22 KiB
Documenting REST API endpoints
This document explains the system for documenting Zulip's REST API.
Zulip's API documentation is an essential resource both for users and for the developers of Zulip's mobile and terminal apps. Our vision is for the documentation to be sufficiently good that developers of Zulip's apps should never need to look at the server's implementation to answer questions about the API's semantics.
To achieve these goals, Zulip leverages the popular OpenAPI format as the data source to ensure that Zulip's API documentation is correct and remains so as Zulip's API evolves.
In particular, the top goal for this system is that all mistakes in verifiable content (i.e. not the English explanations) should cause the Zulip test suite to fail. This is incredibly important, because once you notice one error in API documentation, you no longer trust it to be correct, which ends up wasting the time of its users.
Since it's very difficult to not make little mistakes when writing any untested code, the only good solution to this is a way to test the documentation. We found dozens of errors in the process of adding the validation Zulip has today.
Our API documentation is defined by a few sets of files:
- The primary source of our API documentation is the Zulip server's
OpenAPI description at
zerver/openapi/zulip.yaml
. - The documentation is written the same Markdown framework that powers
our help center docs, with some special
extensions for rendering nice code blocks and example
responses. Most API endpoints share a common template,
api_docs/api-doc-template.md
, which renders the OpenAPI description of the API endpoint. A handful of endpoints that require special content, as well as pages that document general API details rather than specific endpoints, live atapi_docs/*.md
. - We have an extensive set of tests designed to validate that the data
in the OpenAPI file matching the implementation. Specifically,
zerver/tests/test_openapi.py
compares every endpoint's accepted parameters inviews
code with those declared inzulip.yaml
. And the backend test suite checks that every API response served during our extensive backend test suite matches one the declared OpenAPI schema for that endpoint. - The text for the Python examples comes from a test suite for the
Python API documentation (
zerver/openapi/python_examples.py
; run viatools/test-api
). Thegenerate_code_example
macro will magically read content from that test suite and render it as the code example. This structure ensures that Zulip's API documentation is robust to a wide range of possible typos and other bugs in the API documentation. - The JavaScript examples are similarly generated and tested using
zerver/openapi/javascript_examples.js
. - The cURL examples are generated and tested using
zerver/openapi/curl_param_value_generators.py
. - The REST API index
(
api_docs/include/rest-endpoints.md
) in the broader /api left sidebar (api_docs/sidebar_index.md
).
This first section is focused on explaining how the API documentation system is put together; when actually documenting an endpoint, you'll want to also read the Step by step guide.
How it works
Let's use the existing documentation for one of our REST API endpoints to show how the documentation system works: POST /messages/render. We highly recommend looking at these resources while reading the above documentation page:
api_docs/api-doc-template.md
zerver/openapi/zulip.yaml
, specifically the section withoperationId: render-message
zerver/openapi/python_examples.py
If you look at the documentation for existing endpoints, you'll notice that a typical endpoint's documentation is divided into four sections:
- Title and description
- Usage examples
- Parameters
- Response with examples
The rest of this guide describes how each of these sections works.
Title and description
The first line of api-doc-template.md
generates a lot of key
information for our API endpoint documentation:
{generate_api_header(API_ENDPOINT_NAME)}
At the top of the endpoint documentation page is the title, and it
comes from the summary
parameter in the OpenAPI data,
zerver/openapi/zulip.yaml
.
The endpoint description
in the OpenAPI data explains what the
endpoint does in clear English. It should include details on how to
use the endpoint correctly or what it's good or bad for, with links
to any alternative endpoints the user might want to consider.
The description should often contain a link to the documentation of
the relevant feature in the help center, and should
include Changes notes for all feature level updates documented
in the API changelog, see
api_docs/changelog.md
, that reference the endpoint.
Endpoints that only administrators can use should be tagged with the
custom x-requires-administrator
field in the OpenAPI definition.
All of this information is rendered via a Markdown preprocessor,
specifically the APIHeaderPreprocessor
class defined in
zerver/openapi/markdown_extension.py
.
Usage examples
We display usage examples in three languages: Python, JavaScript and curl; we may add more in the future. Every endpoint should have Python and curl documentation; JavaScript is optional as we don't consider that API library to be fully supported.
