Skip to content

GitLab

GitLab architecture overview

Software delivery

There are two software distributions of GitLab:

GitLab is available under different subscriptions.

New versions of GitLab are released from stable branches, and the master branch is used for bleeding-edge development.

For more information, visit the GitLab Release Process.

Both distributions require additional components. These components are described in the Component details section, and all have their own repositories. New versions of each dependent component are usually tags, but staying on the master branch of the GitLab codebase gives you the latest stable version of those components. New versions are generally released around the same time as GitLab releases, with the exception of informal security updates deemed critical.

Components

A typical install of GitLab is on GNU/Linux, but growing number of deployments also use the Kubernetes platform. The largest known GitLab instance is on GitLab.com, which is deployed using our official GitLab Helm chart and the official Linux package.

A typical installation uses NGINX or Apache as a web server to proxy through GitLab Workhorse and into the Puma application server. GitLab serves web pages and the GitLab API using the Puma application server. It uses Sidekiq as a job queue which, in turn, uses Redis as a non-persistent database backend for job information, metadata, and incoming jobs.

By default, communication between Puma and Workhorse is via a Unix domain socket, but forwarding requests via TCP is also supported. Workhorse accesses the gitlab/public directory, bypassing the Puma application server to serve static pages, uploads (for example, avatar images or attachments), and pre-compiled assets.

The GitLab application uses PostgreSQL for persistent database information (for example, users, permissions, issues, or other metadata). GitLab stores the bare Git repositories in the location defined in the configuration file, repositories: section. It also keeps default branch and hook information with the bare repository.

When serving repositories over HTTP/HTTPS GitLab uses the GitLab API to resolve authorization and access and to serve Git objects.

The add-on component GitLab Shell serves repositories over SSH. It manages the SSH keys within the location defined in the configuration file, GitLab Shell section. The file in that location should never be manually edited. GitLab Shell accesses the bare repositories through Gitaly to serve Git objects, and communicates with Redis to submit jobs to Sidekiq for GitLab to process. GitLab Shell queries the GitLab API to determine authorization and access.

Gitaly executes Git operations from GitLab Shell and the GitLab web app, and provides an API to the GitLab web app to get attributes from Git (for example, title, branches, tags, or other metadata), and to get blobs (for example, diffs, commits, or files).

You may also be interested in the production architecture of GitLab.com.

Adapting existing and introducing new components

There are fundamental differences in how the application behaves when it is installed on a traditional Linux machine compared to a containerized platform, such as Kubernetes.

Compared to our official installation methods, some of the notable differences are:

  • Official Linux packages can access files on the same file system with different services. Shared files are not an option for the application running on the Kubernetes platform.
  • Official Linux packages by default have services that have access to the shared configuration and network. This is not the case for services running in Kubernetes, where services might be running in complete isolation, or only accessible through specific ports.

In other words, the shared state between services needs to be carefully considered when architecting new features and adding new components. Services that need to have access to the same files, need to be able to exchange information through the appropriate APIs. Whenever possible, this should not be done with files.

Since components written with the API-first philosophy in mind are compatible with both methods, all new features and services must be written to consider Kubernetes compatibility first.

The simplest way to ensure this, is to add support for your feature or service to the official GitLab Helm chart or reach out to the Distribution team.

Refer to the process for adding new service components for more details.

Simplified component overview

This is a simplified architecture diagram that can be used to understand the GitLab architecture.

A complete architecture diagram is available in our component diagram below.

