16 KiB

Raw Blame History

Writing views in Zulip

What this covers

This page documents how views work in Zulip. You may want to read the new feature tutorial and treat this as a reference.

If you have experience with Django, much of this will be familiar, but you may want to read about how REST requests are dispatched, and how request authentication works.

This document supplements the new feature tutorial and the testing documentation.

What is a view?

A view in Zulip is everything that helps implement a server endpoint. Every path that the Zulip server supports (doesn't show a 404 page for) is a view. The obvious ones are those you can visit in your browser, for example /integrations, which shows the integration documentation. These paths show up in the address bar of the browser. There are other views that are only seen by software, namely the API views. They are used to build the various clients that Zulip has, namely the web client (which is also used by the desktop client) and the mobile clients.

Modifying urls.py

A view is anything with an entry in the appropriate urls.py, usually zproject/urls.py. Zulip views either serve HTML (pages for browsers) or JSON (data for Zulip clients on all platforms, custom bots, and integrations).

The format of the URL patterns in Django is documented here, and the Zulip specific details for these are discussed in detail in the life of a request doc.

We have two Zulip-specific conventions we use for internationalization and for our REST API, respectively.

Writing human-readable views

If you're writing a new page for the website, make sure to add it to i18n_urls in zproject/urls.py

     i18n_urls = [
     ...
+    path('quote-of-the-day', TemplateView.as_view(template_name='zerver/qotd.html')),
+    path('postcards', 'zerver.views.postcards'),
]

As an example, if a request comes in for Spanish, language code es, the server path will be something like: es/features/.

Decorators used for webpage views

This section documents a few simple decorators that we use for webpage views, as an introduction to view decorators.

require_post:


@require_post
def accounts_register(request: HttpRequest) -> HttpResponse:

This decorator ensures that the request was a POST--here, we're checking that the registration submission page is requested with a post, and inside the function, we'll check the form data. If you request this page with GET, you'll get an HTTP 405 METHOD NOT ALLOWED error.

zulip_login_required:

This decorator verifies that the browser is logged in (i.e. has a valid session cookie) before providing the view for this route, or redirects the browser to a login page. This is used in the root path (/) of the website for the web client. If a request comes from a browser without a valid session cookie, they are redirected to a login page. It is a small fork of Django's login_required, adding a few extra checks specific to Zulip.

@zulip_login_required
def home(request: HttpRequest) -> HttpResponse:

Writing a template

Templates for the main website are found in templates/zerver/app.

Writing API REST endpoints

These are code-parsable views that take x-www-form-urlencoded or JSON request bodies, and return JSON-string responses. Almost all Zulip view code is in the implementations of API REST endpoints.

The REST API does authentication of the user through rest_dispatch, which is documented in detail at zerver/lib/rest.py. This method will authenticate the user either through a session token from a cookie on the browser, or from a base64 encoded email:api-key string given via HTTP basic auth for API clients.

>>> import requests
>>> r = requests.get('https://api.github.com/user', auth=('hello@example.com', '0123456789abcdeFGHIJKLmnopQRSTUV'))
>>> r.status_code
-> 200

Request variables

Most API views will have some arguments that are passed as part of the request to control the behavior of the view. In any well-engineered view, you need to write code to parse and validate that the arguments exist and have the correct form. For many applications, this leads to one of several bad outcomes:

The code isn't written, so arguments aren't validated, leading to bugs and confusing error messages for users of the API.
Every function starts with a long list of semi-redundant validation code, usually with highly inconsistent error messages.
Every view function comes with another function that does the validation that has the problems from the last bullet point.

In Zulip, we solve this problem with a special decorator called typed_endpoint which allows a developer to declare the arguments a view function takes and validate their types all within the def line of the function. This framework uses Pydantic V2 to perform data validation and parsing for the view arguments. We like this framework because we have found it makes the validation code compact, readable, and conveniently located in the same place as the method it is validating arguments for.

