Geo database replication (PREMIUM ONLY)

NOTE: If your GitLab installation uses external (not managed by Omnibus) PostgreSQL instances, the Omnibus roles will not be able to perform all necessary configuration steps. In this case, follow the Geo with external PostgreSQL instances document instead.

NOTE: The stages of the setup process must be completed in the documented order. Before attempting the steps in this stage, complete all prior stages.

This document describes the minimal steps you have to take to replicate your primary GitLab database to a secondary node's database. You may have to change some values, based on attributes including your database's setup and size.

You are encouraged to first read through all the steps before executing them in your testing/production environment.

PostgreSQL replication

The GitLab primary node where the write operations happen will connect to the primary database server, and secondary nodes will connect to their own database servers (which are also read-only).

We recommend using PostgreSQL replication slots to ensure that the primary node retains all the data necessary for the secondary nodes to recover. See below for more details.

The following guide assumes that:

  • You are using Omnibus and therefore you are using PostgreSQL 11 or later which includes the pg_basebackup tool.
  • You have a primary node already set up (the GitLab server you are replicating from), running Omnibus' PostgreSQL (or equivalent version), and you have a new secondary server set up with the same versions of the OS, PostgreSQL, and GitLab on all nodes.

WARNING: Geo works with streaming replication. Logical replication is not supported at this time. There is an issue where support is being discussed.

Step 1. Configure the primary server

  1. SSH into your GitLab primary server and login as root:

    sudo -i
  2. Edit /etc/gitlab/gitlab.rb and add a unique name for your node:

    # The unique identifier for the Geo node.
    gitlab_rails['geo_node_name'] = '<node_name_here>'
  3. Reconfigure the primary node for the change to take effect:

    gitlab-ctl reconfigure
  4. Execute the command below to define the node as primary node:

    gitlab-ctl set-geo-primary-node

    This command will use your defined external_url in /etc/gitlab/gitlab.rb.

  5. GitLab 10.4 and up only: Do the following to make sure the gitlab database user has a password defined:

    NOTE: Until FDW settings are removed in GitLab version 14.0, avoid using single or double quotes in the password for PostgreSQL as that will lead to errors when reconfiguring.

    Generate a MD5 hash of the desired password:

    gitlab-ctl pg-password-md5 gitlab
    # Enter password: <your_password_here>
    # Confirm password: <your_password_here>
    # fca0b89a972d69f00eb3ec98a5838484

    Edit /etc/gitlab/gitlab.rb:

    # Fill with the hash generated by `gitlab-ctl pg-password-md5 gitlab`
    postgresql['sql_user_password'] = '<md5_hash_of_your_password>'
    
    # Every node that runs Puma or Sidekiq needs to have the database
    # password specified as below. If you have a high-availability setup, this
    # must be present in all application nodes.
    gitlab_rails['db_password'] = '<your_password_here>'
  6. Omnibus GitLab already has a replication user called gitlab_replicator. You must set the password for this user manually. You will be prompted to enter a password:

    gitlab-ctl set-replication-password

    This command will also read the postgresql['sql_replication_user'] Omnibus setting in case you have changed gitlab_replicator username to something else.

    If you are using an external database not managed by Omnibus GitLab, you need to create the replicator user and define a password to it manually:

    --- Create a new user 'replicator'
    CREATE USER gitlab_replicator;
    
    --- Set/change a password and grants replication privilege
    ALTER USER gitlab_replicator WITH REPLICATION ENCRYPTED PASSWORD '<replication_password>';
  7. Configure PostgreSQL to listen on network interfaces:

    For security reasons, PostgreSQL does not listen on any network interfaces by default. However, Geo requires the secondary node to be able to connect to the primary node's database. For this reason, we need the address of each node.

    NOTE: For external PostgreSQL instances, see additional instructions.

    If you are using a cloud provider, you can lookup the addresses for each Geo node through your cloud provider's management console.

    To lookup the address of a Geo node, SSH in to the Geo node and execute:

    ##
    ## Private address
    ##
    ip route get 255.255.255.255 | awk '{print "Private address:", $NF; exit}'
    
    ##
    ## Public address
    ##
    echo "External address: $(curl --silent "ipinfo.io/ip")"

    In most cases, the following addresses will be used to configure GitLab Geo:

    Configuration Address
    postgresql['listen_address'] Primary node's public or VPC private address.
    postgresql['md5_auth_cidr_addresses'] Secondary node's public or VPC private addresses.

    If you are using Google Cloud Platform, SoftLayer, or any other vendor that provides a virtual private cloud (VPC) you can use the primary and secondary nodes private addresses (corresponds to "internal address" for Google Cloud Platform) for postgresql['md5_auth_cidr_addresses'] and postgresql['listen_address'].

    The listen_address option opens PostgreSQL up to network connections with the interface corresponding to the given address. See the PostgreSQL documentation for more details.

    NOTE: If you need to use 0.0.0.0 or * as the listen_address, you will also need to add 127.0.0.1/32 to the postgresql['md5_auth_cidr_addresses'] setting, to allow Rails to connect through 127.0.0.1. For more information, see omnibus-5258.

    Depending on your network configuration, the suggested addresses may not be correct. If your primary node and secondary nodes connect over a local area network, or a virtual network connecting availability zones like Amazon's VPC or Google's VPC you should use the secondary node's private address for postgresql['md5_auth_cidr_addresses'].

    Edit /etc/gitlab/gitlab.rb and add the following, replacing the IP addresses with addresses appropriate to your network configuration:

    ##
    ## Geo Primary role
    ## - configure dependent flags automatically to enable Geo
    ##
    roles ['geo_primary_role']
    
    ##
    ## Primary address
    ## - replace '<primary_node_ip>' with the public or VPC address of your Geo primary node
    ##
    postgresql['listen_address'] = '<primary_node_ip>'
    
    ##
    # Allow PostgreSQL client authentication from the primary and secondary IPs. These IPs may be
    # public or VPC addresses in CIDR format, for example ['198.51.100.1/32', '198.51.100.2/32']
    ##
    postgresql['md5_auth_cidr_addresses'] = ['<primary_node_ip>/32', '<secondary_node_ip>/32']
    
    ##
    ## Replication settings
    ## - set this to be the number of Geo secondary nodes you have
    ##
    postgresql['max_replication_slots'] = 1
    # postgresql['max_wal_senders'] = 10
    # postgresql['wal_keep_segments'] = 10
    
    ##
    ## Disable automatic database migrations temporarily
    ## (until PostgreSQL is restarted and listening on the private address).
    ##
    gitlab_rails['auto_migrate'] = false
  8. Optional: If you want to add another secondary node, the relevant setting would look like:

    postgresql['md5_auth_cidr_addresses'] = ['<primary_node_ip>/32', '<secondary_node_ip>/32', '<another_secondary_node_ip>/32']

    You may also want to edit the wal_keep_segments and max_wal_senders to match your database replication requirements. Consult the PostgreSQL - Replication documentation for more information.

  9. Save the file and reconfigure GitLab for the database listen changes and the replication slot changes to be applied:

    gitlab-ctl reconfigure

    Restart PostgreSQL for its changes to take effect:

    gitlab-ctl restart postgresql
  10. Re-enable migrations now that PostgreSQL is restarted and listening on the private address.

    Edit /etc/gitlab/gitlab.rb and change the configuration to true:

    gitlab_rails['auto_migrate'] = true

    Save the file and reconfigure GitLab:

    gitlab-ctl reconfigure
  11. Now that the PostgreSQL server is set up to accept remote connections, run netstat -plnt | grep 5432 to make sure that PostgreSQL is listening on port 5432 to the primary server's private address.

  12. A certificate was automatically generated when GitLab was reconfigured. This will be used automatically to protect your PostgreSQL traffic from eavesdroppers, but to protect against active ("man-in-the-middle") attackers, the secondary node needs a copy of the certificate. Make a copy of the PostgreSQL server.crt file on the primary node by running this command:

    cat ~gitlab-psql/data/server.crt

    Copy the output into a clipboard or into a local file. You will need it when setting up the secondary node! The certificate is not sensitive data.

Step 2. Configure the secondary server

  1. SSH into your GitLab secondary server and login as root:

    sudo -i
  2. Stop application server and Sidekiq

    gitlab-ctl stop puma
    gitlab-ctl stop sidekiq

    NOTE: This step is important so we don't try to execute anything before the node is fully configured.

  3. Check TCP connectivity to the primary node's PostgreSQL server:

    gitlab-rake gitlab:tcp_check[<primary_node_ip>,5432]

    NOTE: If this step fails, you may be using the wrong IP address, or a firewall may be preventing access to the server. Check the IP address, paying close attention to the difference between public and private addresses and ensure that, if a firewall is present, the secondary node is permitted to connect to the primary node on port 5432.

  4. Create a file server.crt in the secondary server, with the content you got on the last step of the primary node's setup:

    editor server.crt
  5. Set up PostgreSQL TLS verification on the secondary node:

    Install the server.crt file:

    install \
       -D \
       -o gitlab-psql \
       -g gitlab-psql \
       -m 0400 \
       -T server.crt ~gitlab-psql/.postgresql/root.crt

    PostgreSQL will now only recognize that exact certificate when verifying TLS connections. The certificate can only be replicated by someone with access to the private key, which is only present on the primary node.

  6. Test that the gitlab-psql user can connect to the primary node's database (the default Omnibus database name is gitlabhq_production):

    sudo \
       -u gitlab-psql /opt/gitlab/embedded/bin/psql \
       --list \
       -U gitlab_replicator \
       -d "dbname=gitlabhq_production sslmode=verify-ca" \
       -W \
       -h <primary_node_ip>

    When prompted enter the password you set in the first step for the gitlab_replicator user. If all worked correctly, you should see the list of primary node's databases.

    A failure to connect here indicates that the TLS configuration is incorrect. Ensure that the contents of ~gitlab-psql/data/server.crt on the primary node match the contents of ~gitlab-psql/.postgresql/root.crt on the secondary node.

  7. Configure PostgreSQL:

    This step is similar to how we configured the primary instance. We need to enable this, even if using a single node.

    Edit /etc/gitlab/gitlab.rb and add the following, replacing the IP addresses with addresses appropriate to your network configuration:

    ##
    ## Geo Secondary role
    ## - configure dependent flags automatically to enable Geo
    ##
    roles ['geo_secondary_role']
    
    ##
    ## Secondary address
    ## - replace '<secondary_node_ip>' with the public or VPC address of your Geo secondary node
    ##
    postgresql['listen_address'] = '<secondary_node_ip>'
    postgresql['md5_auth_cidr_addresses'] = ['<secondary_node_ip>/32']
    
    ##
    ## Database credentials password (defined previously in primary node)
    ## - replicate same values here as defined in primary node
    ##
    postgresql['sql_user_password'] = '<md5_hash_of_your_password>'
    gitlab_rails['db_password'] = '<your_password_here>'

    For external PostgreSQL instances, see additional instructions. If you bring a former primary node back online to serve as a secondary node, then you also need to remove roles ['geo_primary_role'] or geo_primary_role['enable'] = true.

  8. Reconfigure GitLab for the changes to take effect:

    gitlab-ctl reconfigure
  9. Restart PostgreSQL for the IP change to take effect:

    gitlab-ctl restart postgresql

Step 3. Initiate the replication process

Below we provide a script that connects the database on the secondary node to the database on the primary node, replicates the database, and creates the needed files for streaming replication.

The directories used are the defaults that are set up in Omnibus. If you have changed any defaults, configure it as you see fit replacing the directories and paths.

WARNING: Make sure to run this on the secondary server as it removes all PostgreSQL's data before running pg_basebackup.

  1. SSH into your GitLab secondary server and login as root:

    sudo -i
  2. Choose a database-friendly name to use for your secondary node to use as the replication slot name. For example, if your domain is secondary.geo.example.com, you may use secondary_example as the slot name as shown in the commands below.

  3. Execute the command below to start a backup/restore and begin the replication

    WARNING: Each Geo secondary node must have its own unique replication slot name. Using the same slot name between two secondaries will break PostgreSQL replication.

    gitlab-ctl replicate-geo-database \
       --slot-name=<secondary_node_name> \
       --host=<primary_node_ip>

    NOTE: Replication slot names must only contain lowercase letters, numbers, and the underscore character.

    When prompted, enter the plaintext password you set up for the gitlab_replicator user in the first step.

    This command also takes a number of additional options. You can use --help to list them all, but here are a couple of tips:

    • If PostgreSQL is listening on a non-standard port, add --port= as well.
    • If your database is too large to be transferred in 30 minutes, you will need to increase the timeout, e.g., --backup-timeout=3600 if you expect the initial replication to take under an hour.
    • Pass --sslmode=disable to skip PostgreSQL TLS authentication altogether (e.g., you know the network path is secure, or you are using a site-to-site VPN). This is not safe over the public Internet!
    • You can read more details about each sslmode in the PostgreSQL documentation; the instructions above are carefully written to ensure protection against both passive eavesdroppers and active "man-in-the-middle" attackers.
    • Change the --slot-name to the name of the replication slot to be used on the primary database. The script will attempt to create the replication slot automatically if it does not exist.
    • If you're repurposing an old server into a Geo secondary node, you'll need to add --force to the command line.
    • When not in a production machine you can disable backup step if you really sure this is what you want by adding --skip-backup

The replication process is now complete.

PgBouncer support (optional)

PgBouncer may be used with GitLab Geo to pool PostgreSQL connections. We recommend using PgBouncer if you use GitLab in a high-availability configuration with a cluster of nodes supporting a Geo primary node and another cluster of nodes supporting a Geo secondary node. For more information, see High Availability with Omnibus GitLab.

Patroni support

Support for Patroni is intended to replace repmgr as a highly availabile PostgreSQL solution on the primary node, but it can also be used for PostgreSQL HA on a secondary node.

Starting with GitLab 13.5, Patroni is available for experimental use with Geo primary and secondary nodes. Due to its experimental nature, Patroni support is subject to change without notice.

This experimental implementation has the following limitations:

  • Whenever a new Leader is elected, the PgBouncer instance must be reconfigured to point to the new Leader.
  • Whenever a new Leader is elected on the primary node, the Standby Leader on the secondary needs to be reconfigured to point to the new Leader.
  • Whenever gitlab-ctl reconfigure runs on a Patroni Leader instance, there's a chance the node will be demoted due to the required short-time restart. To avoid this, you can pause auto-failover by running gitlab-ctl patroni pause. After a reconfigure, it unpauses on its own.

For instructions about how to set up Patroni on the primary node, see the PostgreSQL replication and failover with Omnibus GitLab page.

If you are currently using repmgr on your Geo primary, see these instructions for migrating from repmgr to Patroni.

A production-ready and secure setup requires at least three Patroni instances on the primary site, and a similar configuration on the secondary sites. Be sure to use password credentials and other database best practices.

Similar to repmgr, using Patroni on a secondary node is optional.

Step 1. Configure Patroni permanent replication slot on the primary site

To set up database replication with Patroni on a secondary node, we need to configure a permanent replication slot on the primary node's Patroni cluster, and ensure password authentication is used.

For each Patroni instance on the primary site starting on the Patroni Leader instance:

  1. SSH into your Patroni instance and login as root:

    sudo -i
  2. Edit /etc/gitlab/gitlab.rb and add the following:

    consul['enable'] = true
    consul['configuration'] = {
      retry_join: %w[CONSUL_PRIMARY1_IP CONSULT_PRIMARY2_IP CONSULT_PRIMARY3_IP]
    }
    
    repmgr['enable'] = false
    
    # You need one entry for each secondary, with a unique name following PostgreSQL slot_name constraints:
    #
    # Configuration syntax will be: 'unique_slotname' => { 'type' => 'physical' },
    # We don't support setting a permanent replication slot for logical replication type
    patroni['replication_slots'] = {
      'geo_secondary' => { 'type' => 'physical' }
    }
    
    patroni['use_pg_rewind'] = true
    patroni['postgresql']['max_wal_senders'] = 8 # Use double of the amount of patroni/reserved slots (3 patronis + 1 reserved slot for a Geo secondary).
    patroni['postgresql']['max_replication_slots'] = 8 # Use double of the amount of patroni/reserved slots (3 patronis + 1 reserved slot for a Geo secondary).
    
