Fortunately, I had the chance to experiment it's beta release in a controlled Vagrant environment and check those claim! Bear in mind that those are synthetic benchmarks and real world throughput of real applications should give us different behavior. With that having being said, let's see how the main Ruby servers perform a very simple Rack app returning a simple HTTP 200 with "Hello World":
==> Benchmark parameters: Application : hello_app Operating system : Ubuntu 14.04 LTS (x86_64) Virtual CPU cores : 4 MRI Ruby : ruby 2.1.3p242 (2014-09-19 revision 45877) [x86_64-linux-gnu] JRuby : jruby 1.7.13 (1.9.3p392) 2014-06-24 43f133c on OpenJDK 64-Bit Server VM 1.7.0_65-b32 [linux-amd64] Unicorn workers : 8 Puma workers (MRI) : 8 (16 threads each) Puma workers (JRuby): 1 (32 threads each) Torquebox threads : 32 Raptor workers : 8 Concurrent clients : 16 ==> Benchmark summary: Unicorn (MRI Ruby) : 23015.36 req/sec Puma (MRI Ruby) : 32538.62 req/sec Puma (JRuby) : 399.14 req/sec Torquebox (JRuby) : 29773.21 req/sec Raptor (MRI Ruby) : 95309.35 req/sec Raptor (JRuby) : 87523.65 req/sec
Now, this sounds very impressive indeed! And what about a more complex example: a simple Rails 4.1.6 application rendering a default scaffold index page with a model fetching 20 rows from MySQL:
==> Benchmark summary: Unicorn (MRI Ruby) : 326.49 req/sec Puma (MRI Ruby) : 327.36 req/sec Puma (JRuby) : 221.78 req/sec Torquebox (JRuby) : 226.57 req/sec Raptor (MRI Ruby) : 79617.78 req/sec Raptor (JRuby) : 73948.59 req/sec
Again, super impressive. My first impression when reading these numbers is that Raptor must have builtin response caching (which is great!). So I tweaked the example to make it much heavier than what's considered "normal" for a small cache, rendering a table of 100 rows from the database in the index page, and the results are still competitive:
==> Benchmark summary: Unicorn (MRI Ruby) : 85.98 req/sec Puma (MRI Ruby) : 89.93 req/sec Puma (JRuby) : 79.42 req/sec Torquebox (JRuby) : 77.98 req/sec Raptor (MRI Ruby) : 82.63 req/sec Raptor (JRuby) : 88.92 req/sec
This could mean that at the best case scenario, you could get almost 4 times the throughput from your application, and in the worst case scenario, you're still getting similar throughputs. This means that Raptor is adaptable and reasonably "smart". So overall it's a win-win situation!
The way they are able to achieve these superior numbers can be explained by the way they approached the implementation. This is the breakdown that they have released so far:
- Protection against slow clients: Slow clients on the Internet can block your app, like people standing still in front of your door. Unicorn only has a single small door, so it requires you to attach a buffering reverse proxy in front of it, like Nginx. Puma has as many small doors as you configure threads, so you still need to attach Nginx in front to be safe. But Raptor has a door that's almost infinitely wide, and fully protects your app against slow clients without additional software.
- Optional integration with Nginx. Nginx is a great web server and provides useful features like compression, security, etc. But using Puma and Unicorn with Nginx requires you to configure Puma/Unicorn separately, to configure Nginx separately, to configure monitoring separately, and to manage each component separately. Doesn't it make sense to consolidate all these moving parts into a single package? Enter Raptor's optional Nginx integration mode: configuration, management and daemon monitoring all directly from Nginx.
- Multitenancy. If you run multiple apps on a same server, then managing all those Puma and Unicorn instances quickly becomes tedious. Raptor's Nginx integration mode allows you to manage all your apps from a single Nginx instance.
- Security. Improve your server's security by easily running multiple apps under multiple user accounts. This way, a vulnerability in one app won't affect all your other apps.
- JRuby support. Like Puma (and unlike Unicorn), Raptor has great JRuby support.
So it's not only a different implementation but also features for the future, supporting transparently handling slow clients without causing bottlenecks to your application and avoiding having extra layers of protection. And everything in an easy to use package (according to their release: you will be able to replace your current server in the Gemfile and bundle install it!)
Raptor is about to be released, so keep tuned for more news and when it's available for everybody to test their applications with. From what we have seen so far, this one looks like a winner!