Here's an example:

from zerver.decorator import require_realm_admin
from zerver.lib.typed_endpoint import typed_endpoint

@require_realm_admin
@typed_endpoint
def create_user_backend(
        request: HttpRequest,
        user_profile: UserProfile,
        *,
        email: str,
        password: str,
        full_name: str,
    ):
    # ... code here

The typed_endpoint decorator parses the declared keyword-only arguments of the decorated function, and for each argument that has been declared, it extracts the HTTP parameter with that name from the request, parses it according to the type annotation, and then passes it to the function. It will return an nicely JSON formatted HTTP 400 error in the event that an argument is missing, doesn't parse as JSON, or otherwise is invalid.

require_realm_admin is another decorator which checks the authorization of the given user_profile to make sure it belongs to a realm administrator (and thus has permission to create a user); we show it here primarily to show how typed_endpoint should be the inner decorator.

The implementation of typed_endpoint is documented in detail in zerver/lib/typed_endpoint.py

Pydantic also helps us with request variable validation. For example:

msg_ids: Json[list[int]] will check that the msg_ids HTTP parameter is a list of integers, marshalled as JSON, and pass it into the function as the msg_ids Python keyword argument.
streams_raw: Annotated[Json[list[str]], ApiParamConfig("subscriptions")] will check that the "subscriptions" HTTP parameter is a list of strings, marshalled as JSON, and pass it into the function with the Python keyword argument streams_raw.
message_id: Json[NonNegativeInt] will check that the message_id HTTP parameter is a string containing a JSON encoded non-negative integer.

Annotated can be used in combination with Pydantic's validators to provide additional validation for the arguments.

name: Annotated[str, StringConstraints(max_length=60)] will check that the name HTTP parameter is a string containing up to 60 characters.
Since there is no need to JSON-encode strings (lists, integers, bools and complex objects require JSON encoding), usually simply my_string: str is correct. One can pass, for example, Annotated[str, check_string_in_validator(...)] where one wants to run a validator on the value of a string.

Default values can be specified for optional arguments similar to how we would specify default values in regular python function.

is_default_stream: Json[bool] = False will assign False to the is_default_stream argument if no value is specified when making a request to the endpoint.
We can use None as the default value for optional arguments when we don't want to specify any specific default value, for example, narrow: Json[list[NarrowParameter]] | None = None. This does not allow the caller to pass None as the value, the only way narrow can be set to None is by using the default value.

Pydantic models can be used to define the schema of complex objects that can be passed to the endpoint.

Here's an example:

from typing import Annotated

from pydantic import BaseModel, StringConstraints, model_validator

class AddSubscriptionData(BaseModel):
    name: str
    color: str | None = None
    description: (
        Annotated[str, StringConstraints(max_length=Stream.MAX_DESCRIPTION_LENGTH)] | None
    ) = None

    @model_validator(mode="after")
    def validate_terms(self) -> "AddSubscriptionData":
        # ... Validation logic here
        return self

add: Json[list[AddSubscriptionData]] will require the add argument to be a list of objects having the keys that are specified in the AddSubscriptionData model.
@model_validator can be used to specify additional validation logic for the model.

See zerver/lib/typed_endpoint_validators.py for more validators and their documentation.

Deciding which HTTP verb to use

When writing a new API view, you should writing a view to do just one type of thing. Usually that's either a read or write operation.

If you're reading data, GET is the best option. Other read-only verbs are HEAD, which should be used for testing if a resource is available to be read with GET, without the expense of the full GET. OPTIONS is also read-only, and used by clients to determine which HTTP verbs are available for a given path. This isn't something you need to write, as it happens automatically in the implementation of rest_dispatch--see zerver/lib/rest.py for more.

