zulip/docs/events-system.md

Real-time Push and Events

Zulip's "events system" is the server-to-client push system that powers our real-time sync. This document explains how it works; to read an example of how a complete feature using this system works, check out the new application feature tutorial.

Any single-page web application like Zulip needs a story for how changes made by one client are synced to other clients, though having a good architecture for this is particularly important for a chat tool like Zulip, since the state is constantly changing. When we talk about clients, think a browser tab, mobile app, or API bot that needs to receive updates to the Zulip data. The simplest example is a new message being sent by one client; other clients must be notified in order to display the message. But a complete application like Zulip has dozens of different types of data that need to be synced to other clients, whether it be new streams, changes in a user's name or avatar, settings changes, etc. In Zulip, we call these updates that need to be sent to other clients events.

An important thing to understand when designing such a system is that events need to be synced to every client that has a copy of the old data if one wants to avoid clients displaying inaccurate data to users. So if a user has two browser windows open and sends a message, every client controlled by that user as well as any recipients of the message, including both of those two browser windows, will receive that event. (Technically, we don't need to send events to the client that triggered the change, but this approach saves a bunch of unnecessary duplicate UI update code, since the client making the change can just use the same code as every other client, maybe plus a little notification that the operation succeeded).

Architecturally, there are a few things needed to make a successful real-time sync system work:

• Generation. Generating events when changes happen to data, and determining which users should receive each event.
• Delivery. Efficiently delivering those events to interested clients, ideally in an exactly-once fashion.
• UI updates. Updating the UI in the client once it has received events from the server.

Reactive JavaScript libraries like React and Vue can help simplify the last piece, but there aren't good standard systems for doing generation and delivery, so we have to build them ourselves.

This document discusses how Zulip solves the generation and delivery problems in a scalable, correct, and predictable way.

Generation system

Zulip's generation system is built around a Python function, send_event(event, users). It accepts an event data structure (just a Python dictionary with some keys and value; type is always one of the keys but the rest depends on the specific event) and a list of user IDs for the users whose clients should receive the event. In special cases such as message delivery, the list of users will instead be a list of dicts mapping user IDs to user-specific data like whether that user was mentioned in that message. The data passed to send_event are simply marshalled as JSON and placed in the notify_tornado RabbitMQ queue to be consumed by the delivery system.

Usually, this list of users is one of 3 things:

• A single user (e.g. for user-level settings changes).
• Everyone in the realm (e.g. for organization-level settings changes, like new realm emoji).
• Everyone who would receive a given message (for messages, emoji reactions, message editing, etc.); i.e. the subscribers to a stream or the people on a private message thread.

It is the responsibility of the caller of send_event to choose the list of user IDs correctly. There can be security problems if e.g. an event containing private message content is sent to the entire organization. However, if an event isn't sent to enough clients, there will likely be user-visible real-time sync bugs.

Most of the hard work in event generation is about defining consistent event dictionaries that are clear, readable, will be useful to the wide range of possible clients, and make it easy for developers.

Delivery system

Zulip's event delivery (real-time push) system is based on Tornado, which is ideal for handling a large number of open requests. Details on Tornado are available in the architecture overview, but in short it is good at holding open a large number of connections for a long time. The complete system is about 1500 lines of code in zerver/tornado/, primarily zerver/tornado/event_queue.py.

Zulip's event delivery system is based on "long-polling"; basically clients make GET /json/events calls to the server, and the server doesn't respond to the request until it has an event to deliver to the client. This approach is reasonably efficient and works everywhere (unlike websockets, which have a decreasing but nonzero level of client compatibility problems).

For each connected client, the Event Queue Server maintain an event queue, which contains any events that are to be delivered to that client which have not yet been acknowledged by that client. Ignoring the subtle details around error handling, the protocol is pretty simple; when a client does a GET /json/events call, the server checks if there are any events in the queue. If there are, it returns the events immediately. If there aren't, it records that queue as having a waiting client (often called a handler in the code).

When it pulls an event off the notify_tornado RabbitMQ queue, it simply delivers the event to each queue associated with one of the target users. If the queue has a waiting client, it breaks the long-poll connection by returning an HTTP response to the waiting client request. If there is no waiting client, it simply pushes the event onto the queue.

