In this page you will find information about the settings that are used on

SSH host keys fingerprints

Below are the fingerprints for’s SSH host keys. The first time you connect to a repository, you’ll see one of these keys in the output.

Algorithm MD5 (deprecated) SHA256
DSA (deprecated) 7a:47:81:3a:ee:89:89:64:33:ca:44:52:3d:30:d4:87 p8vZBUOR0XQz6sYiaWSMLmh0t9i8srqYKool/Xfdfqw
ECDSA f1:d0:fb:46:73:7a:70:92:5a:ab:5d:ef:43:e2:1c:35 HbW3g8zUjNSksFbqTiUWPWg2Bq1x8xdGUrliXFzSnUw
ED25519 2e:65:6a:c8:cf:bf:b2:8b:9a:bd:6d:9f:11:5c:12:16 eUXGGm1YGsMAS7vkcx6JOJdOGHPem5gQp4taiCfCLB8
RSA b6:03:0e:39:97:9e:d0:e7:24:ce:a3:77:3e:01:42:09 ROQFvPThGrW4RuWLoL9tq9I9zJ42fK4XywyRtbOz/EQ

SSH known_hosts entries

Add the following to .ssh/known_hosts to skip manual fingerprint confirmation in SSH: ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIAfuCHKVTjquxvt6CM6tdG4SLp1Btn/nOeHHE5UOzRdf ssh-rsa AAAAB3NzaC1yc2EAAAADAQABAAABAQCsj2bNKTBSpIYDEGk9KxsGh3mySTRgMtXL583qmBpzeQ+jqCMRgBqB98u3z++J1sKlXHWfM9dyhSevkMwSbhoR8XIq/U0tCNyokEi/ueaBMCvbcTHhO7FcwzY92WK4Yt0aGROY5qX2UKSeOvuP4D6TPqKF1onrSzH9bx9XUf2lEdWT/ia1NEKjunUqu1xOB/StKDHMoX4/OKyIzuS0q/T1zOATthvasJFoPrAjkohTyaDUz2LN5JoH839hViyEG82yB+MjcFV5MU3N1l1QL3cVUCh93xSaua1N85qivl+siMkPGbO5xR/En4iEY6K2XPASUEMaieWVNTRCtJ4S8H+9 ecdsa-sha2-nistp256 AAAAE2VjZHNhLXNoYTItbmlzdHAyNTYAAAAIbmlzdHAyNTYAAABBBFSMqzJeV9rUzU4kWitGjeR4PWSa29SPqJ1fVkhtj3Hw9xjLVXVYrU9QlYWrOLXBpQ6KWjbjTDTdDkoohFzgbEY=

Mail configuration sends emails from the domain via Mailgun and has its own dedicated IP address (


See our backup strategy.

Alternative SSH port can be reached via a different SSH port for git+ssh.

Port 443

An example ~/.ssh/config is the following:

  User git
  Port 443
  PreferredAuthentications publickey
  IdentityFile ~/.ssh/gitlab

GitLab Pages

Below are the settings for GitLab Pages.

設定 Default
Domain name -
IP address -
Custom domains support yes いいえ
TLS certificates support yes いいえ
Maximum size (uncompressed) 1G 100M
Note: The maximum size of your Pages site is regulated by the artifacts maximum size which is part of GitLab CI/CD.

GitLab CI/CD

Below are the current settings regarding GitLab CI/CD.

設定 Default
Artifacts maximum size (uncompressed) 1G 100M
Artifacts expiry time From June 22, 2020, deleted after 30 days unless otherwise specified (artifacts created before that date have no expiry). deleted after 30 days unless otherwise specified
Scheduled Pipeline Cron */5 * * * * 19 * * * *
Max jobs in active pipelines 500 for Free tier, unlimited otherwise Unlimited
Max pipeline schedules in projects 10 for Free tier, 50 for all paid tiers Unlimited
Max number of instance level variables 25 25
Scheduled Job Archival 3 months Never

Repository size limit has the following account limits enabled. If a setting is not listed, it is set to the default value.

If you are near or over the repository size limit, you can reduce your repository size with Git.

設定 Default
Repository size including LFS 10 GB Unlimited
Note: git push and GitLab project imports are limited to 5 GB per request through Cloudflare. Git LFS and imports other than a file upload are not affected by this limit.

IP range is using the IP range for traffic from its Web/API fleet. This whole range is solely allocated to GitLab. You can expect connections from webhooks or repository mirroring to come from those IPs and allow them. is fronted by Cloudflare. For incoming connections to you might need to allow CIDR blocks of Cloudflare (IPv4 and IPv6).

