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.

note
Starting with GitLab 13.0, Puma is the default web server and Unicorn has been disabled. In GitLab 14.0, Unicorn was removed from the Linux package and only Puma is available.

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

Puma forks worker processes as part of a strategy to reduce memory use.

Each time a worker is created, it shares memory with the primary process and only uses additional memory when it makes changes or additions to its memory pages.

Memory use by workers therefore increases over time, and Puma Worker Killer is the mechanism that recovers this memory.

By default:

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
    

There are costs associated with killing and replacing workers including reduced capacity to run GitLab, and CPU that is consumed restarting the workers. per_worker_max_memory_mb should be set to a higher value if the worker killer is replacing workers too often.

Worker count is calculated based on CPU cores, so a small GitLab deployment with 4-8 workers may experience performance issues if workers are being restarted frequently, once or more per minute. This is too often.

A higher value of 1200 or more would be beneficial if the server has free memory.

The worker killer checks every 20 seconds, and can be monitored using the Puma log /var/log/gitlab/puma/puma_stdout.log. For example, for GitLab 13.5:

PumaWorkerKiller: Out of memory. 4 workers consuming total: 4871.23828125 MB
out of max: 4798.08 MB. Sending TERM to pid 26668 consuming 1001.00390625 MB.

From this output:

  • The formula that calculates the maximum memory value results in workers being killed before they reach the per_worker_max_memory_mb value.
  • The default values for the formula before GitLab 13.5 were 550MB for the primary and per_worker_max_memory_mb specified 850MB for each worker.
  • As of GitLab 13.5 the values are primary: 800MB, worker: 1024MB.
  • The threshold for workers to be killed is set at 98% of the limit:

    0.98 * ( 800 + ( worker_processes * 1024MB ) )
    
  • In the log output above, 0.98 * ( 800 + ( 4 * 1024 ) ) returns the max: 4798.08 MB value.

Increasing the maximum to 1200, for example, would set a max: 5488 MB value.

Workers use additional memory on top of the shared memory, how much depends on a site’s use of GitLab.

Worker timeout

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

note
Unlike Unicorn, the puma['worker_timeout'] setting does not set the maximum request duration.

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:

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

note
Starting with GitLab 13.0, Puma is the default web server and Unicorn has been disabled by default. In GitLab 14.0, Unicorn was removed from the Linux package and only Puma is available.

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.

Unicorn Puma
unicorn['enable'] puma['enable']
unicorn['worker_timeout'] puma['worker_timeout']
unicorn['worker_processes'] puma['worker_processes']
n/a puma['ha']
n/a puma['min_threads']
n/a puma['max_threads']
unicorn['listen'] puma['listen']
unicorn['port'] puma['port']
unicorn['socket'] puma['socket']
unicorn['pidfile'] puma['pidfile']
unicorn['tcp_nopush'] n/a
unicorn['backlog_socket'] n/a
unicorn['somaxconn'] puma['somaxconn']
n/a puma['state_path']
unicorn['log_directory'] puma['log_directory']
unicorn['worker_memory_limit_min'] n/a
unicorn['worker_memory_limit_max'] puma['per_worker_max_memory_mb']
unicorn['exporter_enabled'] puma['exporter_enabled']
unicorn['exporter_address'] puma['exporter_address']
unicorn['exporter_port'] puma['exporter_port']