The examples are defined using a special Markdown extension, see
zerver/openapi/markdown_extension.py
. Here's the Markdown file
block that uses this in api-doc-template.md
:
{start_tabs}
{generate_code_example(python)|API_ENDPOINT_NAME|example}
{generate_code_example(javascript)|API_ENDPOINT_NAME|example}
{tab|curl}
{generate_code_example(curl)|API_ENDPOINT_NAME|example}
{end_tabs}
In some cases, one wants to configure specific parameters to be
included or excluded from the example curl requests for readability
reasons. One can do that using the x-curl-examples-parameters
parameter in the OpenAPI data.
Writing Python examples
For the Python examples, you'll write the example in
zerver/openapi/python_examples.py
, and it'll be run and verified
automatically in Zulip's automated test suite. The code for our
example API endpoint looks like this:
@openapi_test_function('/messages/render:post')
def render_message(client: Client) -> None:
# {code_example|start}
# Render a message
request = {
'content': '**foo**'
}
result = client.render_message(request)
# {code_example|end}
validate_against_openapi_schema(result, '/messages/render', 'post', '200')
This is an actual Python function which will be run as part of the
tools/test-api
test suite. The validate_against_opanapi_schema
function will verify that the result of that request is as defined in
the examples in zerver/openapi/zulip.yaml
.
To run as part of the test suite, the render_message
function needs
to be called from test_messages
(or one of the other functions at
the bottom of the file). The final function, test_the_api
, is what
actually runs the tests. Tests with the openapi_test_function
decorator that are not called will fail tests, as will new endpoints
that are not covered by an openapi_test_function
-decorated test.
You will still want to manually test the example using Zulip's Python API client by copy-pasting from the website; it's easy to make typos and other mistakes where variables are defined outside the tested block, and the tests are not foolproof.
The code that renders API documentation pages will extract the block
between the # {code_example|start}
and # {code_example|end}
comments,
and substitute it in place of
{generate_code_example(python)|/messages/render:post|example}
. Note
that here the API_ENDPOINT_NAME
has been filled in with our example
endpoint's information.
Additional Python imports can be added using the custom
x-python-examples-extra-imports
field in the OpenAPI definition.
Parameters
We have a separate Markdown extension to document the parameters that
an API endpoint supports. Implemented in
zerver/lib/markdown/api_arguments_table_generator.py
, you can see
this in api-doc-template.md
after the Parameters header:
{generate_api_arguments_table|zulip.yaml|API_ENDPOINT_NAME}
This generates the information from the endpoint's parameter definition in the OpenAPI data.
Additional content that you'd like to appear in the parameter
description area can be declared using the custom
x-parameter-description
field in the OpenAPI definition.
Response with examples
Similar to the parameters section above, there is a separate Markdown
extension to document the endpoint's return values and generate the
example response(s) from the OpenAPI data. Implemented in
zerver/lib/markdown/api_return_values_table_generator.py
, you can
see this in after the Response header in api-doc-template.md
:
{generate_return_values_table|zulip.yaml|API_ENDPOINT_NAME}
To generate the example responses from the OpenAPI data, we again
use the special Markdown extension from the Usage examples
discussed above, except with the fixture
argument instead of the
example
argument:
{generate_code_example|API_ENDPOINT_NAME|fixture}
Additional content that you'd like to appear in the responses part of
the page can be added using the custom x-response-description
field
in the OpenAPI definition.
Step by step guide
This section offers a step-by-step process for adding documentation for a new API endpoint. It assumes you've read and understood the above.
-
Start by adding OpenAPI format data to
zerver/openapi/zulip.yaml
for the endpoint. If you copy-paste (which is helpful to get the indentation structure right), be sure to update all the content that you copied to correctly describe your endpoint!In order to do this, you need to figure out how the endpoint in question works by reading the code! To understand how arguments are specified in Zulip backend endpoints, read our REST API tutorial, paying special attention to the details of
typed_endpoint
.Once you understand that, the best way to determine the supported arguments for an API endpoint is to find the corresponding URL pattern in
zprojects/urls.py
, look up the backend function for that endpoint inzerver/views/
, and inspect its keyword-only arguments.You can check your formatting using these helpful tools.
tools/check-openapi
will verify the syntax ofzerver/openapi/zulip.yaml
.tools/test-backend zerver/tests/test_openapi.py
; this test compares your documentation against the code and can find many common mistakes in how arguments are declared.test-backend
: The full Zulip backend test suite will fail if any actual API responses generated by the tests don't match your defined OpenAPI schema. Usetest-backend --rerun
for a fast edit/refresh cycle when debugging.
