debian-mirror-gitlab/doc/administration/gitaly/praefect.md

---
stage: Systems
group: Gitaly
info: To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/product/ux/technical-writing/#assignments
---

# Configure Gitaly Cluster **(FREE SELF)**

Configure Gitaly Cluster using either:

- Gitaly Cluster configuration instructions available as part of
  [reference architectures](../reference_architectures/index.md) for installations of up to:
  - [3000 users](../reference_architectures/3k_users.md#configure-gitaly-cluster).
  - [5000 users](../reference_architectures/5k_users.md#configure-gitaly-cluster).
  - [10,000 users](../reference_architectures/10k_users.md#configure-gitaly-cluster).
  - [25,000 users](../reference_architectures/25k_users.md#configure-gitaly-cluster).
  - [50,000 users](../reference_architectures/50k_users.md#configure-gitaly-cluster).
- The custom configuration instructions that follow on this page.

Smaller GitLab installations may need only [Gitaly itself](index.md).

## Requirements

The minimum recommended configuration for a Gitaly Cluster requires:

- 1 load balancer
- 1 PostgreSQL server (PostgreSQL 11 or newer)
- 3 Praefect nodes
- 3 Gitaly nodes (1 primary, 2 secondary)

NOTE:
[Disk requirements](index.md#disk-requirements) apply to Gitaly nodes.

You should configure an odd number of Gitaly nodes so that transactions have a tie-breaker in case one of the
Gitaly nodes fails in a mutating RPC call.

See the [design document](https://gitlab.com/gitlab-org/gitaly/-/blob/master/doc/design_ha.md)
for implementation details.

NOTE:
If not set in GitLab, feature flags are read as false from the console and Praefect uses their
default value. The default value depends on the GitLab version.

### Network latency and connectivity

Network latency for Gitaly Cluster should ideally be measurable in single-digit milliseconds. Latency is particularly
important for:

- Gitaly node health checks. Nodes must be able to respond 1 second or faster.
- Reference transactions that enforce [strong consistency](index.md#strong-consistency). Lower latencies mean Gitaly
  nodes can agree on changes faster.

Achieving acceptable latency between Gitaly nodes:

- On physical networks generally means high bandwidth, single location connections.
- On the cloud generally means in the same region, including allowing cross availability zone replication. These links
  are designed for this type of synchronization. Latency of less than 2 ms should be sufficient for Gitaly Cluster.

If you can't provide low network latencies for replication (for example, between distant locations), consider Geo. For
more information, see [Comparison to Geo](index.md#comparison-to-geo).

Gitaly Cluster [components](index.md#components) communicate with each other over many routes. Your firewall rules must
allow the following for Gitaly Cluster to function properly:

| From                   | To                     | Default port | TLS port |
|:-----------------------|:-----------------------|:-------------|:---------|
| GitLab                 | Praefect load balancer | `2305`       | `3305`   |
| Praefect load balancer | Praefect               | `2305`       | `3305`   |
| Praefect               | Gitaly                 | `8075`       | `9999`   |
| Praefect               | GitLab (internal API)  | `80`         | `443`    |
| Gitaly                 | GitLab (internal API)  | `80`         | `443`    |
| Gitaly                 | Praefect load balancer | `2305`       | `3305`   |
| Gitaly                 | Praefect               | `2305`       | `3305`   |
| Gitaly                 | Gitaly                 | `8075`       | `9999`   |

NOTE:
Gitaly does not directly connect to Praefect. However, requests from Gitaly to the Praefect
load balancer may still be blocked unless firewalls on the Praefect nodes allow traffic from
the Gitaly nodes.

### Praefect database storage

The requirements are relatively low because the database contains only metadata of:

- Where repositories are located.
- Some queued work.

It depends on the number of repositories, but a good minimum is 5-10 GB, similar to the main
GitLab application database.

## Setup Instructions

If you [installed](https://about.gitlab.com/install/) GitLab using the Omnibus GitLab package
(highly recommended), follow the steps below:

1. [Preparation](#preparation)
1. [Configuring the Praefect database](#postgresql)
1. [Configuring the Praefect proxy/router](#praefect)
1. [Configuring each Gitaly node](#gitaly) (once for each Gitaly node)
1. [Configure the load balancer](#load-balancer)
1. [Updating the GitLab server configuration](#gitlab)
1. [Configure Grafana](#grafana)

### Preparation

Before beginning, you should already have a working GitLab instance.
[Learn how to install GitLab](https://about.gitlab.com/install/).

Provision a PostgreSQL server. You should use the PostgreSQL that is shipped
with Omnibus GitLab and use it to configure the PostgreSQL database. You can use an
external PostgreSQL server (version 11 or newer) but you must set it up [manually](#manual-database-setup).

