We just wrote some new documentation on how Gitlab uses Unicorn and unicorn-worker-killer, available on doc.gitlab.com but also included below. We would love to hear from the community if you have other questions so we can improve this documentation resource!
Update 19:29 CEST: made link to doc.gitlab.com more specific.
GitLab uses Unicorn, a pre-forking Ruby web server, to handle web requests (web browsers and Git HTTP clients). Unicorn is a daemon written in Ruby and C that can load and run a Ruby on Rails application; in our case the Rails application is GitLab Community Edition or GitLab Enterprise Edition.
Unicorn has a multi-process architecture to make better use of available CPU cores (processes can run on different cores) and to have stronger fault tolerance (most failures stay isolated in only one process and cannot take down GitLab entirely). On startup, the Unicorn 'master' process loads a clean Ruby environment with the GitLab application code, and then spawns 'workers' which inherit this clean initial environment. The 'master' never handles any requests, that is left to the workers. The operating system network stack queues incoming requests and distributes them among the workers.
In a perfect world, the master would spawn its pool of workers once, and then the workers handle incoming web requests one after another until the end of time. In reality, worker processes can crash or time out: if the master notices that a worker takes too long to handle a request it will terminate the worker process with SIGKILL ('kill -9'). No matter how the worker process ended, the master process will replace it with a new 'clean' process again. Unicorn is designed to be able to replace 'crashed' workers without dropping user requests.
This is what a Unicorn worker timeout looks like in
unicorn_stderr.log. The master process has PID 56227 below.
[2015-06-05T10:58:08.660325 #56227] ERROR -- : worker=10 PID:53009 timeout (61s > 60s), killing [2015-06-05T10:58:08.699360 #56227] ERROR -- : reaped #<Process::Status: pid 53009 SIGKILL (signal 9)> worker=10 [2015-06-05T10:58:08.708141 #62538] INFO -- : worker=10 spawned pid=62538 [2015-06-05T10:58:08.708824 #62538] INFO -- : worker=10 ready
The main tunables for Unicorn are the number of worker processes and the request timeout after which the Unicorn master terminates a worker process. See the omnibus-gitlab Unicorn settings documentation if you want to adjust these settings.
GitLab has memory leaks. These memory leaks manifest themselves in long-running processes, such as Unicorn workers. (The Unicorn master process is not known to leak memory, probably because it does not handle user requests.)
To make these memory leaks manageable, GitLab comes with the unicorn-worker-killer gem. This gem monkey-patches the Unicorn workers to do a memory self-check after every 16 requests. If the memory of the Unicorn worker exceeds a pre-set limit then the worker process exits. The Unicorn master then automatically replaces the worker process.
This is a robust way to handle memory leaks: Unicorn is designed to handle workers that 'crash' so no user requests will be dropped. The unicorn-worker-killer gem is designed to only terminate a worker process in between requests, so no user requests are affected.
This is what a Unicorn worker memory restart looks like in unicorn_stderr.log. You see that worker 4 (PID 125918) is inspecting itself and decides to exit. The threshold memory value was 254802235 bytes, about 250MB. With GitLab this threshold is a random value between 200 and 250 MB. The master process (PID 117565) then reaps the worker process and spawns a new 'worker 4' with PID 127549.
[2015-06-05T12:07:41.828374 #125918] WARN -- : #<Unicorn::HttpServer:0x00000002734770>: worker (pid: 125918) exceeds memory limit (256413696 bytes > 254802235 bytes) [2015-06-05T12:07:41.828472 #125918] WARN -- : Unicorn::WorkerKiller send SIGQUIT (pid: 125918) alive: 23 sec (trial 1) [2015-06-05T12:07:42.025916 #117565] INFO -- : reaped #<Process::Status: pid 125918 exit 0> worker=4 [2015-06-05T12:07:42.034527 #127549] INFO -- : worker=4 spawned pid=127549 [2015-06-05T12:07:42.035217 #127549] INFO -- : worker=4 ready
One other thing that stands out in the log snippet above, taken from Gitlab.com, is that 'worker 4' was serving requests for only 23 seconds. This is a normal value for our current GitLab.com setup and traffic.
The high frequency of Unicorn memory restarts on some GitLab sites can be a source of confusion for administrators. Usually they are a red herring.