Simplified Component Overview

Component diagram

graph TB

HTTP[HTTP/HTTPS] -- TCP 80, 443 --> NGINX[NGINX]
SSH -- TCP 22 --> GitLabShell[GitLab Shell]
SMTP[SMTP Gateway]
Geo[GitLab Geo Node] -- TCP 22, 80, 443 --> NGINX

GitLabShell --TCP 8080 -->Puma["Puma (GitLab Rails)"]
GitLabShell --> Praefect
Puma --> PgBouncer[PgBouncer]
Puma --> Redis
Puma --> Praefect
Sidekiq --> Redis
Sidekiq --> PgBouncer
Sidekiq --> Praefect
GitLabWorkhorse[GitLab Workhorse] --> Puma
GitLabWorkhorse --> Redis
GitLabWorkhorse --> Praefect
Praefect --> Gitaly
NGINX --> GitLabWorkhorse
NGINX -- TCP 8090 --> GitLabPages[GitLab Pages]
NGINX --> Grafana[Grafana]
NGINX -- TCP 8150 --> GitLabKas[GitLab Kubernetes Agent Server]
GitLabKas --> Praefect
Grafana -- TCP 9090 --> Prometheus[Prometheus]
Prometheus -- TCP 80, 443 --> Puma
RedisExporter[Redis Exporter] --> Redis
Prometheus -- TCP 9121 --> RedisExporter
PostgreSQLExporter[PostgreSQL Exporter] --> PostgreSQL
PgBouncerExporter[PgBouncer Exporter] --> PgBouncer
Prometheus -- TCP 9187 --> PostgreSQLExporter
Prometheus -- TCP 9100 --> NodeExporter[Node Exporter]
Prometheus -- TCP 9168 --> GitLabExporter[GitLab Exporter]
Prometheus -- TCP 9127 --> PgBouncerExporter
GitLabExporter --> PostgreSQL
GitLabExporter --> GitLabShell
GitLabExporter --> Sidekiq
PgBouncer --> Consul
PostgreSQL --> Consul
PgBouncer --> PostgreSQL
NGINX --> Registry
Puma --> Registry
NGINX --> Mattermost
Mattermost --- Puma
Prometheus --> Alertmanager
Migrations --> PostgreSQL
Runner -- TCP 443 --> NGINX
Puma -- TCP 9200 --> Elasticsearch
Sidekiq -- TCP 9200 --> Elasticsearch
Sidekiq -- TCP 80, 443 --> Sentry
Puma -- TCP 80, 443 --> Sentry
Sidekiq -- UDP 6831 --> Jaeger
Puma -- UDP 6831 --> Jaeger
Gitaly -- UDP 6831 --> Jaeger
GitLabShell -- UDP 6831 --> Jaeger
GitLabWorkhorse -- UDP 6831 --> Jaeger
Alertmanager -- TCP 25 --> SMTP
Sidekiq -- TCP 25 --> SMTP
Puma -- TCP 25 --> SMTP
Puma -- TCP 369 --> LDAP
Sidekiq -- TCP 369 --> LDAP
Puma -- TCP 443 --> ObjectStorage["Object Storage"]
Sidekiq -- TCP 443 --> ObjectStorage
GitLabWorkhorse -- TCP 443 --> ObjectStorage
Registry -- TCP 443 --> ObjectStorage
Geo -- TCP 5432 --> PostgreSQL

click Alertmanager "./architecture.html#alertmanager"
click Praefect "./architecture.html#praefect"
click Geo "./architecture.html#gitlab-geo"
click NGINX "./architecture.html#nginx"
click Runner "./architecture.html#gitlab-runner"
click Registry "./architecture.html#registry"
click ObjectStorage "./architecture.html#minio"
click Mattermost "./architecture.html#mattermost"
click Gitaly "./architecture.html#gitaly"
click Jaeger "./architecture.html#jaeger"
click GitLabWorkhorse "./architecture.html#gitlab-workhorse"
click LDAP "./architecture.html#ldap-authentication"
click Puma "./architecture.html#puma"
click GitLabShell "./architecture.html#gitlab-shell"
click SSH "./architecture.html#ssh-request-22"
click Sidekiq "./architecture.html#sidekiq"
click Sentry "./architecture.html#sentry"
click GitLabExporter "./architecture.html#gitlab-exporter"
click Elasticsearch "./architecture.html#elasticsearch"
click Migrations "./architecture.html#database-migrations"
click PostgreSQL "./architecture.html#postgresql"
click Consul "./architecture.html#consul"
click PgBouncer "./architecture.html#pgbouncer"
click PgBouncerExporter "./architecture.html#pgbouncer-exporter"
click RedisExporter "./architecture.html#redis-exporter"
click Redis "./architecture.html#redis"
click Prometheus "./architecture.html#prometheus"
click Grafana "./architecture.html#grafana"
click GitLabPages "./architecture.html#gitlab-pages"
click PostgreSQLExporter "./architecture.html#postgresql-exporter"
click SMTP "./architecture.html#outbound-email"
click NodeExporter "./architecture.html#node-exporter"

