The Django Speed Handbook: making a Django app faster

February 25, 2020

Over the course of developing several Django apps, I’ve learned quite a bit about speed optimizations. Some parts of this process, whether on the backend or frontend, are not well-documented. I’ve decided to collect most of what I know in this article.

If you haven’t taken a close look at the performance of your web-app yet, you’re bound to find something good here.

What’s in this article?

Why speed is important

On the web, 100 milliseconds can make a significant difference and 1 second is a lifetime. Countless studies indicate that faster loading times are associated with better conversion-rates, user-retention, and organic traffic from search engines. Most importantly, they provide a better user experience.

Different apps, different bottlenecks

There are many techniques and practices to optimize your web-app’s performance. It’s easy to get carried away. Look for the highest return-to-effort ratio. Different web-apps have different bottlenecks and therefore will gain the most when those bottlenecks are taken care of. Depending on your app, some tips will be more useful than others.

While this article is catered to Django developers, the speed optimization tips here can be adjusted to pretty much any stack. On the frontend side, it’s especially useful for people hosting with Heroku and who do not have access to a CDN service.

Analyzing and debugging performance issues

On the backend, I recommend the tried-and-true django-debug-toolbar. It will help you analyze your request/response cycles and see where most of the time is spent. Especially useful because it provides database query execution times and provides a nice SQL EXPLAIN in a separate pane that appears in the browser.

Google PageSpeed will display mainly frontend related advice, but some can apply to the backend as well (like server response times). PageSpeed scores do not directly correlate with loading times but should give you a good picture of where the low-hanging fruits for your app are. In development environments, you can use Google Chrome’s Lighthouse which provides the same metrics but can work with local network URIs. GTmetrix is another detail-rich analysis tool.


Some people will tell you that some of the advice here is wrong or lacking. That’s okay; this is not meant to be a bible or the ultimate go-to-guide. Treat these techniques and tips as ones you may use, not should or must use. Different needs call for different setups.

Backend: the database layer

Starting with the backend is a good idea since it’s usually the layer that’s supposed to do most of the heavy lifting behind the scenes.

There’s little doubt in my mind which two ORM functionalities I want to mention first: these are select_related and prefetch_related. They both deal specifically with retrieving related objects and will usually improve speed by minimizing the number of database queries.

Let’s take a music web-app for example, which might have these models:

# music/, some fields & code omitted for brevity
class RecordLabel(models.Model):
    name = models.CharField(max_length=560)

class MusicRelease(models.Model):
    title = models.CharField(max_length=560)
    release_date = models.DateField()

class Artist(models.Model):
    name = models.CharField(max_length=560)
    label = models.ForeignKey(
    music_releases = models.ManyToManyField(

So each artist is related to one and only one record company and each record company can sign multiple artists: a classic one-to-many relationship. Artists have many music-releases, and each release can belong to one artist or more.

I’ve created some dummy data:

  • 20 record labels
  • each record label has 25 artists
  • each artist has 100 music releases

Overall, we have ~50,500 of these objects in our tiny database.

Now let’s wire-up a fairly standard function that pulls our artists and their label. django_query_analyze is a decorator I wrote to count the number of database queries and time to run the function. Its implementation can be found in the appendix.

# music/
def get_artists_and_labels():
    result = []
    artists = Artist.objects.all()
    for artist in artists:
        result.append({"name":, "label":})
    return result

get_artists_and_labels is a regular function which you may use in a Django view. It returns a list of dictionaries, each contains the artist’s name and their label. I’m accessing to force-evaluate the Django QuerySet; you can equate this to trying to access these objects in a Jinja template:

{% for artist in artists_and_labels %}
    <p>Name: {{ }}, Label: {{ }}</p>
{% endfor %}

Now let’s run this function:

ran function get_artists_and_labels
number of queries: 501
Time of execution: 0.3585s

So we’ve pulled 500 artists and their labels in 0.36 seconds, but more interestingly — we’ve hit the database 501 times. Once for all the artists, and 500 more times: once for each of the artists’ labels. This is called The N+1 problem”. Let’s tell Django to retrieve each artist’s label in the same query with select_related:

def get_artists_and_labels_select_related():
    result = []
    artists = Artist.objects.select_related("label") # select_related
    for artist in artists:
            {"name":, "label": if artist.label else "N/A"}
    return result

Now let’s run this:

ran function get_artists_and_labels_select_related
number of queries: 1
Time of execution: 0.01481s

500 queries less and a 96% speed improvement.

