Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
HTTP for Great Good
Search
Sponsored
·
Ship Features Fearlessly
Turn features on and off without deploys. Used by thousands of Ruby developers.
→
Matt Robenolt
September 05, 2013
Programming
200k
85
Share
HTTP for Great Good
Scaling Django with HTTP
DjangoCon US 2013
http://www.youtube.com/watch?v=HAjOQ09I1UY
Matt Robenolt
September 05, 2013
More Decks by Matt Robenolt
See All by Matt Robenolt
Everything is broken and I don't know why.
mattrobenolt
0
44
I am bad at my job.
mattrobenolt
0
240
Everything is broken, and I don't know why. Python edition.
mattrobenolt
1
230
Everything is broken, and I don't know why. Python edition.
mattrobenolt
2
630
Varnish: How We Do It
mattrobenolt
1
240
Everything is broken, and I don't know why.
mattrobenolt
7
1.6k
Cheating Your Way to Webscale
mattrobenolt
13
1.4k
Caching is Hard: Varnish @ Disqus
mattrobenolt
52
2.1M
Developing & Deploying "Large" Scale Web Applications
mattrobenolt
25
1.3k
Other Decks in Programming
See All in Programming
Augmenting AI with the Power of Jakarta EE
ivargrimstad
0
400
運用エージェントは "作る" から "育てる" へ - 記憶と自己進化の3層設計パターン / self-evolving-agents-three-layer-agent-design
gawa
12
3.3k
Java × distroless で 軽量なコンテナイメージを / Java on Distroless
contour_gara
0
420
JJUG CCC 2026 Spring: JSpecify で実現する Kotlin フレンドリーな Java API 設計
ternbusty
1
110
プロパティの順序で型推論が壊れる!? TypeScript6.0の修正からContext-Sensitivityの仕組みを追う
bicstone
2
1.3k
AIとRubyの静的型付け
ukin0k0
0
470
TypeScriptだけでAIエージェントを作る フロント・エージェント・インフラのフルスタック実践
har1101
6
1.2k
今さら聞けないCancellationToken
htkym
0
200
TSKaigi2026-静的解析への投資がAI時代のコード品質を支える ── カスタムESLintルールの設計と運用
hayatokudou
7
1.3k
関係性から理解する"同一性"の型用語たち
pvcresin
2
620
サーバーレスで作る、動画データ管理基盤
oyasumipants
0
320
ユニットテストの先へ:テスト技法で要求・仕様を整理するJava開発実践 / Beyond_Unit_Testing_Practical_Java_Development_Techniques_for_Organizing_Requirements_and_Specifications
shimashima35
0
310
Featured
See All Featured
State of Search Keynote: SEO is Dead Long Live SEO
ryanjones
0
200
Bootstrapping a Software Product
garrettdimon
PRO
307
120k
Navigating Team Friction
lara
192
16k
A Guide to Academic Writing Using Generative AI - A Workshop
ks91
PRO
1
310
Build your cross-platform service in a week with App Engine
jlugia
234
18k
16th Malabo Montpellier Forum Presentation
akademiya2063
PRO
0
130
Visual Storytelling: How to be a Superhuman Communicator
reverentgeek
2
540
XXLCSS - How to scale CSS and keep your sanity
sugarenia
250
1.3M
SEO for Brand Visibility & Recognition
aleyda
0
4.6k
DBのスキルで生き残る技術 - AI時代におけるテーブル設計の勘所
soudai
PRO
65
55k
A Modern Web Designer's Workflow
chriscoyier
698
190k
Being A Developer After 40
akosma
91
590k
Transcript
HTTP for Great Good Scaling Django with HTTP DjangoCon US
September 5th 2013 Matt Robenolt
Hello < me irl
Site Reliability Engineer
“DJANGO ALL THE THINGS!”
“...but
“...but
The slowest part of a web application is typically not
your code.
Between databases and memcaches and Redises and Cassandras and MongoDBs
and networks, Django is not the problem.
“...everything
None
A few vanity metrics.
Monthly Unique Visitors 1,115,080,411
Monthly Page Views 7,516,761,301
Inbound Traffic 42k total req/s 15k app req/s not my
fault
36% of all requests actually hit a Django server 15k/42k
= 36%
...what happened to the other 64%?
