Chapter 1. Lessons learnt from designing the Red Hat Enterprise Virtualization API

Index

F
from_resource() (built-in function), 3

T
to_resource() (built-in function), 3

21