For outgoing connections from CI/CD runners we are not providing static IP addresses. All our runners are deployed into Google Cloud Platform (GCP) - any IP based firewall can be configured by looking up all IP address ranges or CIDR blocks for GCP.

Maximum number of webhooks

A limit of:

  • 100 webhooks applies to projects.
  • 50 webhooks applies to groups.


GitLab offers Linux and Windows shared runners hosted on for executing your pipelines.

Note: Shared Runners provided by GitLab are not configurable. Consider installing your own Runner if you have specific configuration needs.

Linux Shared Runners

Linux Shared Runners on run in autoscale mode and are powered by Google Cloud Platform. Autoscaling means reduced waiting times to spin up CI/CD jobs, and isolated VMs for each project, thus maximizing security. They’re free to use for public open source projects and limited to 2000 CI minutes per month per group for private projects. More minutes can be purchased, if needed. Read about all plans.

All your CI/CD jobs run on n1-standard-1 instances with 3.75GB of RAM, CoreOS and the latest Docker Engine installed. Instances provide 1 vCPU and 25GB of HDD disk space. The default region of the VMs is US East1. Each instance is used only for one job, this ensures any sensitive data left on the system can’t be accessed by other people their CI jobs.

The fleet of Runners are dedicated for GitLab projects as well as community forks of them. They use a slightly larger machine type (n1-standard-2) and have a bigger SSD disk size. They will not run untagged jobs and unlike the general fleet of shared Runners, the instances are re-used up to 40 times.

Jobs handled by the shared Runners on (, will be timed out after 3 hours, regardless of the timeout configured in a project. Check the issues 4010 and 4070 for the reference.

Below are the shared Runners settings.

設定 Default
GitLab Runner Runner versions dashboard -
Executor docker+machine -
Default Docker image ruby:2.5 -
privileged (run Docker in Docker) true false

Pre-clone script

Linux Shared Runners on provide a way to run commands in a CI job before the Runner attempts to run git init and git fetch to download a GitLab repository. The pre_clone_script can be used for:

  • Seeding the build directory with repository data
  • Sending a request to a server
  • Downloading assets from a CDN
  • Any other commands that must run before the git init

To use this feature, define a CI/CD variable called CI_PRE_CLONE_SCRIPT that contains a bash script.

This example demonstrates how you might use a pre-clone step to seed the build directory.


The full contents of our config.toml are:

Note: Settings that are not public are shown as X.

Google Cloud Platform

concurrent = X
check_interval = 1
metrics_server = "X"
sentry_dsn = "X"

  name = "docker-auto-scale"
  request_concurrency = X
  url = ""
  pre_clone_script = "eval \"$CI_PRE_CLONE_SCRIPT\""
  executor = "docker+machine"
  environment = [
  limit = X
    image = "ruby:2.5"
    privileged = true
    volumes = [
      "/dummy-sys-class-dmi-id:/sys/class/dmi/id:ro" # Make kaniko builds work on GCP.
    IdleCount = 50
    IdleTime = 3600
    OffPeakPeriods = ["* * * * * sat,sun *"]
    OffPeakTimezone = "UTC"
    OffPeakIdleCount = 15
    OffPeakIdleTime = 3600
    MaxBuilds = 1 # For security reasons we delete the VM after job has finished so it's not reused.
    MachineName = "srm-%s"
    MachineDriver = "google"
    MachineOptions = [
      "engine-opt=mtu=1460", # Set MTU for container interface, for more information check
      "engine-opt=ipv6", # This will create IPv6 interfaces in the containers.
      "google-operation-backoff-initial-interval=2" # Custom flag from forked docker-machine, for more information check
    Type = "gcs"
    Shared = true
      CredentialsFile = "/path/to/file"
      BucketName = "bucket-name"

Windows Shared Runners (beta)

The Windows Shared Runners are currently in beta and should not be used for production workloads.

During the beta period, the shared runner pipeline quota will apply for groups and projects in the same way as Linux Runners. This may change when the beta period ends, as discussed in this related issue.

Windows Shared Runners on automatically autoscale by launching virtual machines on the Google Cloud Platform. This solution uses a new autoscaling driver developed by GitLab for the custom executor. Windows Shared Runners execute your CI/CD jobs on n1-standard-2 instances with 2 vCPUs and 7.5GB RAM. You can find a full list of available Windows packages in the package documentation.

