• Configure Puma
• Puma Worker Killer
• Worker timeout
• Memory-constrained environments
• Performance caveat when using Puma with Rugged
• Convert Unicorn settings to Puma

Puma

Puma is a simple, fast, multi-threaded, and highly concurrent HTTP 1.1 server for Ruby applications. It’s the default GitLab web server since GitLab 13.0 and has replaced Unicorn. From GitLab 14.0, Unicorn is no longer supported.

Configure Puma

To configure Puma:

1. Determine suitable Puma worker and thread settings.
2. If you’re switching from Unicorn, convert any custom settings to Puma.
3. For multi-node deployments, configure the load balancer to use the readiness check.

4. Reconfigure GitLab so the above changes take effect:

   sudo gitlab-ctl reconfigure
   

For Helm-based deployments, see the webservice chart documentation.

For more details about the Puma configuration, see the Puma documentation.

Puma Worker Killer

By default:

• The Puma Worker Killer restarts a worker if it exceeds a memory limit.
• Rolling restarts of Puma workers are performed every 12 hours.

To change the memory limit setting:

1. Edit /etc/gitlab/gitlab.rb:

   puma['per_worker_max_memory_mb'] = 1024
   

2. Reconfigure GitLab for the changes to take effect:

   sudo gitlab-ctl reconfigure
   

Worker timeout

A timeout of 60 seconds is used when Puma is enabled.

To change the worker timeout:

1. Edit /etc/gitlab/gitlab.rb:

   gitlab_rails['env'] = {
      'GITLAB_RAILS_RACK_TIMEOUT' => 600
    }
   

2. Reconfigure GitLab for the changes to take effect:

   sudo gitlab-ctl reconfigure
   

Memory-constrained environments

In a memory-constrained environment with less than 4GB of RAM available, consider disabling Puma Clustered mode.

Configuring Puma by setting the amount of workers to 0 could reduce memory usage by hundreds of MB. For details on Puma worker and thread settings, see the Puma requirements.

Unlike in a Clustered mode, which is set up by default, only a single Puma process would serve the application.

The downside of running Puma with such configuration is the reduced throughput, which could be considered as a fair tradeoff in a memory-constraint environment.

When running Puma in Single mode, some features are not supported:

• Phased restart do not work: issue
• Phased restart
• Puma Worker Killer

To learn more, visit epic 5303.

Performance caveat when using Puma with Rugged

For deployments where NFS is used to store Git repository, we allow GitLab to use direct Git access to improve performance using Rugged.

Rugged usage is automatically enabled if direct Git access is available and Puma is running single threaded, unless it is disabled by feature flags.

MRI Ruby uses a GVL. This allows MRI Ruby to be multi-threaded, but running at most on a single core. Since Rugged can use a thread for long periods of time (due to intensive I/O operations of Git access), this can starve other threads that might be processing requests. This is not a case for Unicorn or Puma running in a single thread mode, as concurrently at most one request is being processed.

We are actively working on removing Rugged usage. Even though performance without Rugged is acceptable today, in some cases it might be still beneficial to run with it.

Given the caveat of running Rugged with multi-threaded Puma, and acceptable performance of Gitaly, we disable Rugged usage if Puma multi-threaded is used (when Puma is configured to run with more than one thread).

This default behavior may not be the optimal configuration in some situations. If Rugged plays an important role in your deployment, we suggest you benchmark to find the optimal configuration:

• The safest option is to start with single-threaded Puma. When working with Rugged, single-threaded Puma works the same as Unicorn.
• To force Rugged to be used with multi-threaded Puma, you can use feature flags.

Convert Unicorn settings to Puma

Puma has a multi-thread architecture which uses less memory than a multi-process application server like Unicorn. On GitLab.com, we saw a 40% reduction in memory consumption. Most Rails applications requests normally include a proportion of I/O wait time.

During I/O wait time MRI Ruby releases the GVL (Global VM Lock) to other threads. Multi-threaded Puma can therefore still serve more requests than a single process.

When switching to Puma, any Unicorn server configuration will not carry over automatically, due to differences between the two application servers.

The table below summarizes which Unicorn configuration keys correspond to those in Puma when using the Linux package, and which ones have no corresponding counterpart.