Code style and conventions

One can summarize Zulip’s coding philosophy as a relentless focus on making the codebase easy to understand and difficult to make dangerous mistakes in. The majority of work in any large software development project is understanding the existing code so one can debug or modify it, and investments in code readability usually end up paying for themselves when someone inevitably needs to debug or improve the code.

When there’s something subtle or complex to explain or ensure in the implementation, we try hard to make it clear, through a combination of clean and intuitive interfaces, well-named variables and functions, comments/docstrings, and commit messages (roughly in order of priority – if you can make something clear with a good interface, that’s a lot better than writing a comment explaining how the bad interface works).

This page documents code style policies that every Zulip developer should understand. We aim for this document to be short and focused only on details that cannot be easily enforced another way (e.g. through linters, automated tests, subsystem design that makes classes of mistakes unlikely, etc.). This approach minimizes the cognitive load of ensuring a consistent coding style for both contributors and maintainers.

Be consistent!

Look at the surrounding code, or a similar part of the project, and try to do the same thing. If you think the other code has actively bad style, fix it (in a separate commit).

When in doubt, ask in #development help.

Lint tools

You can run them all at once with


You can set this up as a local Git commit hook with


The Vagrant setup process runs this for you.

lint runs many lint checks in parallel, including


Please don’t put any passwords, secret access keys, etc. inline in the code. Instead, use the get_secret function or the get_mandatory_secret function in zproject/ to read secrets from /etc/zulip/secrets.conf.

Dangerous constructs

Too many database queries

Look out for Django code like this:

bars = Bar.objects.filter(...)
for bar in bars:
    foo =
    # Make use of foo

…because it equates to:

bars = Bar.objects.filter(...)
for bar in bars:
    foo = Foo.objects.get(
    # Make use of foo

…which makes a database query for every Bar. While this may be fast locally in development, it may be quite slow in production! Instead, tell Django’s QuerySet API to prefetch the data in the initial query:

bars = Bar.objects.filter(...).select_related()
for bar in bars:
    foo =  # This doesn't take another query, now!
    # Make use of foo

If you can’t rewrite it as a single query, that’s a sign that something is wrong with the database schema. So don’t defer this optimization when performing schema changes, or else you may later find that it’s impossible.

UserProfile.objects.get() / Client.objects.get() / etc.

In our Django code, never do direct UserProfile.objects.get(email=foo) database queries. Instead always use get_user_profile_by_{email,id}. There are 3 reasons for this:

  1. It’s guaranteed to correctly do a case-inexact lookup

  2. It fetches the user object from remote cache, which is faster

  3. It always fetches a UserProfile object which has been queried using .select_related() (see above!), and thus will perform well when one later accesses related models like the Realm.

Similarly we have get_client and access_stream_by_id / access_stream_by_name functions to fetch those commonly accessed objects via remote cache.

Using Django model objects as keys in sets/dicts

Don’t use Django model objects as keys in sets/dictionaries – you will get unexpected behavior when dealing with objects obtained from different database queries:

For example, the following will, surprisingly, fail:

# Bad example -- will raise!
obj: UserProfile = get_user_profile_by_id(17)
some_objs = UserProfile.objects.get(id=17)
assert obj in set([some_objs])

You should work with the IDs instead:

obj: UserProfile = get_user_profile_by_id(17)
some_objs = UserProfile.objects.get(id=17)
assert in set([ for i in some_objs])

You should always pass the update_fields keyword argument to .save() when modifying an existing Django model object. By default, .save() will overwrite every value in the column, which results in lots of race conditions where unrelated changes made by one thread can be accidentally overwritten by another thread that fetched its UserProfile object before the first thread wrote out its change.

Using raw saves to update important model objects

In most cases, we already have a function in zerver.actions with a name like do_activate_user that will correctly handle lookups, caching, and notifying running browsers via the event system about your change. So please check whether such a function exists before writing new code to modify a model object, since your new code has a good chance of getting at least one of these things wrong.

Naive datetime objects

Python allows datetime objects to not have an associated time zone, which can cause time-related bugs that are hard to catch with a test suite, or bugs that only show up during daylight savings time.

Good ways to make time-zone-aware datetimes are below. We import time zone libraries as from datetime import datetime, timezone and from django.utils.timezone import now as timezone_now.


  • timezone_now() to get a datetime when Django is available, such as in zerver/.

  • when Django is not available, such as for bots and scripts.

  • datetime.fromtimestamp(timestamp, tz=timezone.utc) if creating a datetime from a timestamp. This is also available as zerver.lib.timestamp.timestamp_to_datetime.

  • datetime.strptime(date_string, format).replace(tzinfo=timezone.utc) if creating a datetime from a formatted string that is in UTC.