-
Add a function for the endpoint you'd like to document to
zerver/openapi/python_examples.py
, decorated with@openapi_test_function
.render_message
is a good example to follow. There are generally two key pieces to your test: (1) doing an API query and (2) verifying its result has the expected format usingvalidate_against_openapi_schema
. -
Make the desired API call inside the function. If our Python bindings don't have a dedicated method for a specific API call, you may either use
client.call_endpoint
or add a dedicated function to the zulip PyPI package. Ultimately, the goal is for every endpoint to be documented the latter way, but it's useful to be able to write working documentation for an endpoint that isn't supported bypython-zulip-api
yet. -
Add the function to one of the
test_*
functions at the end ofzerver/openapi/python_examples.py
; this will ensure your function will be called when runningtest-api
. -
Capture the JSON response returned by the API call (the test "fixture"). The easiest way to do this is add an appropriate print statement (usually
json.dumps(result, indent=4, sort_keys=True)
), and then runtools/test-api
. You can also use https://jsonformatter.curiousconcept.com/ to format the JSON fixtures. Add the fixture to theexample
subsection of theresponses
section for the endpoint inzerver/openapi/zulip.yaml
. -
Run
./tools/test-api
to make sure your new test function is being run and the tests pass. -
Now, inside the function, isolate the lines of code that call the API and could be displayed as a code example. Wrap the relevant lines in
# {code_example|start} ... relevant lines go here ... # {code_example|end}
comments. The lines inside these comments are what will be displayed as the code example on our/api
page. -
Finally, if the API docs page of the endpoint doesn't follow the common API docs template in
api_docs/api-docs-template.md
, then add its custom Markdown file underapi_docs/
. However, it is a goal to minimize the number of files that diverse from the common template, so only do this if there's a good reason. -
Add the endpoint to the index in
api_docs/include/rest-endpoints.md
. The URL should match theoperationId
for the endpoint, and the link text should match the title of the endpoint from the OpenAPIsummary
field. -
Test your endpoint, pretending to be a new user in a hurry, by visiting it via the links on
http://localhost:9991/api/
(the API docs are rendered from the Markdown source files on page load, so just reload to see an updated version as you edit). You should make sure that copy-pasting the code in your examples works, and post an example of the output in the pull request. -
Document the new API in
api_docs/changelog.md
and bump theAPI_FEATURE_LEVEL
inversion.py
. Also, make sure to add a**Changes**
entry in the description of the new API/event inzerver/openapi/zulip.yaml
, which mentions the API feature level at which they were added.
Why a custom system?
Given that our documentation is written in large part using the OpenAPI format, why maintain a custom Markdown system for displaying it? There's several major benefits to this system:
- It is extremely common for API documentation to become out of date as an API evolves; this automated testing system helps make it possible for Zulip to maintain accurate documentation without a lot of manual management.
- Every Zulip server can host correct API documentation for its version, with the key variables (like the Zulip server URL) already pre-substituted for the user.
- We're able to share implementation language and visual styling with our Help Center, which is especially useful for the extensive non-REST API documentation pages (e.g., our bot framework).
Using the standard OpenAPI format gives us flexibility, though; if we later choose to migrate to third-party tools, we don't need to redo the actual documentation work in order to migrate tools.
Debugging schema validation errors
A common function used to validate and test Zulip's REST API is
validate_against_openapi_schema
. It is used to verify that every
successful API response returned in the backend and documentation test
suites are a documented possibility in the API documentation.
Therefore, when you add a new feature or setting to Zulip, you will most
likely need to update the API documentation (zerver/openapi/zulip.yaml
)
in order to pass existing tests that use this function. Additionally, if
you're writing documentation for a new or undocumented REST API endpoint,
you'll want to use this function to validate and test your changes in
zerver/openapi/python_examples.py
.
Below are some examples to help you when debugging the schema validation
errors produced by validate_against_openapi_schema
. Before reading
through the examples, we recommend reviewing the
OpenAPI configuration documentation if you're unfamiliar
with the format.
If you use Visual Studio Code, an OpenAPI extension can be very helpful in
navigating Zulip's large and detailed OpenAPI file; see
.vscode/extensions.json
.
Deconstructing the error output
To start with a clear example, let's imagine that we are writing the documentation for the REST API endpoint for uploading a file, POST /api/v1/user_uploads.
There are no parameters for this endpoint, and only one return value
specific to this endpoint, uri
, which is the URL of the uploaded file.