Prepare all your new nodes by [installing GitLab](https://about.gitlab.com/install/). You need:

- 1 PostgreSQL node
- 1 PgBouncer node (optional)
- At least 1 Praefect node (minimal storage required)
- 3 Gitaly nodes (high CPU, high memory, fast storage)
- 1 GitLab server

You also need the IP/host address for each node:

1. `PRAEFECT_LOADBALANCER_HOST`: the IP/host address of Praefect load balancer
1. `POSTGRESQL_HOST`: the IP/host address of the PostgreSQL server
1. `PGBOUNCER_HOST`: the IP/host address of the PostgreSQL server
1. `PRAEFECT_HOST`: the IP/host address of the Praefect server
1. `GITALY_HOST_*`: the IP or host address of each Gitaly server
1. `GITLAB_HOST`: the IP/host address of the GitLab server

If you are using Google Cloud Platform, SoftLayer, or any other vendor that provides a virtual private cloud (VPC) you can use the private addresses for each cloud instance (corresponds to "internal address" for Google Cloud Platform) for `PRAEFECT_HOST`, `GITALY_HOST_*`, and `GITLAB_HOST`.

#### Secrets

The communication between components is secured with different secrets, which
are described below. Before you begin, generate a unique secret for each, and
make note of it. This enables you to replace these placeholder tokens
with secure tokens as you complete the setup process.

1. `GITLAB_SHELL_SECRET_TOKEN`: this is used by Git hooks to make callback HTTP
   API requests to GitLab when accepting a Git push. This secret is shared with
   GitLab Shell for legacy reasons.
1. `PRAEFECT_EXTERNAL_TOKEN`: repositories hosted on your Praefect cluster can
   only be accessed by Gitaly clients that carry this token.
1. `PRAEFECT_INTERNAL_TOKEN`: this token is used for replication traffic inside
   your Praefect cluster. This token is distinct from `PRAEFECT_EXTERNAL_TOKEN`
   because Gitaly clients must not be able to access internal nodes of the
   Praefect cluster directly; that could lead to data loss.
1. `PRAEFECT_SQL_PASSWORD`: this password is used by Praefect to connect to
   PostgreSQL.
1. `PRAEFECT_SQL_PASSWORD_HASH`: the hash of password of the Praefect user.
   Use `gitlab-ctl pg-password-md5 praefect` to generate the hash. The command
   asks for the password for `praefect` user. Enter `PRAEFECT_SQL_PASSWORD`
   plaintext password. By default, Praefect uses `praefect` user, but you can
   change it.
1. `PGBOUNCER_SQL_PASSWORD_HASH`: the hash of password of the PgBouncer user.
   PgBouncer uses this password to connect to PostgreSQL. For more details
   see [bundled PgBouncer](../postgresql/pgbouncer.md) documentation.

We note in the instructions below where these secrets are required.

NOTE:
Omnibus GitLab installations can use `gitlab-secrets.json` for `GITLAB_SHELL_SECRET_TOKEN`.

### Customize time server setting

By default, Gitaly and Praefect nodes use the time server at `pool.ntp.org` for time synchronization checks. You can customize this setting by adding the
following to `gitlab.rb` on each node:

- `gitaly['env'] = { "NTP_HOST" => "ntp.example.com" }`, for Gitaly nodes.
- `praefect['env'] = { "NTP_HOST" => "ntp.example.com" }`, for Praefect nodes.

### PostgreSQL

NOTE:
Do not store the GitLab application database and the Praefect
database on the same PostgreSQL server if using [Geo](../geo/index.md).
The replication state is internal to each instance of GitLab and should
not be replicated.

These instructions help set up a single PostgreSQL database, which creates a single point of failure. To avoid this, you can configure your own clustered
PostgreSQL. Support for PostgreSQL replication and failover using Omnibus GitLab is proposed in [epic 7814](https://gitlab.com/groups/gitlab-org/-/epics/7814).
Clustered database support for other databases (for example, Praefect and Geo databases) is proposed in
[issue 7292](https://gitlab.com/gitlab-org/omnibus-gitlab/-/issues/7292).

The following options are available:

- For non-Geo installations, either:
  - Use one of the documented [PostgreSQL setups](../postgresql/index.md).
  - Use your own third-party database setup. This requires [manual setup](#manual-database-setup).
- For Geo instances, either:
  - Set up a separate [PostgreSQL instance](https://www.postgresql.org/docs/11/high-availability.html).
  - Use a cloud-managed PostgreSQL service. AWS
     [Relational Database Service](https://aws.amazon.com/rds/) is recommended.

Setting up PostgreSQL creates empty Praefect tables. For more information, see the
[relevant troubleshooting section](troubleshooting.md#relation-does-not-exist-errors).

#### Running GitLab and Praefect databases on the same server

The GitLab application database and the Praefect database can be run on the same server. However, Praefect should have
its own database server when using Omnibus GitLab PostgreSQL. If there is a failover, Praefect isn't aware and starts to
fail as the database it's trying to use would either:

- Be unavailable.
- In read-only mode.

#### Manual database setup

To complete this section you need:

- One Praefect node
- One PostgreSQL node (version 11 or newer)
  - A PostgreSQL user with permissions to manage the database server

In this section, we configure the PostgreSQL database. This can be used for both external
and Omnibus-provided PostgreSQL server.