Idioms that result in time-zone-naive datetimes, and should be avoided, are and datetime.fromtimestamp(timestamp) without a tz parameter, datetime.utcnow() and datetime.utcfromtimestamp(), and datetime.strptime(date_string, format) without replacing the tzinfo at the end.

Additional notes:

  • Especially in scripts and puppet configuration where Django is not available, using time.time() to get timestamps can be cleaner than dealing with datetimes.

  • All datetimes on the backend should be in UTC, unless there is a good reason to do otherwise.

x.attr('zid') vs.

Our message row DOM elements have a custom attribute zid which contains the numerical message ID. Don’t access this directly as x.attr('zid') ! The result will be a string and comparisons (e.g. with <=) will give the wrong result, occasionally, just enough to make a bug that’s impossible to track down.

You should instead use the id function from the rows module, as in This returns a number. Even in cases where you do want a string, use the id function, as it will simplify future code changes. In most contexts in JavaScript where a string is needed, you can pass a number without any explicit conversion.

JavaScript const and let

Always declare JavaScript variables using const or let rather than var.

JavaScript and TypeScript for (i in myArray)

Don’t use it: [1], [2], [3]

Translation tags

Remember to tag all user-facing strings for translation, whether they are in HTML templates or JavaScript/TypeScript editing the HTML (e.g. error messages).

Paths to state or log files

When writing out state or log files, always pass an absolute path through zulip_path (found in zproject/, which will do the right thing in both development and production.

JS array/object manipulation

For functions that operate on arrays or JavaScript objects, you should generally use modern ECMAScript primitives such as for of loops, Array.prototype.{entries, every, filter, find, indexOf, map, some}, Object.{assign, entries, keys, values}, spread syntax, and so on. Our Babel configuration automatically transpiles and polyfills these using core-js when necessary. We used to use the Underscore library, but that should be avoided in new code.

More arbitrary style things

Line length

We have an absolute hard limit on line length only for some files, but we should still avoid extremely long lines. A general guideline is: refactor stuff to get it under 85 characters, unless that makes the code a lot uglier, in which case it’s fine to go up to 120 or so.

JavaScript and TypeScript

Our JavaScript and TypeScript code is formatted with Prettier. You can ask Prettier to reformat all code via our linter tool with tools/lint --only=prettier --fix. You can also integrate it with your editor.

Combine adjacent on-ready functions, if they are logically related.

The best way to build complicated DOM elements is a Mustache template like static/templates/message_reactions.hbs. For simpler things you can use jQuery DOM building APIs like so:

var new_tr = $('<tr />').attr('id',;

Passing a HTML string to jQuery is fine for simple hardcoded things that don’t need internationalization:

foo.append('<p id="selected">/</p>');

but avoid programmatically building complicated strings.

We used to favor attaching behaviors in templates like so:

<p onclick="select_zerver({{id}})">

but there are some reasons to prefer attaching events using jQuery code:

  • Potential huge performance gains by using delegated events where possible

  • When calling a function from an onclick attribute, this is not bound to the element like you might think

  • jQuery does event normalization

Either way, avoid complicated JavaScript code inside HTML attributes; call a helper function instead.


Our CSS is formatted with Prettier. You can ask Prettier to reformat all code via our linter tool with tools/lint --only=prettier --fix. You can also integrate it with your editor.

Avoid using the style= attribute unless the styling is actually dynamic. Instead, define logical classes and put your styles in external CSS files such as zulip.css.

Don’t use the tag name in a selector unless you have to. In other words, use .foo instead of We shouldn’t have to care if the tag type changes in the future.


  • Our Python code is formatted with Black and isort. The linter tool enforces this by running Black and isort in check mode, or in write mode with tools/lint --only=black,isort --fix. You may find it helpful to integrate Black and isort with your editor.

  • Don’t put a shebang line on a Python file unless it’s meaningful to run it as a script. (Some libraries can also be run as scripts, e.g. to run a test suite.)

  • Scripts should be executed directly (./, so that the interpreter is implicitly found from the shebang line, rather than explicitly overridden (python

  • Put all imports together at the top of the file, absent a compelling reason to do otherwise.

  • Unpacking sequences doesn’t require list brackets:

    [x, y] = xs    # unnecessary
    x, y = xs      # better
  • For string formatting, use x % (y,) rather than x % y, to avoid ambiguity if y happens to be a tuple.


Clear, readable code is important for tests; familiarize yourself with our testing frameworks so that you can write clean, readable tests. Comments about anything subtle about what is being verified are appreciated.

Third party code

See our docs on dependencies for discussion of rules about integrating third-party projects.