Re: Scaling Django

Wed, 03 Feb 2016 11:19:16 -0800

Joshua,

My team is producing a Django app with a small number of
users, so we haven't worried too much about performance yet,
but we know we may have to some day, so I've accumulated a
list of ways to improve performance in a JIRA ticket for if/when
it becomes a priority. 

We've done some of them already, with a quick and easy
thousand-fold increase in speed.

Here's a cut/paste of our ideas.  I hope you find it useful.

Optimize the site for speed

  • CPU speed
    • Do as much as possible in JS on the local device, not in Python/Django on the shared server
  • Disk speed
    • Use the AWS Console to change the disk drives of the server from magnetic
      disks to SSDs.
    • Set up a RAID volume
  • DB speed
    • Do all selects/joins/filtering in the DB. Do not query lots of data
      from the DB and then iterate over it or otherwise filter it in Python.
      Especially, do not query lots of data from the DB, pass it all to the
      browser, and iterate over it or otherwise filter it in _javascript_.
    • Create DB indexes for all fields used frequently in SELECTs
      • But don't over-index
    • Move the DB server to a different Linux server than the
      Web server, to split the load between 2 servers, if the DB
      is the bottleneck. Use Amazon RDS
    • http://www.revsys.com/12days/finding-sources-of-slowness/
      • EXPLAIN PLAN
      • EXPLAIN ANALYZE
  • Ajax
    • Use Ajax partial page loads instead of full page loads
    • Use Ajax to load rest of page while the user already has something to look at.
    • Don't get too chatty
      • Going back to the server on each keystroke for autocomplete is a heavy load. Might be better to wait a second or so after each keystroke to see if the user is still typing, using the same kind of algorithm Fred has used in the past, and that Google uses for auto-suggestions at Google Search.
        • 12/10/2012: Changed to use a 300 ms delay for now.
        • Should perhaps also delay search till 3 chars typed, instead of current
          2 chars
  • Django server-side caching
    • Prime candidate #1. May be costing us at least a couple seconds.
    • Django Template Caching
      • Prime candidate #1a. May be costing us at least a couple seconds.
      • Status: Done. Dramatic improvement.
    • Django page caching
    • Also, "Varnish" as an alternative to standard Django caching. See:
    • Also caching of template pages via template tag {% cache %}
      • Or is this just part of standard Django caching?
    • Could also re-order the INSTALLED_APPS list so that the apps with
      the most frequently loaded templates come first. That would make
      'django.template.loaders.app_directories.Loader' more efficient.
      • Not likely to matter since we're already doing template file
        caching. But if we ever stop using that...
    • Could also give up on the less efficient
      'django.template.loaders.app_directories.Loader', and just use
      'django.template.loaders.filesystem.Loader' which only looks in
      one location, not one per app. Especially since we include the
      app name in the reference to each template already. Could just
      create a single templates folder with app names as subfolders.
      • Not likely to matter since we're already doing template file
        caching. But if we ever stop using that...
  • Web Server speed
    • Apache httpd server
      • Optimal number of processes, threads, connections, open files
    • nginx instead of Apache
  • Caching on the phone
    • We may need to consider some caching of data on the phones
    • Have to be prepared for sluggish performance because some phones will be in areas with poor cell phone reception.
    • Especially for repetitive actions like Event Checkout.
  • Re-write specific pages to be faster
  • Scaling (vertical and horizontal):
    • Prime candidate #5. We really should not need to do this, but it's
      a cheap and easy way to get a quick fix, if none of the other
      prime candidates work well enough.
    • Vertically -- larger servers
    • Horizontally -- more servers
Enjoy!
--Fred
Fred Stluka -- mailto:f...@bristle.com -- http://bristle.com/~fred/
Bristle Software, Inc -- http://bristle.com -- Glad to be of service!
Open Source: Without walls and fences, we need no Windows or Gates.
On 2/3/16 10:30 AM, Joshua Pokotilow wrote:
At the startup where I work, we've written a lot of our server code in Django. So far, we've adopted a "build it fast" mentality, so we invested very little time in optimizing our code. A small amount of load testing has revealed our codebase / infrastructure as it stands today needs to run faster and support more users.

We recently hired some new engineers who are extremely skeptical that we should optimize our existing code. Their main concerns are:

- We need to move to a service-oriented infrastructure because Django is too monolithic (monolithic = technology lock-in & difficult to troubleshoot)
- It's too easy to write slow queries using the Django ORM
- It's hard to hire Django engineers
- While Instagram and DISQUS use Django to service large numbers of people, they don't use it for any serious backend work

After having worked with Django for the last 3 years, I'm a big believer in it, and I believe it would scale. To defend my position, I've pointed out to my colleagues that it's easy to identify bottlenecks with tools like the Django Debug Toolbar and Yet Another Django Profiler. With my colleagues present, I've isolated and fixed significant speed problems inside of a few hours. I don't believe the Django ORM is inherently bad, although I do think that coders who use it should Know What They're Doing. Finally, I've referenced blog entries that talk about how Instagram and Disqus use Django on the backend for backend-y tasks.

Despite my best efforts, my colleagues are still pushing to have us rewrite large portions of our infrastructure as separate services before we try to fix them. For example, we have one slow REST endpoint that returns a boatload of user data, and so there's talk about using a new microservice for users in lieu of our existing Django models. Even if we are able to fix bottlenecks we encounter in a timely fashion, my colleagues fear that Django won't scale with the business.

I'm writing this post to garner additional evidence that Django will scale. Anything compelling (and preferably not obvious) that would help shed some light on Django's ability to scale would be *greatly* appreciated, as it's very difficult for me to defend my position that Django is a viable long-term solution without solid evidence to back up my claims. It certainly doesn't help that I don't have any experience scaling Django myself!

Thank you.
--
You received this message because you are subscribed to the Google Groups "Django users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to django-users+unsubscr...@googlegroups.com.
To post to this group, send email to django-users@googlegroups.com.
Visit this group at https://groups.google.com/group/django-users.
To view this discussion on the web visit https://groups.google.com/d/msgid/django-users/83968c41-d415-4189-b33b-9f99b10b1c41%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "Django users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to django-users+unsubscr...@googlegroups.com.
To post to this group, send email to django-users@googlegroups.com.
Visit this group at https://groups.google.com/group/django-users.
To view this discussion on the web visit https://groups.google.com/d/msgid/django-users/56B25293.3090205%40bristle.com.
For more options, visit https://groups.google.com/d/optout.

Reply via email to