If you're creating new data, try to figure out if the thing you are creating is uniquely identifiable. For example, if you're creating a user, there's only one user per email. If you can find a unique ID, you should use PUT for the view. If you want to create the data multiple times for multiple requests (for example, requesting the send_message view multiple times with the same content should send multiple messages), you should use POST.

If you're updating existing data, use PATCH.

If you're removing data, use DELETE.

Idempotency

When writing a new API endpoint, with the exception of things like sending messages, requests should be safe to repeat, without impacting the state of the server. This is idempotency.

You will often want to return an error if a request to change something would do nothing because the state is already as desired, to make debugging Zulip clients easier. This means that the response for repeated requests may not be the same, but the repeated requests won't change the server more than once or cause unwanted side effects.

Making changes to the database

If the view does any modification to the database, that change is done in a helper function in zerver/actions/*.py. Those functions are responsible for doing a complete update to the state of the server, which often entails both updating the database and sending any events to notify clients about the state change. When possible, we prefer to design a clean boundary between the view function and the actions function is such that all user input validation happens in the view code (i.e. all 400 type errors are thrown there), and the actions code is responsible for atomically executing the change (this is usually signalled by having the actions function have a name starting with do_). So in most cases, errors in an actions function will be the result of an operational problem (e.g., lost connection to the database) and lead to a 500 error. If an actions function is responsible for validation as well, it should have a name starting with check_.

For example, in zerver/views/realm.py:

@require_realm_admin
@typed_endpoint
def update_realm(
    request: HttpRequest,
    user_profile: UserProfile,
    *,
    name: Annotated[str | None, StringConstraints(max_length=Realm.MAX_REALM_NAME_LENGTH)] = None,
    # ...
):
    realm = user_profile.realm
    # ...
            do_set_realm_property(realm, k, v, acting_user=user_profile)
    # ...

realm.save() actually saves the changes to the realm to the database, and send_event sends the event to active clients belonging to the provided list of users (in this case, all active users in the Zulip realm).

Calling from the web application

You should always use channel.<method> to make an HTTP <method> call to the Zulip JSON API. As an example, in web/src/admin.js

var url = "/json/realm";
var data = {
    name: JSON.stringify(new_name),
}
channel.patch({
    url: url,
    data: data,
    success: function (response_data) {
        if (response_data.name !== undefined) {
            ui_report.success($t({defaultMessage: "Name changed!"}), name_status);
        }
        ...

Calling from an API client

Here's how you might manually make a call from python:

payload = {'name': new_name}

# email and API key
api_auth = ('hello@example.com', '0123456789abcdeFGHIJKLmnopQRSTUV')

r = requests.patch(SERVER_URL + 'api/v1/realm',
                   data=json.dumps(payload),
                   auth=api_auth,
                  )

This is simply an illustration; we recommend making use of the Zulip Python API bindings since they provide a nice interface for accessing the API.

Legacy endpoints used by the web client

New features should conform the REST API style. The legacy, web-only endpoints can't effectively enforce usage of a browser, so they aren't preferable from a security perspective, and it is generally a good idea to make your feature available to other clients, especially the mobile clients.

These endpoints make use the older authentication decorator authenticated_json_view, so you may see it in the code.

Incoming webhook integrations

Incoming webhooks are called by other services, often to send a message as part of those services' integrations. They are most often POST requests, and often there is very little you can customize about them. Usually you can expect that the webhook for a service will allow specification for the target server for the webhook, and an API key.

If the webhook does not have an option to provide a bot email, use the webhook_view decorator, to fill in the user_profile and request.client fields of a request:

@webhook_view("PagerDuty", all_event_types=ALL_EVENT_TYPES)
@typed_endpoint
def api_pagerduty_webhook(
    request: HttpRequest,
    user_profile: UserProfile,
    *,
    payload: JsonBodyPayload[WildValue],

request.client will be the result of get_client("ZulipPagerDutyWebhook") in this example and it will be passed to check_send_stream_message. For more information, see Clients in Zulip.

16 KiB Raw Blame History