Component legend

  • ✅ - Installed by default
  • ⚙ - Requires additional configuration, or GitLab Managed Apps
  • ⤓ - Manual installation required
  • ❌ - Not supported or no instructions available
  • N/A - Not applicable

Component statuses are linked to configuration documentation for each component.

Component list

Component Description Omnibus GitLab GitLab Environment Toolkit (GET) GitLab chart Minikube Minimal GitLab.com Source GDK CE/EE
Certificate Management TLS Settings, Let's Encrypt ✅ ✅ ✅ ⚙ ✅ ⚙ ⚙ CE & EE
Consul Database node discovery, failover ⚙ ✅ ❌ ❌ ✅ ❌ ❌ EE Only
Database Migrations Database migrations ✅ ✅ ✅ ✅ ✅ ⚙ ✅ CE & EE
Elasticsearch Improved search within GitLab ⤓ ⚙ ⤓ ⤓ ✅ ⤓ ⤓ EE Only
Gitaly Git RPC service for handling all Git calls made by GitLab ✅ ✅ ✅ ✅ ✅ ⚙ ✅ CE & EE
GitLab Exporter Generates a variety of GitLab metrics ✅ ✅ ✅ ✅ ✅ ❌ ❌ CE & EE
GitLab Geo Node Geographically distributed GitLab nodes ⚙ ⚙ ❌ ❌ ✅ ❌ ⚙ EE Only
GitLab Kubernetes Agent Integrate Kubernetes clusters in a cloud-native way ⚙ ⚙ ⚙ ❌ ❌ ⤓ ⚙ EE Only
GitLab Managed Apps Deploy Helm, Ingress, Cert-Manager, Prometheus, GitLab Runner, JupyterHub, or Knative to a cluster ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ CE & EE
GitLab Pages Hosts static websites ⚙ ⚙ ❌ ❌ ✅ ⚙ ⚙ CE & EE
GitLab Kubernetes Agent Integrate Kubernetes clusters in a cloud-native way ⚙ ⚙ ⚙ ❌ ❌ ⤓ ⚙ EE Only
GitLab self-monitoring: Alertmanager Deduplicates, groups, and routes alerts from Prometheus ⚙ ⚙ ✅ ⚙ ✅ ❌ ❌ CE & EE
GitLab self-monitoring: Grafana Metrics dashboard ✅ ✅ ⚙ ⤓ ✅ ❌ ❌ CE & EE
GitLab self-monitoring: Jaeger View traces generated by the GitLab instance ❌ ⚙ ⚙ ❌ ❌ ⤓ ⚙ CE & EE
GitLab self-monitoring: Prometheus Time-series database, metrics collection, and query service ✅ ✅ ✅ ⚙ ✅ ❌ ❌ CE & EE
GitLab self-monitoring: Sentry Track errors generated by the GitLab instance ⤓ ⤓ ⤓ ❌ ✅ ⤓ ⤓ CE & EE