To run the following instructions, you can use the Praefect node, where `psql` is installed
by Omnibus GitLab (`/opt/gitlab/embedded/bin/psql`). If you are using the Omnibus-provided
PostgreSQL you can use `gitlab-psql` on the PostgreSQL node instead:

1. Create a new user `praefect` to be used by Praefect:

   ```sql
   CREATE ROLE praefect WITH LOGIN PASSWORD 'PRAEFECT_SQL_PASSWORD';
   ```

   Replace `PRAEFECT_SQL_PASSWORD` with the strong password you generated in the preparation step.

1. Create a new database `praefect_production` that is owned by `praefect` user.

   ```sql
   CREATE DATABASE praefect_production WITH OWNER praefect ENCODING UTF8;
   ```

For using Omnibus-provided PgBouncer you need to take the following additional steps. We strongly
recommend using the PostgreSQL that is shipped with Omnibus as the backend. The following
instructions only work on Omnibus-provided PostgreSQL:

1. For Omnibus-provided PgBouncer, you need to use the hash of `praefect` password instead the of the
   actual password:

   ```sql
   ALTER ROLE praefect WITH PASSWORD 'md5<PRAEFECT_SQL_PASSWORD_HASH>';
   ```

   Replace `<PRAEFECT_SQL_PASSWORD_HASH>` with the hash of the password you generated in the
   preparation step. It is prefixed with `md5` literal.

1. The PgBouncer that is shipped with Omnibus is configured to use [`auth_query`](https://www.pgbouncer.org/config.html#generic-settings)
   and uses `pg_shadow_lookup` function. You need to create this function in `praefect_production`
   database:

   ```sql
   CREATE OR REPLACE FUNCTION public.pg_shadow_lookup(in i_username text, out username text, out password text) RETURNS record AS $$
   BEGIN
       SELECT usename, passwd FROM pg_catalog.pg_shadow
       WHERE usename = i_username INTO username, password;
       RETURN;
   END;
   $$ LANGUAGE plpgsql SECURITY DEFINER;

   REVOKE ALL ON FUNCTION public.pg_shadow_lookup(text) FROM public, pgbouncer;
   GRANT EXECUTE ON FUNCTION public.pg_shadow_lookup(text) TO pgbouncer;
   ```

The database used by Praefect is now configured.

You can now configure Praefect to use the database:

```ruby
praefect['configuration'] = {
   # ...
   database: {
      # ...
      host: POSTGRESQL_HOST,
      port: 5432,
      password: PRAEFECT_SQL_PASSWORD,
      dbname: 'praefect_production',
   }
}
```

If you see Praefect database errors after configuring PostgreSQL, see
[troubleshooting steps](troubleshooting.md#relation-does-not-exist-errors).

#### Reads distribution caching

Praefect performance can be improved by additionally configuring the `database_direct`
settings:

```ruby
praefect['configuration'] = {
   # ...
   database: {
      # ...
      session_pooled: {
         # ...
         host: POSTGRESQL_HOST,
         port: 5432

         # Use the following to override parameters of direct database connection.
         # Comment out where the parameters are the same for both connections.
         user: 'praefect',
         password: PRAEFECT_SQL_PASSWORD,
         dbname: 'praefect_production',
         # sslmode: '...',
         # sslcert: '...',
         # sslkey: '...',
         # sslrootcert: '...',
      }
   }
}
```

When configured, this connection is automatically used for the
[SQL LISTEN](https://www.postgresql.org/docs/11/sql-listen.html) feature and
allows Praefect to receive notifications from PostgreSQL for cache invalidation.

Verify this feature is working by looking for the following log entry in the Praefect
log:

```plaintext
reads distribution caching is enabled by configuration
```

#### Use PgBouncer

To reduce PostgreSQL resource consumption, you should set up and configure [PgBouncer](https://www.pgbouncer.org/) in
front of the PostgreSQL instance. However, PgBouncer isn't required because
Praefect makes a low number of connections. If you choose to use PgBouncer, you can use the same PgBouncer instance for
both the GitLab application database and the Praefect database.

To configure PgBouncer in front of the PostgreSQL instance, you must point Praefect to PgBouncer by setting database
parameters on Praefect configuration:

```ruby
praefect['configuration'] = {
   # ...
   database: {
      # ...
      host: PGBOUNCER_HOST,
      port: 6432,
      user: 'praefect',
      password: PRAEFECT_SQL_PASSWORD,
      dbname: 'praefect_production',
      # sslmode: '...',
      # sslcert: '...',
      # sslkey: '...',
      # sslrootcert: '...',
   }
}
```

Praefect requires an additional connection to the PostgreSQL that supports the
[LISTEN](https://www.postgresql.org/docs/11/sql-listen.html) feature. With PgBouncer
this feature is only available with `session` pool mode (`pool_mode = session`).
It is not supported in `transaction` pool mode (`pool_mode = transaction`).

To configure the additional connection, you must either:

- Configure a new PgBouncer database that uses to the same PostgreSQL database endpoint,
  but with different pool mode (`pool_mode = session`).
- Connect Praefect directly to PostgreSQL and bypass PgBouncer.

#### Configure a new PgBouncer database with `pool_mode = session`

You should use PgBouncer with `session` pool mode. You can use the
[bundled PgBouncer](../postgresql/pgbouncer.md) or use an external PgBouncer and
[configure it manually](https://www.pgbouncer.org/config.html).

