Disqus: Serving 400 million people with Python | Hacker News Reader

Disqus: Serving 400 million people with Python

ontwik.com

21 pointsahmicro9 comments

After reading http://www.yesodweb.com/blog/preliminary-warp-cross-language... I wish it were feasible to implement it again in a more performant language/framework. Would be really fascinating to get solid cost analysis numbers. If you can handle 6x more requests in even Node that could result in a near 6x cost decrease in hardware, is that worth it? I'd love to know.

I'm curious why the downvotes here. Is that information not interesting to anyone else?

I'm not one of the downvoters. But I have a remark: the only Python webserver they looked at in the linked article is Tornado. That's hardly representative, as far as these kind of microbenchmarks are representative in general.

To be fair one of the selling points of Tornado when it came out was Speed. But that isn't really the point, my comment was about understanding how much choosing framework X is or isn't costing you relative to other choices. I'm not saying people should use Warp, I simply read the post recently so it was fresh in my mind.

Python has gevent, which is much faster than Tornado, at least at microbenchmarks (see http://nichol.as/benchmark-of-python-web-servers ). It's probably in the same league as Node.

In the article they also give the reason why Warp is so much faster: it auto-scales to four cores. The Python frameworks don't do this, you have to use multiprocessing for that (or just run an instance for every core).

Again I think you're missing what I was saying. I'm not saying you should chose Node over Tornado or making any kind of value judgement, I'm saying it would be interesting to state the costs of choosing various frameworks to help make better decisions.

On a side note I'm very skeptical of the claim that gevent is in the same league as Node, the numbers in this site are close to the Warp numbers and gevent is still about 6x slower than Node. (Unless I'm reading the microbenchmark wrong). But that's the problem with microbenchmarks, they don't really tell you much.

EDIT: Clarity

I would say it is mainly because network latency dominates response time. It's really cool that Warp can serve static pages that fast but the frameworks start having the same speed when you use a database. Additionally, static pages should be served by an upstream cache which would be faster than Warp.

Amusingly, the comments on that article/benchmark are served by Disqus, which is obviously Python/Django.

It was staggeringly dull. A list of frameworks they've used, but no enthusiasm and no actually saying anything deep about their architecture.

Reply on news.ycombinator.com