Let’s look at another function, for getting each artist’s first 100 music releases:

def get_artists_and_releases():
    result = []
    artists = Artist.objects.all()[:100]
    for artist in artists:
                "releases": [release.title for release in artist.music_releases.all()],
    return result

How long does it take to fetch 100 artists and 100 releases for each one of them?

ran function get_artists_and_releases
number of queries: 101
Time of execution: 0.18245s

Let’s change the artists variable in this function and add select_related so we can bring the number of queries down and hopefully get a speed boost:

artists = Artist.objects.select_related("music_releases")

If you actually do that, you’ll get an error:

django.core.exceptions.FieldError: Invalid field name(s) given in select_related: 'music_releases'. Choices are: label

That’s because select_related can only be used to cache ForeignKey or OneToOneField attributes. The relationship between Artist and MusicRelease is many-to-many though, and that’s where prefetch_ related comes in:

def get_artists_and_releases_prefetch_related():
    result = []
    artists = Artist.objects.all()[:100].prefetch_related("music_releases") # prefetch_related
    for artist in artists:
                "releases": [rel.title for rel in artist.music_releases.all()],
    return result

select_related can only cache the one” side of the one-to-many” relationship, or either side of a one-to-one” relationship. You can use prefetch_related for all other caching, including the many side in one-to-many relationships, and many-to-many relationships. Here’s the improvement in our example:

ran function get_artists_and_releases_prefetch_related
number of queries: 2
Time of execution: 0.13239s


Things to keep in mind about select_related and prefetch_related:

  • If you aren’t pooling your database connections, the gains will be even bigger because of fewer roundtrips to the database.
  • For very large result-sets, running prefetch_related can actually make things slower.
  • One database query isn’t necessarily faster than two or more.


Indexing your database columns can have a big impact on query performance. Why then, is it not the first clause of this section? Because indexing is more complicated than simply scattering db_index=True on your model fields.

Creating an index on frequently accessed columns can improve the speed of look-ups pertaining to them. Indexing comes at the cost of additional writes and storage space though, so you should always measure your benefit:cost ratio. In general, creating indices on a table will slow down inserts/updates.

Take only what you need

When possible, use values() and especially values_list() to only pull the needed properties of your database objects. Continuing our example, if we only want to display a list of artist names and don’t need the full ORM objects, it’s usually better to write the query like so:

artist_names = Artist.objects.values('name') 
# <QuerySet [{'name': 'Chet Faker'}, {'name': 'Billie Eilish'}]>

artist_names = Artist.objects.values_list('name') 
# <QuerySet [('Chet Faker',), ('Billie Eilish',)]>

artist_names = Artist.objects.values_list('name', flat=True) 
# <QuerySet ['Chet Faker', 'Billie Eilish']>

Haki Benita, a true database expert (unlike me), reviewed some parts of this section. You should read Haki’s blog.

Backend: the request layer

The next layer we’re going to look at is the request layer. These are your Django views, context processors, and middleware. Good decisions here will also lead to better performance.


In the section about select_related we were using the function to return 500 artists and their labels. In many situations returning this many objects is either unrealistic or undesirable. The section about pagination in the Django docs is crystal clear on how to work with the Paginator object. Use it when you don’t want to return more than N objects to the user, or when doing so makes your web-app too slow.

Asynchronous execution/background tasks

There are times when a certain action inevitably takes a lot of time. For example, a user requests to export a big number of objects from the database to an XML file. If we’re doing everything in the same process, the flow looks like this:

web: user requests file -> process file -> return response

Say it takes 45 seconds to process this file. You’re not really going to let the user wait all this time for a response. First, because it’s a horrible experience from a UX standpoint, and second, because some hosts will actually cut the process short if your app doesn’t respond with a proper HTTP response after N seconds.