GitLab Shell Handles git over SSH sessions ✅ ✅ ✅ ✅ ✅ ⚙ ✅ CE & EE
GitLab Workhorse Smart reverse proxy, handles large HTTP requests ✅ ✅ ✅ ✅ ✅ ⚙ ✅ CE & EE
Inbound email (SMTP) Receive messages to update issues ⤓ ⤓ ⚙ ⤓ ✅ ⤓ ⤓ CE & EE
Jaeger integration Distributed tracing for deployed apps ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ EE Only
LDAP Authentication Authenticate users against centralized LDAP directory ⤓ ⤓ ⤓ ⤓ ❌ ⤓ ⤓ CE & EE
Mattermost Open-source Slack alternative ⚙ ⚙ ⤓ ⤓ ⤓ ❌ ❌ CE & EE
MinIO Object storage service ⤓ ⤓ ✅ ✅ ✅ ❌ ⚙ CE & EE
NGINX Routes requests to appropriate components, terminates SSL ✅ ✅ ✅ ⚙ ✅ ⤓ ❌ CE & EE
Node Exporter Prometheus endpoint with system metrics ✅ ✅ N/A N/A ✅ ❌ ❌ CE & EE
Outbound email (SMTP) Send email messages to users ⤓ ⤓ ⚙ ⤓ ✅ ⤓ ⤓ CE & EE
Patroni Manage PostgreSQL HA cluster leader selection and replication ⚙ ✅ ❌ ❌ ✅ ❌ ❌ EE Only
PgBouncer Exporter Prometheus endpoint with PgBouncer metrics ⚙ ✅ ❌ ❌ ✅ ❌ ❌ CE & EE
PgBouncer Database connection pooling, failover ⚙ ✅ ❌ ❌ ✅ ❌ ❌ EE Only
PostgreSQL Exporter Prometheus endpoint with PostgreSQL metrics ✅ ✅ ✅ ✅ ✅ ❌ ❌ CE & EE
PostgreSQL Database ✅ ✅ ✅ ✅ ✅ ⤓ ✅ CE & EE
Praefect A transparent proxy between any Git client and Gitaly storage nodes. ✅ ✅ ⚙ ❌ ✅ ⚙ ✅ CE & EE
Puma (GitLab Rails) Handles requests for the web interface and API ✅ ✅ ✅ ✅ ✅ ⚙ ✅ CE & EE
Redis Exporter Prometheus endpoint with Redis metrics ✅ ✅ ✅ ✅ ✅ ❌ ❌ CE & EE
Redis Caching service ✅ ✅ ✅ ✅ ✅ ⤓ ✅ CE & EE
Registry Container registry, allows pushing and pulling of images ⚙ ⚙ ✅ ✅ ✅ ⤓ ⚙ CE & EE
Runner Executes GitLab CI/CD jobs ⤓ ⤓ ✅ ⚙ ✅ ⚙ ⚙ CE & EE
Sentry integration Error tracking for deployed apps ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ ⤓ CE & EE
Sidekiq Background jobs processor ✅ ✅ ✅ ✅ ✅ ✅ ✅ CE & EE

