High Performance Django
Uploaded by
Best Tech VideosHigh Performance Django
Uploaded by
Best Tech VideosHigh Performance Django
David Cramer
http://www.davidcramer.net/
http://www.ibegin.com/
Curse
• Peak daily traffic of approx. 15m pages, 150m hits.
• Average monthly traffic 120m pages, 6m uniques.
• Python, MySQL, Squid, memcached, mod_python, lighty.
• Most developers came strictly from PHP (myself included).
• 12 web servers, 4 database servers, 2 squid caches.
iBegin
• Massive amounts of data, 100m+ rows.
• Python, PHP, MySQL, mod_wsgi.
• Small team of developers.
• Complex database partitioning/synchronization tasks.
• Attempting to not branch off of Django.
Areas of Concern
• Database (ORM)
• Webserver (Resources, Handling Millions of Reqs)
• Caching (Invalidation, Cache Dump)
• Template Rendering (Logic Separation)
• Profiling
Tools of the Trade
• Webserver (Apache, Nginx, Lighttpd)
• Object Cache (memcached)
• Database (MySQL, PostgreSQL, …)
• Page Cache (Squid, Nginx, Varnish)
• Load Balancing (Nginx, Perlbal)
How We Did It
• “Primary” web servers serving Django using mod_python.
• Media servers using Django on lighttpd.
• Static served using additional instances of lighttpd.
• Load balancers passing requests to multiple Squids.
• Squids passing requests to multiple web servers.
Lessons Learned
• Don’t be afraid to experiment. You’re not limited to a one.
• mod_wsgi is a huge step forward from mod_python.
• Serving static files using different software can help.
• Send proper HTTP headers where they are needed.
• Use services like S3, Akamai, Limelight, etc..
Webserver Software
Python Scripts Static Content
• Apache (wsgi, mod_py, • Apache
fastcgi) • Lighttpd
• Lighttpd (fastcgi) • Tinyhttpd
• Nginx (fastcgi) • Nginx
Reverse Proxies Software Load Balancers
• Nginx • Nginx
• Squid • Perlbal
• Varnish
Database (ORM)
• Won’t make your queries efficient. Make your own indexes.
• select_related() can be good, as well as bad.
• Inherited ordering (Meta: ordering) will get you.
• Hundreds of queries on a page is never a good thing.
• Know when to not use the ORM.
Handling JOINs
class Category(models.Model):
name = models.CharField()
created_by = models.ForeignKey(User)
class Poll(models.Model):
name = models.CharField()
category = models.ForeignKey(Category)
created_by = models.ForeignKey(User)
# We need to output a page listing all Poll's with
# their name and category's name.
def a_bad_example(request):
# We have just caused Poll to JOIN with User and Category,
# which will also JOIN with User a second time.
my_polls = Poll.objects.all().select_related()
return render_to_response('polls.html', locals(), request)
def a_good_example(request):
# Use select_related explicitly in each case.
poll = Poll.objects.all().select_related('category')
return render_to_response('polls.html', locals(), request)
Template Rendering
• Sandboxed engines are typically slower by nature.
• Keep logic in views and template tags.
• Be aware of performance in loops, and groupby (regroup).
• Loaded templates can be cached to avoid disk reads.
• Switching template engines is easy, but may not give you
any worthwhile performance gain.
Template Engines
Caching
• Two flavors of caching: object cache and browser cache.
• Django provides built-in support for both.
• Invalidation is a headache without a well thought out plan.
• Caching isn’t a solution for slow loading pages or improper indexes.
• Use a reverse proxy in between the browser and your web servers:
Squid, Varnish, Nginx, etc..
Cache With a Plan
• Build your pages to use proper cache headers.
• Create a plan for object cache expiration, and invalidation.
• For typical web apps you can serve the same cached page
for both anonymous and authenticated users.
• Contain commonly used querysets in managers for
transparent caching and invalidation.
Cache Commonly Used Items
def my_context_processor(request):
# We access object_list every time we use our context processors so
# it makes sense to cache this, no?
cache_key = ‘mymodel:all’
object_list = cache.get(cache_key)
if object_list is None:
object_list = MyModel.objects.all()
cache.set(cache_key, object_list)
return {‘object_list’: object_list}
# Now that we are caching the object list we are going to want to invalidate it
class MyModel(models.Model):
name = models.CharField()
def save(self, *args, **kwargs):
super(MyModel, self).save(*args, **kwargs)
# save it before you update the cache
cache.set(‘mymodel:all’, MyModel.objects.all())
Profiling Code
• Finding the bottleneck can be time consuming.
• Tools exist to help identify common problematic areas.
– cProfile/Profile Python modules.
– PDB (Python Debugger)
Profiling Code With cProfile
import sys
try: import cProfile as profile
except ImportError: import profile
try: from cStringIO import StringIO
except ImportError: import StringIO
from django.conf import settings
class ProfilerMiddleware(object):
def can(self, request):
return settings.DEBUG and 'prof' in request.GET and (not settings.INTERNAL_IPS or request.META['REMOTE_ADDR'] in
settings.INTERNAL_IPS)
def process_view(self, request, callback, callback_args, callback_kwargs):
if self.can(request):
self.profiler = profile.Profile()
args = (request,) + callback_args
return self.profiler.runcall(callback, *args, **callback_kwargs)
def process_response(self, request, response):
if self.can(request):
self.profiler.create_stats()
out = StringIO()
old_stdout, sys.stdout = sys.stdout, out
self.profiler.print_stats(1)
sys.stdout = old_stdout
response.content = '<pre>%s</pre>' % out.getvalue()
return response
http://localhost:8000/?prof
Profiling Database Queries
from django.db import connection
class DatabaseProfilerMiddleware(object):
def can(self, request):
return settings.DEBUG and 'dbprof' in request.GET \
and (not settings.INTERNAL_IPS or \
request.META['REMOTE_ADDR'] in settings.INTERNAL_IPS)
def process_response(self, request, response):
if self.can(request):
out = StringIO()
out.write('time\tsql\n')
total_time = 0
for query in reversed(sorted(connection.queries, key=lambda x: x['time'])):
total_time += float(query['time'])*1000
out.write('%s\t%s\n' % (query['time'], query['sql']))
response.content = '<pre style="white-space:pre-wrap">%d queries executed in %.3f seconds\n\n%s</pre>' %
(len(connection.queries), total_time/1000, out.getvalue())
return response
http://localhost:8000/?dbprof
Summary
• Database efficiency is the typical problem in web apps.
• Develop and deploy a caching plan early on.
• Use profiling tools to find your problematic areas. Don’t pre-
optimize unless there is good reason.
• Find someone who knows more than me to configure your
server software.
Thanks!
Slides and code available online at:
http://www.davidcramer.net/djangocon