The following example uses the bundled PgBouncer and sets up two separate connection pools on PostgreSQL host,
one in `session` pool mode and the other in `transaction` pool mode. For this example to work,
you need to prepare PostgreSQL server as documented in [in the setup instructions](#manual-database-setup):

```ruby
pgbouncer['databases'] = {
  # Other database configuration including gitlabhq_production
  ...

  praefect_production: {
    host: POSTGRESQL_HOST,
    # Use `pgbouncer` user to connect to database backend.
    user: 'pgbouncer',
    password: PGBOUNCER_SQL_PASSWORD_HASH,
    pool_mode: 'transaction'
  },
  praefect_production_direct: {
    host: POSTGRESQL_HOST,
    # Use `pgbouncer` user to connect to database backend.
    user: 'pgbouncer',
    password: PGBOUNCER_SQL_PASSWORD_HASH,
    dbname: 'praefect_production',
    pool_mode: 'session'
  },

  ...
}

# Allow the praefect user to connect to PgBouncer
pgbouncer['users'] = {
  'praefect': {
    'password': PRAEFECT_SQL_PASSWORD_HASH,
  }
}
```

Both `praefect_production` and `praefect_production_direct` use the same database endpoint
(`praefect_production`), but with different pool modes. This translates to the following
`databases` section of PgBouncer:

```ini
[databases]
praefect_production = host=POSTGRESQL_HOST auth_user=pgbouncer pool_mode=transaction
praefect_production_direct = host=POSTGRESQL_HOST auth_user=pgbouncer dbname=praefect_production pool_mode=session
```

Now you can configure Praefect to use PgBouncer for both connections:

```ruby
praefect['configuration'] = {
   # ...
   database: {
      # ...
      host: PGBOUNCER_HOST,
      port: 6432,
      user: 'praefect',
      # `PRAEFECT_SQL_PASSWORD` is the plain-text password of
      # Praefect user. Not to be confused with `PRAEFECT_SQL_PASSWORD_HASH`.
      password: PRAEFECT_SQL_PASSWORD,
      dbname: 'praefect_production',
      session_pooled: {
         # ...
         dbname: 'praefect_production_direct',
         # There is no need to repeat the following. Parameters of direct
         # database connection will fall back to the values above.
         #
         # host: PGBOUNCER_HOST,
         # port: 6432,
         # user: 'praefect',
         # password: PRAEFECT_SQL_PASSWORD,
      },
   },
}
```

With this configuration, Praefect uses PgBouncer for both connection types.

NOTE:
Omnibus GitLab handles the authentication requirements (using `auth_query`), but if you are preparing
your databases manually and configuring an external PgBouncer, you must include `praefect` user and
its password in the file used by PgBouncer. For example, `userlist.txt` if the [`auth_file`](https://www.pgbouncer.org/config.html#auth_file)
configuration option is set. For more details, consult the PgBouncer documentation.

#### Configure Praefect to connect directly to PostgreSQL

As an alternative to configuring PgBouncer with `session` pool mode, Praefect can be configured to use different
connection parameters for direct access to PostgreSQL. This connection supports the `LISTEN` feature.

An example of Praefect configuration that bypasses PgBouncer and directly connects to PostgreSQL:

```ruby
praefect['configuration'] = {
   # ...
   database: {
      # ...
      session_pooled: {
         # ...
         host: POSTGRESQL_HOST,
         port: 5432,

         # Use the following to override parameters of direct database connection.
         # Comment out where the parameters are the same for both connections.
         #
         user: 'praefect',
         password: PRAEFECT_SQL_PASSWORD,
         dbname: 'praefect_production',
         # sslmode: '...',
         # sslcert: '...',
         # sslkey: '...',
         # sslrootcert: '...',
      },
   },
}
```

### Praefect

> [Introduced](https://gitlab.com/gitlab-org/gitaly/-/issues/2634) in GitLab 13.4, Praefect nodes can no longer be designated as `primary`.

If there are multiple Praefect nodes:

1. Designate one node as the deploy node, and configure it using the following steps.
1. Complete the following steps for each additional node.

To complete this section you need a [configured PostgreSQL server](#postgresql), including:

WARNING:
Praefect should be run on a dedicated node. Do not run Praefect on the
application server, or a Gitaly node.

On the **Praefect** node:

1. Disable all other services by editing `/etc/gitlab/gitlab.rb`:

<!--
Updates to example must be made at:
- https://gitlab.com/gitlab-org/gitlab/-/blob/master/doc/administration/gitaly/praefect.md
- all reference architecture pages
-->

   ```ruby
   # Avoid running unnecessary services on the Praefect server
   gitaly['enable'] = false
   postgresql['enable'] = false
   redis['enable'] = false
   nginx['enable'] = false
   puma['enable'] = false
   sidekiq['enable'] = false
   gitlab_workhorse['enable'] = false
   prometheus['enable'] = false
   alertmanager['enable'] = false
   grafana['enable'] = false
   gitlab_exporter['enable'] = false
   gitlab_kas['enable'] = false

   # Enable only the Praefect service
   praefect['enable'] = true

   # Prevent database migrations from running on upgrade automatically
   praefect['auto_migrate'] = false
   gitlab_rails['auto_migrate'] = false
   ```

1. Configure **Praefect** to listen on network interfaces by editing
   `/etc/gitlab/gitlab.rb`:

   ```ruby
   praefect['configuration'] = {
      # ...
      listen_addr: '0.0.0.0:2305',
   }
   ```

1. Configure Prometheus metrics by editing
   `/etc/gitlab/gitlab.rb`:

   ```ruby
   praefect['configuration'] = {
      # ...
      #
      # Enable Prometheus metrics access to Praefect. You must use firewalls
      # to restrict access to this address/port.
      # The default metrics endpoint is /metrics
      prometheus_listen_addr: '0.0.0.0:9652',
      # Some metrics run queries against the database. Enabling separate database metrics allows
      # these metrics to be collected when the metrics are
      # scraped on a separate /db_metrics endpoint.
      prometheus_exclude_database_from_default_metrics: true,
   }
   ```

1. Configure a strong authentication token for **Praefect** by editing
   `/etc/gitlab/gitlab.rb`, which is needed by clients outside the cluster
   (like GitLab Shell) to communicate with the Praefect cluster:

   ```ruby
   praefect['configuration'] = {
      # ...
      auth: {
         # ...
         token: 'PRAEFECT_EXTERNAL_TOKEN',
      },
   }
   ```

1. Configure **Praefect** to [connect to the PostgreSQL database](#postgresql). We
   highly recommend using [PgBouncer](#use-pgbouncer) as well.

   If you want to use a TLS client certificate, the options below can be used:

   ```ruby
   praefect['configuration'] = {
      # ...
      database: {
         # ...
         #
         # Connect to PostgreSQL using a TLS client certificate
         # sslcert: '/path/to/client-cert',
         # sslkey: '/path/to/client-key',
         #
         # Trust a custom certificate authority
         # sslrootcert: '/path/to/rootcert',
      },
   }
   ```

   By default, Praefect refuses to make an unencrypted connection to
   PostgreSQL. You can override this by uncommenting the following line:

   ```ruby
   praefect['configuration'] = {
      # ...
      database: {
         # ...
         # sslmode: 'disable',
      },
   }
   ```

1. Configure the **Praefect** cluster to connect to each Gitaly node in the
   cluster by editing `/etc/gitlab/gitlab.rb`.

   The virtual storage's name must match the configured storage name in GitLab
   configuration. In a later step, we configure the storage name as `default`
   so we use `default` here as well. This cluster has three Gitaly nodes `gitaly-1`,
   `gitaly-2`, and `gitaly-3`, which are intended to be replicas of each other.

   WARNING:
   If you have data on an already existing storage called
   `default`, you should configure the virtual storage with another name and
   [migrate the data to the Gitaly Cluster storage](index.md#migrate-to-gitaly-cluster)
   afterwards.

   Replace `PRAEFECT_INTERNAL_TOKEN` with a strong secret, which is used by
   Praefect when communicating with Gitaly nodes in the cluster. This token is
   distinct from the `PRAEFECT_EXTERNAL_TOKEN`.

   Replace `GITALY_HOST_*` with the IP or host address of the each Gitaly node.

   More Gitaly nodes can be added to the cluster to increase the number of
   replicas. More clusters can also be added for very large GitLab instances.

   NOTE:
   When adding additional Gitaly nodes to a virtual storage, all storage names
   in that virtual storage must be unique. Additionally, all Gitaly node
   addresses referenced in the Praefect configuration must be unique.

   ```ruby
   # Name of storage hash must match storage name in git_data_dirs on GitLab
   # server ('default') and in gitaly['configuration'][:storage][INDEX][:name] on Gitaly nodes ('gitaly-1')
   praefect['configuration'] = {
      # ...
      virtual_storage: [
         {
            # ...
            name: 'default',
            node: [
               {
                  storage: 'gitaly-1',
                  address: 'tcp://GITALY_HOST_1:8075',
                  token: 'PRAEFECT_INTERNAL_TOKEN'
               },
               {
                  storage: 'gitaly-2',
                  address: 'tcp://GITALY_HOST_2:8075',
                  token: 'PRAEFECT_INTERNAL_TOKEN'
               },
               {
                  storage: 'gitaly-3',
                  address: 'tcp://GITALY_HOST_3:8075',
                  token: 'PRAEFECT_INTERNAL_TOKEN'
               },
            ],
         },
      ],
   }
   ```

1. Enable [distribution of reads](index.md#distributed-reads).

1. Save the changes to `/etc/gitlab/gitlab.rb` and
   [reconfigure Praefect](../restart_gitlab.md#omnibus-gitlab-reconfigure):

   ```shell
   gitlab-ctl reconfigure
   ```

1. For:

   - The "deploy node":
     1. Enable Praefect database auto-migration again by setting `praefect['auto_migrate'] = true` in
        `/etc/gitlab/gitlab.rb`.
     1. To ensure database migrations are only run during reconfigure and not automatically on
        upgrade, run:

        ```shell
        sudo touch /etc/gitlab/skip-auto-reconfigure
        ```

   - The other nodes, you can leave the settings as they are. Though
     `/etc/gitlab/skip-auto-reconfigure` isn't required, you may want to set it to prevent GitLab
     running reconfigure automatically when running commands such as `apt-get update`. This way any
     additional configuration changes can be done and then reconfigure can be run manually.

1. Save the changes to `/etc/gitlab/gitlab.rb` and
   [reconfigure Praefect](../restart_gitlab.md#omnibus-gitlab-reconfigure):

   ```shell
   gitlab-ctl reconfigure
   ```

1. To ensure that Praefect
   [has updated its Prometheus listen address](https://gitlab.com/gitlab-org/gitaly/-/issues/2734),
   [restart Praefect](../restart_gitlab.md#omnibus-gitlab-restart):

   ```shell
   gitlab-ctl restart praefect
   ```

1. Verify that Praefect can reach PostgreSQL:

   ```shell
   sudo -u git /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml sql-ping
   ```

   If the check fails, make sure you have followed the steps correctly. If you
   edit `/etc/gitlab/gitlab.rb`, remember to run `sudo gitlab-ctl reconfigure`
   again before trying the `sql-ping` command.

#### Enable TLS support

Praefect supports TLS encryption. To communicate with a Praefect instance that listens
for secure connections, you must:

- Ensure Gitaly is [configured for TLS](configure_gitaly.md#enable-tls-support) and use a `tls://` URL scheme in the `gitaly_address`
  of the corresponding storage entry in the GitLab configuration.
- Bring your own certificates because this isn't provided automatically. The certificate
  corresponding to each Praefect server must be installed on that Praefect server.

Additionally the certificate, or its certificate authority, must be installed on all Gitaly servers
and on all Praefect clients that communicate with it following the procedure described in
[GitLab custom certificate configuration](https://docs.gitlab.com/omnibus/settings/ssl/index.html#install-custom-public-certificates) (and repeated below).

Note the following:

- The certificate must specify the address you use to access the Praefect server. You must add the hostname or IP
  address as a Subject Alternative Name to the certificate.
- When running Praefect sub-commands such as `dial-nodes` and `list-untracked-repositories` from the command line with
  [Gitaly TLS enabled](configure_gitaly.md#enable-tls-support), you must set the `SSL_CERT_DIR` or `SSL_CERT_FILE`
  environment variable so that the Gitaly certificate is trusted. For example:

   ```shell
   sudo SSL_CERT_DIR=/etc/gitlab/trusted-certs /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml dial-nodes
   ```

- You can configure Praefect servers with both an unencrypted listening address
  `listen_addr` and an encrypted listening address `tls_listen_addr` at the same time.
  This allows you to do a gradual transition from unencrypted to encrypted traffic, if
  necessary.

  To disable the unencrypted listener, set:

  ```ruby
  praefect['configuration'] = {
    # ...
    listen_addr: nil,
  }
  ```

To configure Praefect with TLS:

**For Omnibus GitLab**

1. Create certificates for Praefect servers.

1. On the Praefect servers, create the `/etc/gitlab/ssl` directory and copy your key
   and certificate there:

   ```shell
   sudo mkdir -p /etc/gitlab/ssl
   sudo chmod 755 /etc/gitlab/ssl
   sudo cp key.pem cert.pem /etc/gitlab/ssl/
   sudo chmod 644 key.pem cert.pem
   ```

1. Edit `/etc/gitlab/gitlab.rb` and add:

   ```ruby
   praefect['configuration'] = {
      # ...
      tls_listen_addr: '0.0.0.0:3305',
      tls: {
         # ...
         certificate_path: '/etc/gitlab/ssl/cert.pem',
         key_path: '/etc/gitlab/ssl/key.pem',
      },
   }
   ```

1. Save the file and [reconfigure](../restart_gitlab.md#omnibus-gitlab-reconfigure).

1. On the Praefect clients (including each Gitaly server), copy the certificates,
   or their certificate authority, into `/etc/gitlab/trusted-certs`:

   ```shell
   sudo cp cert.pem /etc/gitlab/trusted-certs/
   ```

1. On the Praefect clients (except Gitaly servers), edit `git_data_dirs` in
   `/etc/gitlab/gitlab.rb` as follows:

   ```ruby
   git_data_dirs({
     "default" => {
       "gitaly_address" => 'tls://PRAEFECT_LOADBALANCER_HOST:3305',
       "gitaly_token" => 'PRAEFECT_EXTERNAL_TOKEN'
     }
   })
   ```

1. Save the file and [reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure).

**For installations from source**

1. Create certificates for Praefect servers.
1. On the Praefect servers, create the `/etc/gitlab/ssl` directory and copy your key and certificate
   there:

   ```shell
   sudo mkdir -p /etc/gitlab/ssl
   sudo chmod 755 /etc/gitlab/ssl
   sudo cp key.pem cert.pem /etc/gitlab/ssl/
   sudo chmod 644 key.pem cert.pem
   ```

1. On the Praefect clients (including each Gitaly server), copy the certificates,
   or their certificate authority, into the system trusted certificates:

   ```shell
   sudo cp cert.pem /usr/local/share/ca-certificates/praefect.crt
   sudo update-ca-certificates
   ```

1. On the Praefect clients (except Gitaly servers), edit `storages` in
   `/home/git/gitlab/config/gitlab.yml` as follows:

   ```yaml
   gitlab:
     repositories:
       storages:
         default:
           gitaly_address: tls://PRAEFECT_LOADBALANCER_HOST:3305
           path: /some/local/path
   ```

   NOTE:
   `/some/local/path` should be set to a local folder that exists, however no
   data is stored in this folder. This requirement is scheduled to be removed when
   [this issue](https://gitlab.com/gitlab-org/gitaly/-/issues/1282) is resolved.