Component details

This document is designed to be consumed by systems administrators and GitLab Support Engineers who want to understand more about the internals of GitLab and how they work together.

When deployed, GitLab should be considered the amalgamation of the below processes. When troubleshooting or debugging, be as specific as possible as to which component you are referencing. That should increase clarity and reduce confusion.

Layers

GitLab can be considered to have two layers from a process perspective:

  • Monitoring: Anything from this layer is not required to deliver GitLab the application, but allows administrators more insight into their infrastructure and what the service as a whole is doing.
  • Core: Any process that is vital for the delivery of GitLab as a platform. If any of these processes halt, a GitLab outage results. For the Core layer, you can further divide into:
    • Processors: These processes are responsible for actually performing operations and presenting the service.
    • Data: These services store/expose structured data for the GitLab service.

Alertmanager

Alert manager is a tool provided by Prometheus that "handles alerts sent by client applications such as the Prometheus server. It takes care of deduplicating, grouping, and routing them to the correct receiver integration such as email, PagerDuty, or Opsgenie. It also takes care of silencing and inhibition of alerts." You can read more in issue #45740 about what we alert on.

Certificate management

Consul

Consul is a tool for service discovery and configuration. Consul is distributed, highly available, and extremely scalable.

Database migrations

Elasticsearch

Elasticsearch is a distributed RESTful search engine built for the cloud.

Gitaly

Gitaly is a service designed by GitLab to remove our need for NFS for Git storage in distributed deployments of GitLab (think GitLab.com or High Availability Deployments). As of 11.3.0, this service handles all Git level access in GitLab. You can read more about the project in the project's README.

Praefect

Praefect is a transparent proxy between each Git client and the Gitaly coordinating the replication of repository updates to secondary nodes.

GitLab Geo

Geo is a premium feature built to help speed up the development of distributed teams by providing one or more read-only mirrors of a primary GitLab instance. This mirror (a Geo secondary site) reduces the time to clone or fetch large repositories and projects, or can be part of a Disaster Recovery solution.

GitLab Exporter

GitLab Exporter is a process designed in house that allows us to export metrics about GitLab application internals to Prometheus. You can read more in the project's README.

GitLab Kubernetes Agent

GitLab Kubernetes Agent is an active in-cluster component for solving GitLab and Kubernetes integration tasks in a secure and cloud-native way.

You can use it to sync deployments onto your Kubernetes cluster.

GitLab Pages

GitLab Pages is a feature that allows you to publish static websites directly from a repository in GitLab.

You can use it either for personal or business websites, such as portfolios, documentation, manifestos, and business presentations. You can also attribute any license to your content.

GitLab Runner

GitLab Runner runs jobs and sends the results to GitLab.

GitLab CI/CD is the open-source continuous integration service included with GitLab that coordinates the testing. The old name of this project was GitLab CI Multi Runner but please use GitLab Runner (without CI) from now on.

GitLab Shell

GitLab Shell is a program designed at GitLab to handle SSH-based git sessions, and modifies the list of authorized keys. GitLab Shell is not a Unix shell nor a replacement for Bash or Zsh.

GitLab Workhorse

GitLab Workhorse is a program designed at GitLab to help alleviate pressure from Puma. You can read more about the historical reasons for developing. It's designed to act as a smart reverse proxy to help speed up GitLab as a whole.

Grafana

Grafana is an open source, feature rich metrics dashboard and graph editor for Graphite, Elasticsearch, OpenTSDB, Prometheus, and InfluxDB.

Jaeger

Jaeger, inspired by Dapper and OpenZipkin, is a distributed tracing system. It can be used for monitoring microservices-based distributed systems.

For monitoring deployed apps, see Jaeger tracing documentation

Logrotate

GitLab is comprised of a large number of services that all log. We started bundling our own Logrotate as of GitLab 7.4 to make sure we were logging responsibly. This is just a packaged version of the common open source offering.

Mattermost

Mattermost is an open source, private cloud, Slack-alternative from https://mattermost.com.

MinIO

MinIO is an object storage server released under Apache License v2.0. It is compatible with Amazon S3 cloud storage service. It is best suited for storing unstructured data such as photos, videos, log files, backups, and container / VM images. Size of an object can range from a few KBs to a maximum of 5TB.

NGINX

NGINX has an Ingress port for all HTTP requests and routes them to the appropriate sub-systems within GitLab. We are bundling an unmodified version of the popular open source webserver.

Node Exporter

Node Exporter is a Prometheus tool that gives us metrics on the underlying machine (think CPU/Disk/Load). It's just a packaged version of the common open source offering from the Prometheus project.

Patroni

PgBouncer

Lightweight connection pooler for PostgreSQL.

PgBouncer Exporter

Prometheus exporter for PgBouncer. Exports metrics at 9127/metrics.

PostgreSQL

GitLab packages the popular Database to provide storage for Application meta data and user information.

PostgreSQL Exporter

postgres_exporter is the community provided Prometheus exporter that delivers data about PostgreSQL to Prometheus for use in Grafana Dashboards.

Prometheus

Prometheus is a time-series tool that helps GitLab administrators expose metrics about the individual processes used to provide GitLab the service.

Redis

Redis is packaged to provide a place to store:

  • session data
  • temporary cache information
  • background job queues

See our Redis guidelines for more information about how GitLab uses Redis.

Redis Exporter

Redis Exporter is designed to give specific metrics about the Redis process to Prometheus so that we can graph these metrics in Grafana.

Registry

The registry is what users use to store their own Docker images. The bundled registry uses NGINX as a load balancer and GitLab as an authentication manager. Whenever a client requests to pull or push an image from the registry, it returns a 401 response along with a header detailing where to get an authentication token, in this case the GitLab instance. The client then requests a pull or push auth token from GitLab and retries the original request to the registry. Learn more about token authentication.

An external registry can also be configured to use GitLab as an auth endpoint.

Sentry

Sentry fundamentally is a service that helps you monitor and fix crashes in real time. The server is in Python, but it contains a full API for sending events from any language, in any application.

For monitoring deployed apps, see the Sentry integration docs

Sidekiq

Sidekiq is a Ruby background job processor that pulls jobs from the Redis queue and processes them. Background jobs allow GitLab to provide a faster request/response cycle by moving work into the background.

Puma

Starting with GitLab 13.0, Puma is the default web server and Unicorn has been disabled by default.

Puma is a Ruby application server that is used to run the core Rails Application that provides the user facing features in GitLab. Often this displays in process output as bundle or config.ru depending on the GitLab version.

Unicorn

Starting with GitLab 13.0, Puma is the default web server and Unicorn has been disabled by default.

Unicorn is a Ruby application server that is used to run the core Rails Application that provides the user facing features in GitLab. Often this displays in process output as bundle or config.ru depending on the GitLab version.

LDAP Authentication

Outbound Email

Inbound Email

GitLab Managed Apps

GitLab provides GitLab Managed Apps, a one-click install for various applications which can be added directly to your configured cluster. These applications are needed for Review Apps and deployments when using Auto DevOps. You can install them after you create a cluster. This includes:

GitLab by request type

GitLab provides two "interfaces" for end users to access the service:

  • Web HTTP Requests (Viewing the UI/API)
  • Git HTTP/SSH Requests (Pushing/Pulling Git Data)

It's important to understand the distinction as some processes are used in both and others are exclusive to a specific request type.

GitLab Web HTTP request cycle

When making a request to an HTTP Endpoint (think /users/sign_in) the request takes the following path through the GitLab Service:

  • NGINX - Acts as our first line reverse proxy.
  • GitLab Workhorse - This determines if it needs to go to the Rails application or somewhere else to reduce load on Puma.
  • Puma - Since this is a web request, and it needs to access the application, it routes to Puma.
  • PostgreSQL/Gitaly/Redis - Depending on the type of request, it may hit these services to store or retrieve data.

GitLab Git request cycle

Below we describe the different paths that HTTP vs. SSH Git requests take. There is some overlap with the Web Request Cycle but also some differences.

Web request (80/443)

Git operations over HTTP use the stateless "smart" protocol described in the Git documentation, but responsibility for handling these operations is split across several GitLab components.

Here is a sequence diagram for git fetch. Note that all requests pass through NGINX as well as any other HTTP load balancers, but are not transformed in any way by them. All paths are presented relative to a /namespace/project.git URL.

sequenceDiagram
    participant Git on client
    participant NGINX
    participant Workhorse
    participant Rails
    participant Gitaly
    participant Git on server

    Note left of Git on client: git fetch<br/>info-refs
    Git on client->>+Workhorse: GET /info/refs?service=git-upload-pack
    Workhorse->>+Rails: GET /info/refs?service=git-upload-pack
    Note right of Rails: Auth check
    Rails-->>-Workhorse: Gitlab::Workhorse.git_http_ok
    Workhorse->>+Gitaly: SmartHTTPService.InfoRefsUploadPack request
    Gitaly->>+Git on server: git upload-pack --stateless-rpc --advertise-refs
    Git on server-->>-Gitaly: git upload-pack response
    Gitaly-->>-Workhorse: SmartHTTPService.InfoRefsUploadPack response
    Workhorse-->>-Git on client: 200 OK