We want to keep iterating to get Windows Shared Runners in a stable state and generally available. You can follow our work towards this goal in the related epic.


The full contents of our config.toml are:

Note: Settings that are not public are shown as X.
concurrent = X
check_interval = 3

  name = "windows-runner"
  url = ""
  token = "TOKEN"
  executor = "custom"
  builds_dir = "C:\\GitLab-Runner\\builds"
  cache_dir = "C:\\GitLab-Runner\\cache"
  shell  = "powershell"
    config_exec = "C:\\GitLab-Runner\\autoscaler\\autoscaler.exe"
    config_args = ["--config", "C:\\GitLab-Runner\\autoscaler\\config.toml", "custom", "config"]
    prepare_exec = "C:\\GitLab-Runner\\autoscaler\\autoscaler.exe"
    prepare_args = ["--config", "C:\\GitLab-Runner\\autoscaler\\config.toml", "custom", "prepare"]
    run_exec = "C:\\GitLab-Runner\\autoscaler\\autoscaler.exe"
    run_args = ["--config", "C:\\GitLab-Runner\\autoscaler\\config.toml", "custom", "run"]
    cleanup_exec = "C:\\GitLab-Runner\\autoscaler\\autoscaler.exe"
    cleanup_args = ["--config", "C:\\GitLab-Runner\\autoscaler\\config.toml", "custom", "cleanup"]

The full contents of our autoscaler/config.toml are:

Provider = "gcp"
Executor = "winrm"
OS = "windows"
LogLevel = "info"
LogFormat = "text"
LogFile = "C:\\GitLab-Runner\\autoscaler\\autoscaler.log"
VMTag = "windows"

  ServiceAccountFile = "PATH"
  Project = "some-project-df9323"
  Zone = "us-east1-c"
  MachineType = "n1-standard-2"
  Image = "IMAGE"
  DiskSize = 50
  DiskType = "pd-standard"
  Subnetwork = "default"
  Network = "default"
  Tags = ["TAGS"]
  Username = "gitlab_runner"

  MaximumTimeout = 3600
  ExecutionMaxRetries = 0

  Enabled = true
  Directory = "C:\\GitLab-Runner\\autoscaler\\machines"


Below is a simple .gitlab-ci.yml file to show how to start using the Windows Shared Runners:

    - shared-windows
    - windows
    - windows-1809

  - build
  - test

 - Set-Variable -Name "time" -Value (date -Format "%H:%m")
 - echo ${time}
 - echo "started by ${GITLAB_USER_NAME}"

    - .shared_windows_runners
  stage: build
    - echo "running scripts in the build job"

    - .shared_windows_runners
  stage: test
    - echo "running scripts in the test job"

Limitations and known issues

  • All the limitations mentioned in our beta definition.
  • The average provisioning time for a new Windows VM is 5 minutes. This means that you may notice slower build start times on the Windows Shared Runner fleet during the beta. In a future release we will update the autoscaler to enable the pre-provisioning of virtual machines. This will significantly reduce the time it takes to provision a VM on the Windows fleet. You can follow along in the related issue.
  • The Windows Shared Runner fleet may be unavailable occasionally for maintenance or updates.
  • The Windows Shared Runner virtual machine instances do not use the GitLab Docker executor. This means that you will not be able to specify image or services in your pipeline configuration.
  • For the beta release, we have included a set of software packages in the base VM image. If your CI job requires additional software that’s not included in this list, then you will need to add installation commands to before_script or script to install the required software. Note that each job runs on a new VM instance, so the installation of additional software packages needs to be repeated for each job in your pipeline.
  • The job may stay in a pending state for longer than the Linux shared Runners.
  • There is the possibility that we introduce breaking changes which will require updates to pipelines that are using the Windows Shared Runner fleet.

Sidekiq runs Sidekiq with arguments --timeout=4 --concurrency=4 and the following environment variables:

設定 Default
Note: The SIDEKIQ_MEMORY_KILLER_MAX_RSS setting is 16000000 on Sidekiq import nodes and Sidekiq export nodes.

PostgreSQL being a fairly large installation of GitLab means we have changed various PostgreSQL settings to better suit our needs. For example, we use streaming replication and servers in hot-standby mode to balance queries across different database servers.

