gitlab-org--gitlab-foss/doc/administration/reference_architectures/index.md

type	stage	group	info
reference, concepts	Enablement	Distribution	To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

Reference architectures

You can set up GitLab on a single server or scale it up to serve many users. This page details the recommended Reference Architectures that were built and verified by the GitLab Quality and Support teams.

Below is a chart representing each architecture tier and the number of users they can handle. As your number of users grow with time, it’s recommended that you scale GitLab accordingly.

Testing on these reference architectures was performed with the GitLab Performance Tool at specific coded workloads, and the throughputs used for testing were calculated based on sample customer data. Select the reference architecture that matches your scale.

Each endpoint type is tested with the following number of requests per second (RPS) per 1,000 users:

• API: 20 RPS
• Web: 2 RPS
• Git: 2 RPS

For GitLab instances with less than 2,000 users, it's recommended that you use the default setup by installing GitLab on a single machine to minimize maintenance and resource costs.

If your organization has more than 2,000 users, the recommendation is to scale the GitLab components to multiple machine nodes. The machine nodes are grouped by components. The addition of these nodes increases the performance and scalability of to your GitLab instance.

When scaling GitLab, there are several factors to consider:

• Multiple application nodes to handle frontend traffic.
• A load balancer is added in front to distribute traffic across the application nodes.
• The application nodes connects to a shared file server and PostgreSQL and Redis services on the backend.

NOTE: Depending on your workflow, the following recommended reference architectures may need to be adapted accordingly. Your workload is influenced by factors including how active your users are, how much automation you use, mirroring, and repository/change size. Additionally the displayed memory values are provided by GCP machine types. For different cloud vendors, attempt to select options that best match the provided architecture.

Available reference architectures

The following reference architectures are available:

• Up to 1,000 users
• Up to 2,000 users
• Up to 3,000 users
• Up to 5,000 users
• Up to 10,000 users
• Up to 25,000 users
• Up to 50,000 users

A GitLab Premium or Ultimate license is required to get assistance from Support with troubleshooting the 2,000 users and higher reference architectures. Read more about our definition of scaled architectures.

Availability Components

GitLab comes with the following components for your use, listed from least to most complex:

• Automated backups
• Traffic load balancer
• Zero downtime updates
• Automated database failover
• Instance level replication with GitLab Geo

As you implement these components, begin with a single server and then do backups. Only after completing the first server should you proceed to the next.

Also, not implementing extra servers for GitLab doesn't necessarily mean that you'll have more downtime. Depending on your needs and experience level, single servers can have more actual perceived uptime for your users.

Automated backups (CORE ONLY)

• Level of complexity: Low
• Required domain knowledge: PostgreSQL, GitLab configurations, Git

This solution is appropriate for many teams that have the default GitLab installation. With automatic backups of the GitLab repositories, configuration, and the database, this can be an optimal solution if you don't have strict requirements. Automated backups is the least complex to setup. This provides a point-in-time recovery of a predetermined schedule.

Traffic load balancer (STARTER ONLY)

• Level of complexity: Medium
• Required domain knowledge: HAProxy, shared storage, distributed systems

This requires separating out GitLab into multiple application nodes with an added load balancer. The load balancer will distribute traffic across GitLab application nodes. Meanwhile, each application node connects to a shared file server and database systems on the back end. This way, if one of the application servers fails, the workflow is not interrupted. HAProxy is recommended as the load balancer.

With this added component you have a number of advantages compared to the default installation:

• Increase the number of users.
• Enable zero-downtime upgrades.
• Increase availability.

For more details on how to configure a traffic load balancer with GitLab, you can refer to any of the available reference architectures with more than 1,000 users.

Zero downtime updates (STARTER ONLY)

• Level of complexity: Medium
• Required domain knowledge: PostgreSQL, HAProxy, shared storage, distributed systems

GitLab supports zero-downtime updates. Single GitLab nodes can be updated with only a few minutes of downtime. To avoid this, we recommend to separate GitLab into several application nodes. As long as at least one of each component is online and capable of handling the instance's usage load, your team's productivity will not be interrupted during the update.

Automated database failover (PREMIUM ONLY)

• Level of complexity: High
• Required domain knowledge: PgBouncer, Repmgr or Patroni, shared storage, distributed systems

By adding automatic failover for database systems, you can enable higher uptime with additional database nodes. This extends the default database with cluster management and failover policies. PgBouncer in conjunction with Repmgr or Patroni is recommended.

Instance level replication with GitLab Geo (PREMIUM ONLY)

• Level of complexity: Very High
• Required domain knowledge: Storage replication

GitLab Geo allows you to replicate your GitLab instance to other geographical locations as a read-only fully operational instance that can also be promoted in case of disaster.

Deviating from the suggested reference architectures

As a general rule of thumb, the further away you move from the Reference Architectures, the harder it will be get support for it. With any deviation, you're introducing a layer of complexity that will add challenges to finding out where potential issues might lie.

The reference architectures use the official GitLab Linux packages (Omnibus GitLab) to install and configure the various components (with one notable exception being the suggested select Cloud Native installation method described below). The components are installed on separate machines (virtualized or bare metal), with machine hardware requirements listed in the "Configuration" column and equivalent VM standard sizes listed in GCP/AWS/Azure columns of each available reference architecture.

Running components on Docker (including Compose) with the same specs should be fine, as Docker is well known in terms of support. However, it is still an additional layer and may still add some support complexities, such as not being able to run strace easily in containers.

Other technologies, like Docker swarm are not officially supported, but can be implemented at your own risk. In that case, GitLab Support will not be able to help you.

Configuring select components with Cloud Native Helm

We also provide Helm charts as a Cloud Native installation method for GitLab. For the reference architectures, select components can be set up in this way as an alternative if so desired.

For these kind of setups we support using the charts in an advanced configuration where stateful backend components, such as the database or Gitaly, are run externally - either via Omnibus or reputable third party services. Note that we don't currently support running the stateful components via Helm at large scales.

When designing these environments you should refer to the respective Reference Architecture above for guidance on sizing. Components run via Helm would be similarly scaled to their Omnibus specs, only translated into Kubernetes resources.

For example, if you were to set up a 50k installation with the Rails nodes being run in Helm, then the same amount of resources as given for Omnibus should be given to the Kubernetes cluster with the Rails nodes broken down into a number of smaller Pods across that cluster.