    postgresql['md5_auth_cidr_addresses'] = [
      'PATRONI_PRIMARY1_IP/32', 'PATRONI_PRIMARY2_IP/32', 'PATRONI_PRIMARY3_IP/32', 'PATRONI_PRIMARY_PGBOUNCER/32',
      'PATRONI_SECONDARY1_IP/32', 'PATRONI_SECONDARY2_IP/32', 'PATRONI_SECONDARY3_IP/32', 'PATRONI_SECONDARY_PGBOUNCER/32' # We list all secondary instances as they can all become a Standby Leader
    ]
    
    postgresql['pgbouncer_user_password'] = 'PGBOUNCER_PASSWORD_HASH'
    postgresql['sql_replication_password'] = 'POSTGRESQL_REPLICATION_PASSWORD_HASH'
    postgresql['sql_user_password'] = 'POSTGRESQL_PASSWORD_HASH'
  3. Reconfigure GitLab for the changes to take effect:

    gitlab-ctl reconfigure

Step 2. Configure a Standby cluster on the secondary site

NOTE: If you are converting a secondary site to a Patroni Cluster, you must start on the PostgreSQL instance. It will become the Patroni Standby Leader instance, and then you can switchover to another replica if you need.

For each Patroni instance on the secondary site:

  1. SSH into your Patroni node and login as root:

    sudo -i
  2. Edit /etc/gitlab/gitlab.rb and add the following:

    roles ['consul_role', 'postgres_role']
    
    consul['enable'] = true
    consul['configuration'] = {
      retry_join: %w[CONSUL_SECONDARY1_IP CONSULT_SECONDARY2_IP CONSULT_SECONDARY3_IP]
    }
    
    repmgr['enable'] = false
    
    postgresql['md5_auth_cidr_addresses'] = [
      'PATRONI_SECONDARY1_IP/32', 'PATRONI_SECONDARY2_IP/32', 'PATRONI_SECONDARY3_IP/32', 'PATRONI_SECONDARY_PGBOUNCER/32',
      # Any other instance that needs access to the database as per documentation
    ]
    
    patroni['enable'] = false
    patroni['standby_cluster']['enable'] = true
    patroni['standby_cluster']['host'] = 'PATRONI_PRIMARY_LEADER_IP' # This needs to be changed anytime the primary Leader changes
    patroni['standby_cluster']['port'] = 5432
    patroni['standby_cluster']['primary_slot_name'] = 'geo_secondary' # Or the unique replication slot name you setup before
    patroni['replication_password'] = 'PLAIN_TEXT_POSTGRESQL_REPLICATION_PASSWORD'
    patroni['use_pg_rewind'] = true
    patroni['postgresql']['max_wal_senders'] = 5 # A minimum of three for one replica, plus two for each additional replica
    patroni['postgresql']['max_replication_slots'] = 5 # A minimum of three for one replica, plus two for each additional replica
  3. Reconfigure GitLab for the changes to take effect. This is required to bootstrap PostgreSQL users and settings:

    gitlab-ctl reconfigure
  4. Remove the PostgreSQL data directory:

    WARNING: If you are converting a secondary site to a Patroni Cluster, you must skip this step on the PostgreSQL instance.

    rm -rf /var/opt/gitlab/postgresql/data
  5. Edit /etc/gitlab/gitlab.rb to enable Patroni:

    patroni['enable'] = true
  6. Reconfigure GitLab for the changes to take effect:

    gitlab-ctl reconfigure

Migrating from repmgr to Patroni

  1. Before migrating, it is recommended that there is no replication lag between the primary and secondary sites and that replication is paused. In GitLab 13.2 and later, you can pause and resume replication with gitlab-ctl geo-replication-pause and gitlab-ctl geo-replication-resume on a Geo secondary database node.
  2. Follow the instructions to migrate repmgr to Patroni. When configuring Patroni on each primary site database node, add patroni['replicaton_slots'] = { '<slot_name>' => 'physical' } to gitlab.rb where <slot_name> is the name of the replication slot for your Geo secondary. This will ensure that Patroni recognizes the replication slot as permanent and will not drop it upon restarting.
  3. If database replication to the secondary was paused before migration, resume replication once Patroni is confirmed working on the primary.

Migrating a single PostgreSQL node to Patroni

Before the introduction of Patroni, Geo had no Omnibus support for HA setups on the secondary node.

With Patroni it's now possible to support that. In order to migrate the existing PostgreSQL to Patroni:

  1. Make sure you have a Consul cluster setup on the secondary (similar to how you set it up on the primary).
  2. Configure a permanent replication slot.
  3. Configure a Standby Cluster on that single node machine.

You will end up with a "Standby Cluster" with a single node. That allows you to later on add additional patroni nodes by following the same instructions above.

Troubleshooting

Read the troubleshooting document.