The list of specific settings (and their defaults) is as follows:

設定 Default
archive_command /usr/bin/envdir /etc/wal-e.d/env /opt/wal-e/bin/wal-e wal-push %p empty
archive_mode on off
autovacuum_analyze_scale_factor 0.01 0.01
autovacuum_max_workers 6 3
autovacuum_vacuum_cost_limit 1000 -1
autovacuum_vacuum_scale_factor 0.01 0.02
checkpoint_completion_target 0.7 0.9
checkpoint_segments 32 10
effective_cache_size 338688MB Based on how much memory is available
hot_standby on off
hot_standby_feedback on off
log_autovacuum_min_duration 0 -1
log_checkpoints on off
log_line_prefix %t [%p]: [%l-1] empty
log_min_duration_statement 1000 -1
log_temp_files 0 -1
maintenance_work_mem 2048MB 16 MB
max_replication_slots 5 0
max_wal_senders 32 0
max_wal_size 5GB 1GB
shared_buffers 112896MB Based on how much memory is available
shared_preload_libraries pg_stat_statements empty
shmall 30146560 Based on the server’s capabilities
shmmax 123480309760 Based on the server’s capabilities
wal_buffers 16MB -1
wal_keep_segments 512 10
wal_level replica minimal
statement_timeout 15s 60s
idle_in_transaction_session_timeout 60s 60s

Some of these settings are in the process being adjusted. For example, the value for shared_buffers is quite high and as such we are looking into adjusting it. More information on this particular change can be found at An up to date list of proposed changes can be found at[]=database&label_name[]=change.

Unicorn adjusts the memory limits for the unicorn-worker-killer gem.

Base default:

  • memory_limit_min = 750MiB
  • memory_limit_max = 1024MiB

Web front-ends:

  • memory_limit_min = 1024MiB
  • memory_limit_max = 1280MiB rate limits

Note: See Rate limits for administrator documentation.

IP blocks usually happen when receives unusual traffic from a single IP address that the system views as potentially malicious based on rate limit settings. After the unusual traffic ceases, the IP address will be automatically released depending on the type of block, as described below.

If you receive a 403 Forbidden error for all requests to, please check for any automated processes that may be triggering a block. For assistance, contact GitLab Support with details, such as the affected IP address.

HAProxy API throttle responds with HTTP status code 429 to API requests that exceed 10 requests per second per IP address.

The following example headers are included for all API requests:

RateLimit-Limit: 600
RateLimit-Observed: 6
RateLimit-Remaining: 594
RateLimit-Reset: 1563325137
RateLimit-ResetTime: Wed, 17 Jul 2019 00:58:57 GMT


Rack Attack initializer

Details of rate limits enforced by Rack Attack.

Protected paths throttle responds with HTTP status code 429 to POST requests at protected paths that exceed 10 requests per minute per IP address.

See the source below for which paths are protected. This includes user creation, user confirmation, user sign in, and password reset.

This header is included in responses to blocked requests:

Retry-After: 60

See Protected Paths for more details.

Git and container registry failed authentication ban responds with HTTP status code 403 for 1 hour, if 30 failed authentication requests were received in a 3-minute period from a single IP address.

This applies only to Git requests and container registry (/jwt/auth) requests (combined).

This limit:

  • Is reset by requests that authenticate successfully. For example, 29 failed authentication requests followed by 1 successful request, followed by 29 more failed authentication requests would not trigger a ban.
  • Does not apply to JWT requests authenticated by gitlab-ci-token.

No response headers are provided.

Admin Area settings

Visibility settings

On, projects, groups, and snippets created As of GitLab 12.2 (July 2019), projects, groups, and snippets have the Internal visibility setting disabled on

SSH maximum number of connections defines the maximum number of concurrent, unauthenticated SSH connections by using the MaxStartups setting. If more than the maximum number of allowed connections occur concurrently, they are dropped and users get an ssh_exchange_identification error.


To help avoid abuse, project and group imports, exports, and export downloads are rate limited. See Project import/export rate limits and Group import/export rate limits for details. Logging

We use Fluentd to parse our logs. Fluentd sends our logs to Stackdriver Logging and Cloud Pub/Sub. Stackdriver is used for storing logs long-term in Google Cold Storage (GCS). Cloud Pub/Sub is used to forward logs to an Elastic cluster using pubsubbeat.

You can view more information in our runbooks such as: at scale

In addition to the GitLab Enterprise Edition Omnibus install, uses the following applications and settings to achieve scale. All settings are publicly available at chef cookbooks.

Elastic Cluster

We use Elasticsearch and Kibana for part of our monitoring solution:


We use Fluentd to unify our GitLab logs:


Prometheus complete our monitoring stack:


For the visualization of monitoring data:


Open source error tracking:


Service discovery:


High Performance TCP/HTTP Load Balancer: