Scalable Django Architecture
•
42 likes•12,490 views
Rami Sayar presented on advanced Django architecture techniques for scaling Django applications in production environments. He discussed load balancing Django across multiple servers using solutions like HAProxy. He also covered caching with Varnish to improve performance, and using Redis or Memcached for caching. Finally, he discussed asynchronous task queues like Celery for processing tasks outside the request-response cycle.
Report
Share
Scalable Django Architecture
- 1. Advanced Django Architecture Rami Sayar - @ramisayar Technical Evangelist Microsoft Canada
- 2. Advanced Scalable Django Architecture Rami Sayar - @ramisayar Technical Evangelist Microsoft Canada
- 3. Agenda • Advanced Django Techniques & Patterns • Structuring Complex Applications • Scaling Django Applications • Django Applications in Production CONFOO - @RAMISAYAR
- 4. What is a production web app? • Django is surrounded by tons of applications, services and code in production. • Advanced Django architecture is needed to support Django in production without failing hard. CONFOO - @RAMISAYAR
- 5. Load Balancing Django CONFOO - @RAMISAYAR
- 6. Load Balancing • Load balancing Django improves the performance and reliability by distributing traffic. • Becomes an interesting problem when you have to consider: • SSL Termination • Sticky Sessions? • WebSockets? CONFOO - @RAMISAYAR
- 7. Load Balancing = Django + HAProxy + SSL • HAProxy gained native SSL support in HAProxy 1.5.x => June 2014! • SSL Termination: • Fairly straightforward to set up HAProxy. • Read: How To Implement SSL Termination With HAProxy CONFOO - @RAMISAYAR
- 8. Load Balancing Proxying & Remote Addr class SetRemoteAddrFromForwardedFor(object): def process_request(self, request): try: real_ip = request.META['HTTP_X_FORWARDED_FOR'] except KeyError: pass else: # HTTP_X_FORWARDED_FOR can be a comma-separated list of IPs. # Take just the first one. real_ip = real_ip.split(",")[0] request.META['REMOTE_ADDR'] = real_ip CONFOO - @RAMISAYAR
- 9. Azure Traffic Manager • High Performance Load Balancer + DNS Support. • Sticky Sessions don’t exist. • HTTPS Supported. CONFOO - @RAMISAYAR
- 10. Caching Django CONFOO - @RAMISAYAR
- 11. Caching = Django + Varnish • Varnish has a tendency to just cache EVERYTHING which can be frustrating for the User, unless you set it up properly. • Varnish does not cache content & cookies. • Varnish and Django CSRF protection gotcha. • Varnish does not deal with HTTPS (hence the need of HAProxy in front of Varnish). CONFOO - @RAMISAYAR
- 12. Caching = Django + Varnish • Django-Varnish – “It allows you to monitor certain models and when they are updated, Django Varnish will purge the model's absolute_url on your frontend(s). This ensures that object detail pages are served blazingly fast and are always up to date.” • https://github.com/justquick/django-varnish • Project Idea: Deeper Django/Varnish integration. CONFOO - @RAMISAYAR
- 13. Caching = Django + Varnish • https://www.varnish-cache.org/ • http://www.isvarnishworking.com/ • http://yml-blog.blogspot.ca/2010/01/esi-using-varnish-and- django.html • http://chase-seibert.github.io/blog/2011/09/23/varnish-caching- for-unauthenticated-django-views.html • http://blog.bigdinosaur.org/adventures-in-varnish/ • http://www.nedproductions.biz/wiki/a-perfected-varnish-reverse- caching-proxy-vcl-script CONFOO - @RAMISAYAR
- 14. Serving Django CONFOO - @RAMISAYAR
- 15. Web Server = Django + Apache • Forced to use Apache? Your options: • mod_python (No) • mod_wsgi (Yes) • FastCGI (Deprecated since 1.7) • Phusion Passenger (Yes) CONFOO - @RAMISAYAR
- 16. Web Server = Django + Apache • mod_wsgi • Decent performance on the Apache side. • You can setup Apache any way you like. • You can setup mod_wsgi to use virtualenv. CONFOO - @RAMISAYAR
- 17. Web Server = Django + Apache import os, sys, site # Add the site-packages of the chosen virtualenv to work with site.addsitedir('~/.virtualenvs/myprojectenv/local/lib/python2.7/site-packages') # Add the app's directory to the PYTHONPATH sys.path.append('/home/django_projects/MyProject') sys.path.append('/home/django_projects/MyProject/myproject') os.environ['DJANGO_SETTINGS_MODULE'] = 'myproject.settings' # Activate your virtual env activate_env=os.path.expanduser("~/.virtualenvs/myprojectenv/bin/activate_this.py") execfile(activate_env, dict(__file__=activate_env)) import django.core.handlers.wsgi application = django.core.handlers.wsgi.WSGIHandler() CONFOO - @RAMISAYAR
- 18. Web Server = Django + Alternatives (aka. Nginx) • Nginx + uWSGI [uwsgi] touch-reload = /tmp/newproject socket = 127.0.0.1:3031 workers = 2 chdir = /srv/newproject env = DJANGO_SETTINGS_MODULE=newproject.settings module = django.core.handlers.wsgi:WSGIHander() CONFOO - @RAMISAYAR
- 19. Web Server = Django + Alternatives (aka. Gunicorn) • Nginx & Gunicorn • In theory, you should use Nginx as a reverse proxy here, to serve static files if you’re not uploading everything to a CDN or a separate media server. If you are, you can use HAProxy and Gunicorn as a standalone webserver, one less dependency to care about. • Install Gunicorn directly inside your virtual environment. • Use Gaffer to monitor the Gunicorn. CONFOO - @RAMISAYAR
- 20. Web Server - References • http://blog.kgriffs.com/2012/12/18/uwsgi-vs-gunicorn-vs-node- benchmarks.html CONFOO - @RAMISAYAR
- 21. REST APIs & Django CONFOO - @RAMISAYAR
- 22. Use Django REST Framework! Django REST Framework vs Django Tasty Pie CONFOO - @RAMISAYAR
- 23. Cache & Django CONFOO - @RAMISAYAR
- 24. Caching in Production • Caching – Django has multiple options: • Redis. Make sure you use Hiredis and django-redis-cache. Use a hosted Redis Cache (Azure, AWS, etc). • Memcached. Make sure you use pylibmc and not python-memcached, better performance with the C library. • SSL still matters. CONFOO - @RAMISAYAR
- 25. Azure Redis Cache CONFOO - @RAMISAYAR
- 26. Logging Django CONFOO - @RAMISAYAR
- 27. Logging in Production • Logstash • Collect logs, parse them, and store them for later use. • Free and open source. • Set it up with Redis. CONFOO - @RAMISAYAR
- 28. Logging in Production – Use Python-Logstash LOGGING = { 'handlers': { 'logstash': { 'level': 'DEBUG', 'class': 'logstash.LogstashHandler', 'host': 'localhost', 'port': 5959, # Default value: 5959 'version': 1, # Version of logstash event schema. Default value: 0 (for backward compatibility of the library) 'message_type': 'logstash', # 'type' field in logstash message. Default value: 'logstash'. 'tags': ['tag1', 'tag2'], # list of tags. Default: None. }, }, CONFOO - @RAMISAYAR
- 29. Logging in Production – Use Python-Logstash 'loggers': { 'django.request': { 'handlers': ['logstash'], 'level': 'DEBUG', 'propagate': True, }, }, } CONFOO - @RAMISAYAR
- 30. Logging in Production – Using Sentry • Realtime event logging and aggregation platform. • Specialize in monitoring exceptions and errors. • https://github.com/getsentry/sentry CONFOO - @RAMISAYAR
- 31. PostgreSQL in Production CONFOO - @RAMISAYAR
- 32. PostgreSQL in Production – pgpool & slony • PostgreSQL 9+ has streaming replication assuming you’re running identical databases with identical version on identical architectures (not really an issue). • If not using the Streaming Replication feature (WAL), you’ll want to use either pgpool or slony to scale your PostgreSQL database across multiple machines. • Use Slony if you want master-slave-peer replication. • Use pgpool or pgbouncer for connection pooling. Pgbouncer only does connection pooling. CONFOO - @RAMISAYAR
- 33. Async Tasks & Django CONFOO - @RAMISAYAR
- 34. Async Tasks – Celery • Use Celery for Asynchronous Tasks • Use Celery with Redis and store results in Django ORM • RQ is an alternative for distributing tasks. (Based on Redis) Read: http://www.caktusgroup.com/blog/2014/09/29/celery-production/ CONFOO - @RAMISAYAR
- 35. On Redis • Redis is extremely popular as an easy and simple publish- subscribe. • Redis is very particular about memory. Running out of memory in Redis -> CRASH AND BURN! • If you don’t have enough memory in your VM to run Redis Cache -> Change your approach, use Apache Kafka. CONFOO - @RAMISAYAR
- 36. Tips & Techniques • Don’t forget: • Turn off Debug Mode • Turn off Template Debug Mode • Use different Django Settings Modules depending on the role and environment you are in through environment variables by setting “DJANGO_SETTINGS_MODULE”. CONFOO - @RAMISAYAR
- 37. What did we learn? • Scaling Django Architectures • Most of the work surrounds Django • Minimal changes to your django app for the most part. CONFOO - @RAMISAYAR
- 38. Thank You! Questions? Follow @ramisayar CONFOO - @RAMISAYAR
- 39. Resources • http://blog.disqus.com/post/62187806135/scaling-django-to-8- billion-page-views • http://engineering.hackerearth.com/2013/10/07/scaling- database-with-django-and-haproxy/ • http://www.djangobook.com/en/2.0/chapter12.html • https://www.digitalocean.com/community/tutorials/how-to- implement-ssl-termination-with-haproxy-on-ubuntu-14-04 • https://docs.djangoproject.com/en/1.7/topics/security/ • https://www.digitalocean.com/community/tutorials/how-to-scale- django-beyond-the-basics CONFOO - @RAMISAYAR
- 40. ©2013 Microso- Corpora1on. All rights reserved. Microso-, Windows, Office, Azure, System Center, Dynamics and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The informa1on herein is for informa1onal purposes only and represents the current view of Microso- Corpora1on as of the date of this presenta1on. Because Microso- must respond to changing market condi1ons, it should not be interpreted to be a commitment on the part of Microso-, and Microso- cannot guarantee the accuracy of any informa1on provided a-er the date of this presenta1on. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.
Editor's Notes
- What does a web application look like in production?
- A few small complications, one of which is that every request’s remote IP (request.META["REMOTE_IP"]) will be that of the load balancer, not the actual IP making the request. Load balancers deal with this by setting a special header, X-Forwarded-For, to the actual requesting IP address. Use Middleware.
- A few small complications, one of which is that every request’s remote IP (request.META["REMOTE_IP"]) will be that of the load balancer, not the actual IP making the request. Load balancers deal with this by setting a special header, X-Forwarded-For, to the actual requesting IP address. Use Middleware.