In most cases, the sensible thing to do here is to remove this functionality from the request-response loop and relay it to a different process:

web: user requests file -> delegate to another process -> return response 
background process:        receive job -> process file -> notify user

Background tasks are beyond the scope of this article but if you’ve ever needed to do something like the above I’m sure you’ve heard of libraries like Celery.

Compressing Django’s HTTP responses

This is not to be confused with static-file compression, which is mentioned later in the article.

Compressing Django’s HTTP/JSON responses also stands to save your users some latency. How much exactly? Let’s check the number of bytes in our response’s body without any compression:

Content-Length: 66980
Content-Type: text/html; charset=utf-8

So our HTTP response is around 67KB. Can we do better? Many use Django’s built-in GZipMiddleware for gzip compression, but today the newer and more effective brotli enjoys the same support across browsers (except IE11, of course).

Important: Compression can potentially open your website to security breaches, as mentioned in the GZipMiddleware section of the Django docs.

Let’s install the excellent django-compression-middleware library. It will choose the fastest compression mechanism supported by the browser by checking the request’s Accept-Encoding headers:

pip install django-compression-middleware

Include it in our Django app’s middleware:

    # ...

And inspect the body’s Content-Length again:

Content-Encoding: br
Content-Length: 7239
Content-Type: text/html; charset=utf-8

The body size is now 7.24KB, 89% smaller. You can certainly argue this kind of operation should be delegated to a dedicated server like Ngnix or Apache. I’d argue that everything is a balance between simplicity and resources.


Caching is the process of storing the result of a certain calculation for faster future retrieval. Django has an excellent caching framework that lets you do this on a variety of levels and using different storage backends.

Caching can be tricky in data-driven apps: you’d never want to cache a page that’s supposed to display up-to-date, realtime information at all times. So, the big challenge isn’t so much setting up caching as it is figuring out what should be cached, for how long, and understanding when or how the cache is invalidated.

Before resorting to caching, make sure you’ve made proper optimizations at the database-level and/or on the frontend. If designed and queried properly, databases are ridiculously fast at pulling out relevant information at scale.

Frontend: where it gets hairier

Reducing static files/assets sizes can significantly speed up your web application. Even if you’ve done everything right on the backend, serving your images, CSS, and JavaScript files inefficiently will degrade your application’s speed.

Between compiling, minifying, compressing, and purging, it’s easy to get lost. Let’s try not to.

Serving static-files

You have several options on where and how to serve static files. Django’s docs mention a dedicated server running Ngnix and Apache, Cloud/CDN, or the same-server approach.

I’ve gone with a bit of a hybrid attitude: images are served from a CDN, large file-uploads go to S3, but all serving and handling of other static assets (CSS, JavaScript, etc…) is done using WhiteNoise (covered in-detail later).


Just to make sure we’re on the same page, here’s what I mean when I say:

  • Compiling: If you’re using SCSS for your stylesheets, you’ll first have to compile those to CSS because browsers don’t understand SCSS.
  • Minifying: reducing whitespace and removing comments from CSS and JS files can have a significant impact on their size. Sometimes this process involves uglifying: the renaming of long variable names to shorter ones, etc…
  • Compressing/Combining: for CSS and JS, combining multiple files to one. For images, usually means removing some data from images to make their files size smaller.
  • Purging: remove unneeded/unused code. In CSS for example: removing selectors that aren’t used.

Serving static files from Django with WhiteNoise

WhiteNoise allows your Python web-application to serve static assets on its own. As its author states, it comes in when other options like Nginx/Apache are unavailable or undesired.

Let’s install it:

pip install whitenoise[brotli]

Before enabling WhiteNoise, make sure your STATIC_ROOT is defined in

STATIC_ROOT = os.path.join(BASE_DIR, "staticfiles")

To enable WhiteNoise, add its WhiteNoise middleware right below SecurityMiddleware in

  # ...

In production, you’ll have to run collectstatic for WhiteNoise to work.

While this step is not mandatory, it’s strongly advised to add caching and compression:


Now whenever it encounters a {% static %} tag in templates, WhiteNoise will take care of compressing and caching the file for you. It also takes care of cache-invalidation.

One more important step: To ensure that we get a consistent experience between development and production environments, we add runserver_nostatic:

    # ...

This can be added regardless of whether DEBUG is True or not, because you don’t usually run Django via runserver in production.

I found it useful to also increase the caching time:

# Whitenoise cache policy
WHITENOISE_MAX_AGE = 31536000 if not DEBUG else 0 # 1 year

Wouldn’t this cause problems with cache-invalidation? No, because WhiteNoise creates versioned files when you run collectstatic:

<link rel="stylesheet" href="/static/CACHE/css/4abd0e4b71df.css" type="text/css" media="all">

So when you deploy your application again, your static files are overwritten and will have a different name, thus the previous cache becomes irrelevant.

Compressing and combining with django-compressor

WhiteNoise already compresses static files, so django-compressor is optional. But the latter offers an additional enhancement: combining the files. To use compressor with WhiteNoise we have to take a few extra steps.

Let’s say the user loads an HTML document that links three .css files:

  <link rel="stylesheet" href="base.css" type="text/css" media="all">
  <link rel="stylesheet" href="additions.css" type="text/css" media="all">
  <link rel="stylesheet" href="new_components.css" type="text/css" media="all">

Your browser will make three different requests to these locations. In many scenarios it’s more effective to combine these different files when deploying, and django-compressor does that with its {% compress css %} template tag:


{% load compress %}
  {% compress css %}
    <link rel="stylesheet" href="base.css" type="text/css" media="all">
    <link rel="stylesheet" href="additions.css" type="text/css" media="all">
    <link rel="stylesheet" href="new_components.css" type="text/css" media="all">
  {% compress css %}


  <link rel="stylesheet" href="combined.css" type="text/css" media="all">

Let’s go over the steps to make django-compressor and WhiteNoise play well. Install:

pip install django_compressor

Tell compressor where to look for static files:

COMPRESS_ROOT = os.path.abspath(STATIC_ROOT)

Because of the way these two libraries intercept the request-response cycle, they’re incompatible with their default configurations. We can overcome this by modifying some settings.

I prefer to use environment variables in .env files and have one Django, but if you have settings/ and settings/, you’ll know how to convert these values:


from decouple import config


COMPRESS_OFFLINE is True in production and False in development. COMPRESS_ENABLED is True in both1.

With offline compression, one must run compress on every deployment. On Heroku, you’ll want to disable the platform from automatically running collectstatic for you (on by default) and instead opt to do that in the post_compile hook, which Heroku will run when you deploy. If you don’t already have one, create a folder called bin at the root of your project and inside of it a file called post_compile with the following:

python collectstatic --noinput
python compress --force
python collectstatic --noinput

Another nice thing about compressor is that it can compress SCSS/SASS files:

    ("text/x-sass", "django_libsass.SassCompiler"),
    ("text/x-scss", "django_libsass.SassCompiler"),

Minifying CSS & JS

Another important thing to apply when talking about load-times and bandwidth usage is minifying: the process of (automatically) decreasing your code’s file-size by eliminating whitespace and removing comments.

There are several approaches to take here, but if you’re using django-compressor specifically, you get that for free as well. You just need to add the following (or any other filters compressor supports) to your file:

    "css": [
    "js": ["compressor.filters.jsmin.JSMinFilter"],

Defer-loading JavaScript

Another thing that contributes to slower performance is loading external scripts. The gist of it is that browsers will try to fetch and execute JavaScript files in the <head> tag as they are encountered and before parsing the rest of the page:

    <script src="https://will-block.js"></script>
      <script src="https://will-also-block.js"></script>

We can use the async and defer keywords to mitigate this:

      <script async src="somelib.somecdn.js"></script>

async and defer both allow the script to be fetched asynchronously without blocking. One of the key differences between them is when the script is allowed to execute: With async, once the script has been downloaded, all parsing is paused until the script has finished executing, while with defer the script is executed only after all HTML has been parsed.

I suggest referring to Flavio Copes’ article on the defer and aysnc keywords. Its general conclusion is:

The best thing to do to speed up your page loading when using scripts is to put them in the head, and add a defer attribute to your script tag.

Lazy-loading images

Lazily loading images means that we only request them when or a little before they enter the client’s (user’s) viewport. It saves time and bandwidth ($ on cellular networks) for your users. With excellent, dependency-free JavaScript libraries like LazyLoad, there really isn’t an excuse to not lazy-load images. Moreover, Google Chrome natively supports the lazy attribute since version 76.

Using the aforementioned LazyLoad is fairly simple and the library is very customizable. In my own app, I want it to apply on images only if they have a lazy class, and start loading an image 300 pixels before it enters the viewport:

$(document).ready(function (e) {
  new LazyLoad({
    elements_selector: ".lazy", // classes to apply to
    threshold: 300 // pixel threshold

Now let’s try it with an existing image:

<img class="album-artwork" alt="{{ album.title }}"  src="{{ album.image_url }}">

We replace the src attribute with data-src and add lazy to the class attribute:

<img class="album-artwork lazy" alt="{{ album.title }}"  data-src="{{ album.image_url }}">

Now the client will request this image when the latter is 300 pixels under the viewport.

If you have many images on certain pages, using lazy-loading will dramatically improve your load times.

Optimize & dynamically scale images

Another thing to consider is image-optimization. Beyond compression, there are two more techniques to consider here.

First, file-format optimization. There are newer formats like WebP that are presumably 25-30% smaller than your average JPEG image at the same quality. As of 02/2020 WebP has decent but incomplete browser support, so you’ll have to provide a standard format fallback if you want to use it.

Second, serving different image-sizes to different screen sizes: if some mobile device has a maximum viewport width of 650px, then why serve it the same 1050px image you’re displaying to 13″ 2560px retina display?

Here, too, you can choose the level of granularity and customization that suits your app. For simpler cases, You can use the srcset attribute to control sizing and be done at that, but if for example you’re also serving WebP with JPEG fallbacks for the same image, you may use the <picture> element with multiple sources and source-sets.

If the above sounds complicated for you as it does for me, this guide should help explain the terminology and use-cases.

Unused CSS: Removing imports

If you’re using a CSS framework like Bootstrap, don’t just include all of its components blindly. In fact, I would start with commenting out all of the non-essential components and only add those gradually as the need arises. Here’s a snippet of my bootstrap.scss, where all of its different parts are imported:

// ...

// Components
// ...
@import "bootstrap/dropdowns";
@import "bootstrap/button-groups";
@import "bootstrap/input-groups";
@import "bootstrap/navbar";
// @import "bootstrap/breadcrumbs";
// @import "bootstrap/badges";
// @import "bootstrap/jumbotron";

// Components w/ JavaScript
@import "bootstrap/modals";
@import "bootstrap/tooltip";
@import "bootstrap/popovers";
// @import "bootstrap/carousel";

I don’t use things like badges or jumbotron so I can safely comment those out.

Unused CSS: Purging CSS with PurgeCSS

A more aggressive and more complicated approach is using a library like PurgeCSS, which analyzes your files, detects CSS content that’s not in use, and removes it. PurgeCSS is an NPM package, so if you’re hosting Django on Heroku, you’ll need to install the Node.js buildpack side-by-side with your Python one.


I hope you’ve found at least one area where you can make your Django app faster. If you have any questions, suggestions, or feedback don’t hesitate to drop me a line on Twitter.


Decorator used for QuerySet performance analysis

Below is the code for the django_query_analyze decorator:

from timeit import default_timer as timer
from django.db import connection, reset_queries

def django_query_analyze(func):
    """decorator to perform analysis on Django queries"""

    def wrapper(*args, **kwargs):

        avs = []
        query_counts = []
        for _ in range(20):
            start = timer()
            func(*args, **kwargs)
            end = timer()
            avs.append(end - start)

        print(f"ran function {func.__name__}")
        print(f"-" * 20)
        print(f"number of queries: {int(sum(query_counts) / len(query_counts))}")
        print(f"Time of execution: {float(format(min(avs), '.5f'))}s")
        return func(*args, **kwargs)

    return wrapper

  1. it’s still useful to hold this boolean in the environment