When starting up, each client makes a POST /json/register to the server, which creates a new event queue for that client and return the queue_id as well as an initial last_event_id to the client (it can also, optionally, fetch the initial data to save an RTT and avoid races; see the below section on initial data fetches for details on why this is useful). Once the event queue is registered, the client can just do an infinite loop calling GET /json/events with those parameters, updating last_event_id each time to acknowledge any events it has received (see call_on_each_event in the Zulip Python API bindings for a complete example implementation). When handling each GET /json/events request, the queue server can safely delete any events have have an event ID less than or equal to the client's last_event_id (event IDs are just a counter for the events a given queue has received.)

If network failures were impossible, the last_event_id parameter in the protocol would not be required, but it is important for enabling exactly-once delivery in the presence of potential failures. (Without it, the queue server would have to delete events from the queue as soon as it attempted to send them to the client; if that specific HTTP response didn't reach the client due to a network TCP failure, then those events could be lost).

The queue servers are a very high-traffic system, processing at a minimum a request for every message delivered to every Zulip client. Additionally, as a workaround for low-quality NAT servers that kill HTTP connections that are open without activity for more than 60s, the queue servers also send a heartbeat event to each queue at least once every 45s or so (if no other events have arrived in the meantime).

To avoid a large memory and other resource leak, the queues are garbage collected after (by default) 10 minutes of inactivity from a client, under the theory that the client has likely gone off the Internet (or no longer exists) access; this happens constantly. If the client returns, it will receive a "queue not found" error when requesting events; it's handler for this case should just restart the client / reload the browser so that it refetches initial data the same way it would on startup. Since clients have to implement their startup process anyway, this approach adds minimal technical complexity to clients. A nice side effect is that if the Event Queue Server server (which stores queues in memory) were to crash and lose its data, clients would recover, just as if they had lost Internet access briefly (there is some DoS risk to manage, though).

(The Event Queue Server is designed to save any event queues to disk and reload them when the server is restarted, and catches exceptions carefully, so such incidents are very rare, but it's nice to have a design that handles them without leaving broken out-of-date clients anyway).

The initial data fetch

When a client starts up, it usually wants to get 2 things from the server:

• The "current state" of various pieces of data, e.g. the current settings, set of users in the organization (for typeahead), stream, messages, etc. (aka the "initial state").
• A subscription to receive updates to those data when they are changed by a client (aka an event queue).

Ideally, one would get those two things atomically, i.e. if some other user changes their name, either the name change happens before the fetch (and thus the old name is in the initial state and there will be an event in the queue for the name change) or after (the new name is in the initial state, and there is no event for that name change in the queue).

Achieving this atomicity goals means we save a huge amount of work that the N clients for Zulip don't need to worry about a wide range of potential rare and hard to reproduce race conditions; we just have to implement things correctly once in the Zulip server.

This is quite challenging to do technically, because fetching the initial state for a complex web application like Zulip might involve dozens of queries to the database, caches, etc. over the course of 100ms or more, and it is thus nearly impossible to do all of those things together atomically. So instead, we use a more complicated algorithm that can produce the atomic result from non-atomic subroutines. Here's how it works when you make a register API request. The request is directly handled by Django:

• Django makes an HTTP request to Tornado, requesting that a new event queue be created, and records its queue ID.
• Django does all the various database/cache/etc. queries to fetch the data, non-atomically, from the various data sources (see the fetch_initial_state_data function).
• Django makes a second HTTP request to Tornado, requesting any events that had been added to the Tornado event queue since it was created.
• Finally, Django "applies" the events (see the apply_events function) to the initial state that it fetched. E.g. for a name change event, it finds the user data in the realm_user data struture, and updates it to have the new name.

This achieves everything we desire, at the cost that we need to write the apply_events function. This is a difficult function to implement correctly, because the situations that it tests for almost never happen (being race conditions). So we have a special test class, EventsRegisterTest, that is specifically designed to test this function by ensuring the possible race condition always happens. In particular, it does the following:

• Call fetch_initial_state_data to get the current state.
• Call a state change function that issues an event, e.g. do_change_full_name, and capture any events that are generated.
• Call apply_events(state, events), and compare the result to calling fetch_initial_state_data again now.

The apply_events code is correct if those two results are identical.

The final detail we need to ensure that apply_events always works correctly is to make sure that we have EventsRegisterTest tests for every event type that can be generated by Zulip. This can be tested manually using test-backend --coverage EventsRegisterTest and then checking that all the calls to send_event are covered. Someday we'll add automation that verifies this directly by inspecting the coverage data.

Messages

One exception to the protocol described in the last section is the actual messages. Because Zulip clients usually fetch them in a separate AJAX call after the rest of the site is loaded, we don't need them to be included in the initial state data. To handle those correctly, clients are responsible for discarding events related to messages that the client has not yet fetched.