1. Save the file and [restart GitLab](../restart_gitlab.md#installations-from-source).
1. Copy all Praefect server certificates, or their certificate authority, to the system
   trusted certificates on each Gitaly server so the Praefect server trusts the
   certificate when called by Gitaly servers:

   ```shell
   sudo cp cert.pem /usr/local/share/ca-certificates/praefect.crt
   sudo update-ca-certificates
   ```

1. Edit `/home/git/praefect/config.toml` and add:

   ```toml
   tls_listen_addr = '0.0.0.0:3305'

   [tls]
   certificate_path = '/etc/gitlab/ssl/cert.pem'
   key_path = '/etc/gitlab/ssl/key.pem'
   ```

1. Save the file and [restart GitLab](../restart_gitlab.md#installations-from-source).

#### Service discovery

> [Introduced](https://gitlab.com/groups/gitlab-org/-/epics/8971) in GitLab 15.10.

Prerequisites:

- A DNS server.

GitLab uses service discovery to retrieve a list of Praefect hosts. Service
discovery involves periodic checks of a DNS A or AAAA record, with the IPs
retrieved from the record serving as the addresses of the target nodes.
Praefect does not support service discovery by SRV record.

By default, the minimum time between checks is 5 minutes, regardless of the
records' TTLs. Praefect does not support customizing this interval. When clients
receive an update, they:

- Establish new connections to the new IP addresses.
- Keep existing connections to intact IP addresses.
- Drop connections to removed IP addresses.

In-flight requests on to-be-removed connections are still handled until they
finish. Workhorse has a 10-minute timeout, while other clients do not specify a
graceful timeout.

The DNS server should return all IP addresses instead of load-balancing itself.
Clients can distribute requests to IP addresses in a round-robin fashion.

Before updating client configuration, ensure that DNS service discovery works
correctly. It should return the list of IP addresses correctly. `dig` is a good
tool to use to verify.

```console
❯ dig A praefect.service.consul @127.0.0.1

; <<>> DiG 9.10.6 <<>> A praefect.service.consul @127.0.0.1
;; global options: +cmd
;; Got answer:
;; ->>HEADER<<- opcode: QUERY, status: NOERROR, id: 29210
;; flags: qr aa rd ra; QUERY: 1, ANSWER: 3, AUTHORITY: 0, ADDITIONAL: 1

;; OPT PSEUDOSECTION:
; EDNS: version: 0, flags:; udp: 4096
;; QUESTION SECTION:
;praefect.service.consul.                     IN      A

;; ANSWER SECTION:
praefect.service.consul.              0       IN      A       10.0.0.3
praefect.service.consul.              0       IN      A       10.0.0.2
praefect.service.consul.              0       IN      A       10.0.0.1

;; Query time: 0 msec
;; SERVER: ::1#53(::1)
;; WHEN: Wed Dec 14 12:53:58 +07 2022
;; MSG SIZE  rcvd: 86
```

##### Configure service discovery

By default, Praefect delegates DNS resolution to the operating system. In such
cases, the Gitaly address can be set in either of these formats:

- `dns:[host]:[port]`
- `dns:///[host]:[port]` (note the three slashes)

You can also appoint an authoritative name server by setting it in this format:

- `dns://[authority_host]:[authority_port]/[host]:[port]`

::Tabs

:::TabTitle Linux package (Omnibus)

1. Edit `/etc/gitlab/gitlab.rb` and add:

   ```ruby
   praefect['consul_service_name'] = 'praefect'
   ```

1. Save the file and [reconfigure](../restart_gitlab.md#omnibus-gitlab-reconfigure).
1. On the Praefect clients (except Gitaly servers), edit `git_data_dirs` in
`/etc/gitlab/gitlab.rb` as follows. Replace `PRAEFECT_SERVICE_DISCOVERY_ADDRESS`
with Praefect service discovery address, such as `praefect.service.consul`.

   ```ruby
   git_data_dirs({
     "default" => {
       "gitaly_address" => 'dns:PRAEFECT_SERVICE_DISCOVERY_ADDRESS:2305',
       "gitaly_token" => 'PRAEFECT_EXTERNAL_TOKEN'
     }
   })
   ```

1. Save the file and [reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure).

:::TabTitle Self-compiled (source)

1. Install a DNS service discovery service. Register all Praefect nodes with the service.
1. On the Praefect clients (except Gitaly servers), edit `storages` in
   `/home/git/gitlab/config/gitlab.yml` as follows:

   ```yaml
   gitlab:
     repositories:
       storages:
         default:
           gitaly_address: dns:PRAEFECT_SERVICE_DISCOVERY_ADDRESS:2305
           path: /some/local/path
   ```

   NOTE:
   `/some/local/path` should be set to a local folder that exists, however no
   data is stored in this folder. [Issue 375254](https://gitlab.com/gitlab-org/gitlab/-/issues/375254)
   proposes to remove this requirement.