Let’s talk about HTTP. Hypertext Transport Protocol
$ curl -v disqus.com
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Request
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Method
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Path
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Version
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Headers
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Response
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Status
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Headers
Request in Django request.method # method request.get_full_path() # path request.META['HTTP_USER_AGENT']
request.META['HTTP_ACCEPT']
Response in Django response = HttpResponse(body) response.status_code = 200 response['X-Foo']
= 'bar'
“Cool,
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > < HTTP/1.1 200 OK < Server: nginx < Date: Fri, 30 Aug 2013 06:38:37 GMT < Content-Type: text/html; charset=utf-8 < Content-Length: 10453 < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 06:38:36 GMT < Cache-Control: no-cache, must-revalidate Hmm. What can we do with this information?
> GET / HTTP/1.1 > User-Agent: curl/7.24.0 > Host: disqus.com
> Accept: */* > If-Modified-Since: Fri, 30 Aug 2013 00:32:14 GMT > < HTTP/1.1 304 Not Modified < Server: nginx < Date: Fri, 30 Aug 2013 18:05:38 GMT < Last-Modified: Fri, 30 Aug 2013 00:32:14 GMT < Vary: Accept-Encoding < Expires: Fri, 30 Aug 2013 18:05:37 GMT < Cache-Control: no-cache, must-revalidate
304 Not Modified No body is sent with the response
304 Not Modified Client reuses its cached version
304 Not Modified Usually more efficient to calculate
notbad.gif
But we can do better!
Static files have been promoting good practices for a long
time.
“Far future headers”
$ curl -v a.disquscdn.com/dotcom/d-6203c8f/css/ disqus-web/pages/home.css < HTTP/1.1 200 OK <
Server: nginx < Content-Type: text/css; charset=utf-8 < Last-Modified: Fri, 16 Aug 2013 20:31:05 GMT < Expires: Sun, 15 Sep 2013 20:34:21 GMT < Cache-Control: max-age=2592000 < Content-Length: 30749 < Date: Sun, 18 Aug 2013 03:23:37 GMT < Via: 1.1 varnish < Age: 110956 < Connection: keep-alive < Vary: Accept-Encoding
$ curl -v a.disquscdn.com/dotcom/d-6203c8f/css/ disqus-web/pages/home.css < HTTP/1.1 200 OK <
Server: nginx < Content-Type: text/css; charset=utf-8 < Last-Modified: Fri, 16 Aug 2013 20:31:05 GMT < Expires: Sun, 15 Sep 2013 20:34:21 GMT < Cache-Control: max-age=2592000 < Content-Length: 30749 < Date: Sun, 18 Aug 2013 03:23:37 GMT < Via: 1.1 varnish < Age: 110956 < Connection: keep-alive < Vary: Accept-Encoding 30 days in the future.
Chrome Web Inspector on second visit
No HTTP request 0ms
No HTTP request The client knew to just use its
local cache
No HTTP request Computer actually did something right for once
“I SEE WHAT YOU DID THERE” - Hopefully you
Takeaways Clients behave differently depending on the response headers
Takeaways These usually come with minimal effort with static files
Takeaways We can and should utilize these to our advantage
to improve UX
Same logic can be applied to dynamic content.
What’s this look like in Django?
Last-Modified def lol(request): response = render(request, 'lol.html') response['Last-Modified'] = \
'Fri, 16 Aug 2013 20:31:05 GMT' return response * don’t do this.
Last-Modified from django.views.decorators.http import \ last_modified def post_last_modified(request, slug): return
Post.objects.get(slug=slug).modified @last_modified(post_last_modified) def blog_post_detail(request, slug): # Your view
Cache-Control def lol(request): response = render(request, 'lol.html') response['Cache-Control'] = 'max-age=600'
return response
Cache-Control from django.views.decorators.cache import \ cache_control @cache_control(max_age=600) # Cache for
10m def home(request): return HttpResponse('lol')
“OMG!
None
Well... not really.
How many requests will {{user}} make to the same page
within 10 minutes?
1? 2? 3?
...out of 42k requests per second.
Even with caching, your app is doing a lot of
work.
Parsing HTTP.
WSGI.
Django middleware stack.
Do some stuff.
Render a template?
Back out through the Django middleware.
Transform an HttpResponse into a real HTTP response.
...at 42k requests per second.
You’re gonna have a bad time. me
Until now, “client” has been a user’s browser.
“If only we could utilize this Cache-Control stuff better...”
Introducing
$ apt-get install varnish
$ brew install varnish
tl;dr Varnish sits between Django and your users Internet
tl;dr Caches HTTP responses and respects proper HTTP headers Internet
tl;dr Its sole purpose in life is to be a
cache, so it’s really fast. Internet
Stand back, science is happening.
Stand back, science is happening. benchmarking
Simple, non-scientific “Hello World”
from django.views.decorators.cache import \ cache_control from django.http import HttpResponse @cache_control(max_age=5)
def hello(request): return HttpResponse('Hello world') “Hello World”
$ httperf --server 127.0.0.1 --port 8000 -- uri /hello/ --rate
150 --num-conn 10 --num-call 500 --hog Request rate: 369.6 req/s (2.7 ms/req) Django + gunicorn * on my MacBook Air
$ httperf --server 127.0.0.1 --port 8888 -- uri /hello/ --rate
150 --num-conn 10 --num-call 10000 --hog Request rate: 15633.4 req/s (0.1 ms/req) Varnish * on my MacBook Air
Varnish: How does it work?
First request
First response “Lemme
“Yo,
Next response “wut
Caching: ProMoves™
Augment with JavaScript Update your UI optimistically
Augment with JavaScript Leverage cookies to store non-critical data
Augment with JavaScript Defer fetching user-specific data until needed
Short TTLs are good Most things can be cached for
at least 5s
Short TTLs are good At 10k requests/s, a 5s TTL
absorbs 49,999 requests
Let’s meet: John and Jane Doe.
John and Jane are different users.
John logs into Disqus.
Jane logs into Disqus.
Jane sees John’s stuff.
Jane sees John’s stuff. ^ not
We really want to avoid this from ever happening.
Cookies
How do users even work?
$ curl -vd "username=foo&password=bar" https:// disqus.com/profile/login/ > POST /profile/login/ HTTP/1.1
> User-Agent: curl/7.24.0 > Host: disqus.com > < HTTP/1.1 302 FOUND < Server: nginx < Date: Fri, 30 Aug 2013 21:34:36 GMT < Vary: Cookie < Set-Cookie: sessionid=f7aa9598-11bb-11e3-9eb1-003048d9a288; Domain=.disqus.com; expires=Sun, 29-Sep-2013 21:34:36 GMT; httponly; Max-Age=2592000; Path=/
$ curl -vd "username=foo&password=bar" https:// disqus.com/profile/login/ > POST /profile/login/ HTTP/1.1
> User-Agent: curl/7.24.0 > Host: disqus.com > < HTTP/1.1 302 FOUND < Server: nginx < Date: Fri, 30 Aug 2013 21:34:36 GMT < Vary: Cookie < Set-Cookie: sessionid=f7aa9598-11bb-11e3-9eb1-003048d9a288; Domain=.disqus.com; expires=Sun, 29-Sep-2013 21:34:36 GMT; httponly; Max-Age=2592000; Path=/ Set-Cookie
$ curl -vd "username=foo&password=bar" https:// disqus.com/profile/login/ > POST /profile/login/ HTTP/1.1
> User-Agent: curl/7.24.0 > Host: disqus.com > < HTTP/1.1 302 FOUND < Server: nginx < Date: Fri, 30 Aug 2013 21:34:36 GMT < Vary: Cookie < Set-Cookie: sessionid=f7aa9598-11bb-11e3-9eb1-003048d9a288; Domain=.disqus.com; expires=Sun, 29-Sep-2013 21:34:36 GMT; httponly; Max-Age=2592000; Path=/ Session Id
Unique id that represents a logged in user Session Id
django.contrib.sessions / django.contrib.auth Session Id
Could potentially cache per session id Session Id
By default, Varnish will not cache any request with a
Cookie header at all.
Think about if your endpoint changes based on a user’s
authentication.
If it doesn’t, Varnish can normalize it.
None
Learn: Varnish Configuration Language (VCL) in 30 seconds
sub vcl_recv { // These urls can be stripped of
all // cookies since they serve the same // data for anon and auth'd user if ( req.url == "/" || req.url ~ "^/embed/comments/" ) { unset req.http.Cookie; } }
Basically, caching is hard.
Go make some stuff faster.
We’re hiring people who hate computers. disqus.com/jobs
Questions? I have answers. ^ github.com/mattrobenolt @mattrobenolt some