    Note left of Git on client: git fetch<br/>fetch-pack
    Git on client->>+Workhorse: POST /git-upload-pack
    Workhorse->>+Rails: POST /git-upload-pack
    Note right of Rails: Auth check
    Rails-->>-Workhorse: Gitlab::Workhorse.git_http_ok
    Workhorse->>+Gitaly: SmartHTTPService.PostUploadPack request
    Gitaly->>+Git on server: git upload-pack --stateless-rpc
    Git on server-->>-Gitaly: git upload-pack response
    Gitaly-->>-Workhorse: SmartHTTPService.PostUploadPack response
    Workhorse-->>-Git on client: 200 OK

The sequence is similar for git push, except git-receive-pack is used instead of git-upload-pack.

SSH request (22)

Git operations over SSH can use the stateful protocol described in the Git documentation, but responsibility for handling them is split across several GitLab components.

No GitLab components speak SSH directly - all SSH connections are made between Git on the client machine and the SSH server, which terminates the connection. To the SSH server, all connections are authenticated as the git user; GitLab users are differentiated by the SSH key presented by the client.

Here is a sequence diagram for git fetch, assuming Fast SSH key lookup is enabled. Note that AuthorizedKeysCommand is an executable provided by GitLab Shell:

sequenceDiagram
    participant Git on client
    participant SSH server
    participant AuthorizedKeysCommand
    participant GitLab Shell
    participant Rails
    participant Gitaly
    participant Git on server

    Note left of Git on client: git fetch
    Git on client->>+SSH server: ssh git fetch-pack request
    SSH server->>+AuthorizedKeysCommand: gitlab-shell-authorized-keys-check git AAAA...
    AuthorizedKeysCommand->>+Rails: GET /internal/api/authorized_keys?key=AAAA...
    Note right of Rails: Lookup key ID
    Rails-->>-AuthorizedKeysCommand: 200 OK, command="gitlab-shell upload-pack key_id=1"
    AuthorizedKeysCommand-->>-SSH server: command="gitlab-shell upload-pack key_id=1"
    SSH server->>+GitLab Shell: gitlab-shell upload-pack key_id=1
    GitLab Shell->>+Rails: GET /internal/api/allowed?action=upload_pack&key_id=1
    Note right of Rails: Auth check
    Rails-->>-GitLab Shell: 200 OK, { gitaly: ... }
    GitLab Shell->>+Gitaly: SSHService.SSHUploadPack request
    Gitaly->>+Git on server: git upload-pack request
    Note over Git on client,Git on server: Bidirectional communication between Git client and server
    Git on server-->>-Gitaly: git upload-pack response
    Gitaly -->>-GitLab Shell: SSHService.SSHUploadPack response
    GitLab Shell-->>-SSH server: gitlab-shell upload-pack response
    SSH server-->>-Git on client: ssh git fetch-pack response

The git push operation is very similar, except git receive-pack is used instead of git upload-pack.

If fast SSH key lookups are not enabled, the SSH server reads from the ~git/.ssh/authorized_keys file to determine what command to run for a given SSH session. This is kept up to date by an AuthorizedKeysWorker in Rails, scheduled to run whenever an SSH key is modified by a user.

SSH certificates may be used instead of keys. In this case, AuthorizedKeysCommand is replaced with an AuthorizedPrincipalsCommand. This extracts a username from the certificate without using the Rails internal API, which is used instead of key_id in the /api/internal/allowed call later.

GitLab Shell also has a few operations that do not involve Gitaly, such as resetting two-factor authentication codes. These are handled in the same way, except there is no round-trip into Gitaly - Rails performs the action as part of the internal API call, and GitLab Shell streams the response back to the user directly.

System layout

When referring to ~git in the pictures it means the home directory of the Git user which is typically /home/git.

GitLab is primarily installed within the /home/git user home directory as git user. Within the home directory is where the GitLab server software resides as well as the repositories (though the repository location is configurable).

The bare repositories are located in /home/git/repositories. GitLab is a Ruby on rails application so the particulars of the inner workings can be learned by studying how a Ruby on rails application works.

To serve repositories over SSH there's an add-on application called GitLab Shell which is installed in /home/git/gitlab-shell.

Installation folder summary

To summarize here's the directory structure of the git user home directory.

Processes

ps aux | grep '^git'

GitLab has several components to operate. It requires a persistent database (PostgreSQL) and Redis database, and uses Apache httpd or NGINX to proxypass Puma. All these components should run as different system users to GitLab (for example, postgres, redis, and www-data, instead of git).

As the git user it starts Sidekiq and Puma (a simple Ruby HTTP server running on port 8080 by default). Under the GitLab user there are normally 4 processes: puma master (1 process), puma cluster worker (2 processes), sidekiq (1 process).

Repository access

Repositories get accessed via HTTP or SSH. HTTP cloning/push/pull uses the GitLab API and SSH cloning is handled by GitLab Shell (previously explained).

Troubleshooting

See the README for more information.

Init scripts of the services

The GitLab init script starts and stops Puma and Sidekiq:

/etc/init.d/gitlab
Usage: service gitlab {start|stop|restart|reload|status}

Redis (key-value store/non-persistent database):

/etc/init.d/redis
Usage: /etc/init.d/redis {start|stop|status|restart|condrestart|try-restart}

SSH daemon:

/etc/init.d/sshd
Usage: /etc/init.d/sshd {start|stop|restart|reload|force-reload|condrestart|try-restart|status}

Web server (one of the following):

/etc/init.d/httpd
Usage: httpd {start|stop|restart|condrestart|try-restart|force-reload|reload|status|fullstatus|graceful|help|configtest}

$ /etc/init.d/nginx
Usage: nginx {start|stop|restart|reload|force-reload|status|configtest}

Persistent database:

$ /etc/init.d/postgresql
Usage: /etc/init.d/postgresql {start|stop|restart|reload|force-reload|status} [version ..]

Log locations of the services

GitLab (includes Puma and Sidekiq logs):

  • /home/git/gitlab/log/ contains application.log, production.log, sidekiq.log, puma.stdout.log, git_json.log and puma.stderr.log normally.

GitLab Shell:

  • /home/git/gitlab-shell/gitlab-shell.log

SSH:

  • /var/log/auth.log auth log (on Ubuntu).
  • /var/log/secure auth log (on RHEL).

NGINX:

  • /var/log/nginx/ contains error and access logs.

Apache httpd:

  • Explanation of Apache logs.
  • /var/log/apache2/ contains error and output logs (on Ubuntu).
  • /var/log/httpd/ contains error and output logs (on RHEL).

Redis:

  • /var/log/redis/redis.log there are also log-rotated logs there.

PostgreSQL:

  • /var/log/postgresql/*

GitLab specific configuration files

GitLab has configuration files located in /home/git/gitlab/config/*. Commonly referenced configuration files include:

  • gitlab.yml: GitLab configuration
  • puma.rb: Puma web server settings
  • database.yml: Database connection settings

GitLab Shell has a configuration file at /home/git/gitlab-shell/config.yml.

Maintenance tasks

GitLab provides Rake tasks with which you see version information and run a quick check on your configuration to ensure it is configured properly within the application. See maintenance Rake tasks. In a nutshell, do the following:

sudo -i -u git
cd gitlab
bundle exec rake gitlab:env:info RAILS_ENV=production
bundle exec rake gitlab:check RAILS_ENV=production

It's recommended to sign in to the git user using either sudo -i -u git or sudo su - git. Although the sudo commands provided by GitLab work in Ubuntu, they don't always work in RHEL.

GitLab.com

The GitLab.com architecture is detailed for your reference, but this architecture is only useful if you have millions of users.