The GitLab Rails application code suffers from memory leaks. For web requests this problem is made manageable using unicorn-worker-killer which restarts Unicorn worker processes in between requests when needed. The Sidekiq MemoryKiller applies the same approach to the Sidekiq processes used by GitLab to process background jobs.
Unlike unicorn-worker-killer, which is enabled by default for all GitLab installations since GitLab 6.4, the Sidekiq MemoryKiller is enabled by default only for Omnibus packages. The reason for this is that the MemoryKiller relies on Runit to restart Sidekiq after a memory-induced shutdown and GitLab installations from source do not all use Runit or an equivalent.
With the default settings, the MemoryKiller will cause a Sidekiq restart no more often than once every 15 minutes, with the restart causing about one minute of delay for incoming background jobs.
The MemoryKiller is controlled using environment variables.
SIDEKIQ_MEMORY_KILLER_MAX_RSS: if this variable is set, and its value is greater than 0, then after each Sidekiq job, the MemoryKiller will check the RSS of the Sidekiq process that executed the job. If the RSS of the Sidekiq process (expressed in kilobytes) exceeds SIDEKIQ_MEMORY_KILLER_MAX_RSS, a delayed shutdown is triggered. The default value for Omnibus packages is set in the omnibus-gitlab repository.
SIDEKIQ_MEMORY_KILLER_GRACE_TIME: defaults to 900 seconds (15 minutes). When a shutdown is triggered, the Sidekiq process will keep working normally for another 15 minutes.
SIDEKIQ_MEMORY_KILLER_SHUTDOWN_WAIT: defaults to 30 seconds. When the grace time has expired, the MemoryKiller tells Sidekiq to stop accepting new jobs. Existing jobs get 30 seconds to finish. After that, the MemoryKiller tells Sidekiq to shut down, and an external supervision mechanism (e.g. Runit) must restart Sidekiq.