If we comment out that return value and example from the existing API
documentation in zerver/openapi/zulip.yaml
, for example:
/user_uploads:
post:
operationId: upload-file
...
responses:
"200":
description: Success.
content:
application/json:
schema:
allOf:
- $ref: "#/components/schemas/JsonSuccessBase"
- additionalProperties: false
properties:
result: {}
msg: {}
# uri:
# type: string
# description: |
# The URI of the uploaded file.
example:
{
"msg": "",
"result": "success",
# "uri": "/user_uploads/1/4e/m2A3MSqFnWRLUf9SaPzQ0Up_/zulip.txt",
}
We will now get an error when we run the API documentation test suite
in the development environment (tools/test-api
):
Running API tests...
2022-12-19 15:05:42.347 WARN [django.server] "POST /api/v1/users HTTP/1.1" 400 88
Waiting for children to stop...
Traceback (most recent call last):
File "tools/test-api", line 93, in <module>
test_the_api(client, nonadmin_client, owner_client)
File "/srv/zulip/zerver/openapi/python_examples.py", line 1636, in test_the_api
test_users(client, owner_client)
File "/srv/zulip/zerver/openapi/python_examples.py", line 1555, in test_users
upload_file(client)
File "/srv/zulip/zerver/openapi/python_examples.py", line 52, in _record_calls_wrapper
return test_func(*args, **kwargs)
File "/srv/zulip/zerver/openapi/python_examples.py", line 1284, in upload_file
validate_against_openapi_schema(result, "/user_uploads", "post", "200")
File "/srv/zulip/zerver/openapi/openapi.py", line 489, in validate_against_openapi_schema
raise SchemaError(message) from None
zerver.openapi.openapi.SchemaError: 1 response validation error(s) at post /api/v1/user_uploads (200):
ValidationError: Additional properties are not allowed ('uri' was unexpected)
Failed validating 'additionalProperties' in schema['allOf'][2]:
{'additionalProperties': False,
'example': {'msg': '',
'result': 'success',
'properties': {'msg': {}, 'result': {}}}
On instance:
{'msg': '',
'result': 'success',
'uri': '/user_uploads/2/85/XoqF0K7XEOLVGylgdpof80RB/img.jpg'}
We can see in the traceback that a SchemaError
was raised in
validate_against_openapi_schema
:
File "/srv/zulip/zerver/openapi/openapi.py", line 478, in validate_against_openapi_schema
raise SchemaError(message) from None
The next line in the output, lets us know how many errors were found and for what endpoint.
zerver.openapi.openapi.SchemaError: 1 response validation error(s) at post /api/v1/user_uploads (200):
As expected from commenting out the code above, there was one validation
error for the POST /api/v1/user_uploads
endpoint. The next line gives
more information about that error.
ValidationError: Additional properties are not allowed ('uri' was unexpected)
We see that there was a uri
value returned by the endpoint that hasn't
been documented. The next few lines of output, show us what return values
are documented (again due to our changes) for this endpoint.
Failed validating 'additionalProperties' in schema['allOf'][2]:
{'additionalProperties': False,
'example': {'msg': '',
'result': 'success',
'properties': {'msg': {}, 'result': {}}}
And finally, we see the test instance that did not match our current
documentation, which includes the uri
return value.
On instance:
{'msg': '',
'result': 'success',
'uri': '/user_uploads/2/85/XoqF0K7XEOLVGylgdpof80RB/img.jpg'}
This is a useful example because the endpoint's documentation is short and straightforward, helping to easily identify the parts of the error output that are useful in debugging these errors when testing the API documentation.
Adding a realm setting
Building on the new feature tutorial
example, if the realm setting for mandatory_topics
was not documented
in the POST /api/v1/register
endpoint, running tools/test-api
in the
development environment would result in this error:
...
zerver.openapi.openapi.SchemaError: 1 response validation error(s) at post /api/v1/register (200):
ValidationError: Additional properties are not allowed ('realm_mandatory_topics' was unexpected)
Failed validating 'additionalProperties' in schema['allOf'][2]:
'OpenAPI schema omitted due to length of output.'
On instance:
'Error instance omitted due to length of output.'
Because this endpoint is very long and descriptive, we do not print the entire documentation schema (or test instance, in this case) to the console. Doing so would print thousands of lines of output that are not useful for debugging what is missing from the API documentation.
The key information for debugging this endpoint is in the line beginning
with ValidationError
. There we can see that the documentation does not
include the new realm_mandatory_topics
boolean that we added in the
example feature tutorial, and we can look at other similar realm settings
to add the documentation for that new feature.