1. Save the file and [restart GitLab](../restart_gitlab.md#installations-from-source).

::EndTabs

### Gitaly

NOTE:
Complete these steps for **each** Gitaly node.

To complete this section you need:

- [Configured Praefect node](#praefect)
- 3 (or more) servers, with GitLab installed, to be configured as Gitaly nodes.
  These should be dedicated nodes, do not run other services on these nodes.

Every Gitaly server assigned to the Praefect cluster needs to be configured. The
configuration is the same as a standard [standalone Gitaly server](index.md),
except:

- The storage names are exposed to Praefect, not GitLab
- The secret token is shared with Praefect, not GitLab

The configuration of all Gitaly nodes in the Praefect cluster can be identical,
because we rely on Praefect to route operations correctly.

Particular attention should be shown to:

- The `gitaly['configuration'][:auth][:token]` configured in this section must match the `token`
  value under `praefect['configuration'][:virtual_storage][<index>][:node][<index>][:token]` on the Praefect node. This value was
  set in the [previous section](#praefect). This document uses the placeholder `PRAEFECT_INTERNAL_TOKEN` throughout.
- The storage names in `gitaly['configuration'][:storage]` configured in this section must match the
  storage names under `praefect['configuration'][:virtual_storage]` on the Praefect node. This
  was set in the [previous section](#praefect). This document uses `gitaly-1`,
  `gitaly-2`, and `gitaly-3` as Gitaly storage names.

For more information on Gitaly server configuration, see our
[Gitaly documentation](configure_gitaly.md#configure-gitaly-servers).

1. SSH into the **Gitaly** node and login as root:

   ```shell
   sudo -i
   ```

1. Disable all other services by editing `/etc/gitlab/gitlab.rb`:

   ```ruby
   # Disable all other services on the Gitaly node
   postgresql['enable'] = false
   redis['enable'] = false
   nginx['enable'] = false
   grafana['enable'] = false
   puma['enable'] = false
   sidekiq['enable'] = false
   gitlab_workhorse['enable'] = false
   prometheus_monitoring['enable'] = false
   gitlab_kas['enable'] = false

   # Enable only the Gitaly service
   gitaly['enable'] = true

   # Enable Prometheus if needed
   prometheus['enable'] = true

   # Disable database migrations to prevent database connections during 'gitlab-ctl reconfigure'
   gitlab_rails['auto_migrate'] = false
   ```

1. Configure **Gitaly** to listen on network interfaces by editing
   `/etc/gitlab/gitlab.rb`:

   ```ruby
   gitaly['configuration'] = {
      # ...
      #
      # Make Gitaly accept connections on all network interfaces.
      # Use firewalls to restrict access to this address/port.
      listen_addr: '0.0.0.0:8075',
      # Enable Prometheus metrics access to Gitaly. You must use firewalls
      # to restrict access to this address/port.
      prometheus_listen_addr: '0.0.0.0:9236',
   }
   ```

1. Configure a strong `auth_token` for **Gitaly** by editing
   `/etc/gitlab/gitlab.rb`, which is needed by clients to communicate with
   this Gitaly nodes. Typically, this token is the same for all Gitaly
   nodes.

   ```ruby
   gitaly['configuration'] = {
      # ...
      auth: {
         # ...
         token: 'PRAEFECT_INTERNAL_TOKEN',
      },
   }
   ```

1. Configure the GitLab Shell secret token, which is needed for `git push` operations. Either:

   - Method 1:

     1. Copy `/etc/gitlab/gitlab-secrets.json` from the Gitaly client to same path on the Gitaly
        servers and any other Gitaly clients.
     1. [Reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure) on Gitaly servers.

   - Method 2:

     1. Edit `/etc/gitlab/gitlab.rb`.
     1. Replace `GITLAB_SHELL_SECRET_TOKEN` with the real secret.

        ```ruby
        gitlab_shell['secret_token'] = 'GITLAB_SHELL_SECRET_TOKEN'
        ```

1. Configure an `internal_api_url`, which is also needed for `git push` operations:

   ```ruby
   # Configure the gitlab-shell API callback URL. Without this, `git push` will
   # fail. This can be your front door GitLab URL or an internal load balancer.
   # Examples: 'https://gitlab.example.com', 'http://1.2.3.4'
   gitlab_rails['internal_api_url'] = 'http://GITLAB_HOST'
   ```

1. Configure the storage location for Git data by setting `gitaly['configuration'][:storage]` in
   `/etc/gitlab/gitlab.rb`. Each Gitaly node should have a unique storage name
   (such as `gitaly-1`).

   Instead of configuring `gitaly['configuration'][:storage]` uniquely for each Gitaly node, it is
   often easier to have include the configuration for all Gitaly nodes on every
   Gitaly node. You can do this because the Praefect `virtual_storage`
   configuration maps each storage name (such as `gitaly-1`) to a specific node, and
   requests are routed accordingly. This means every Gitaly node in your fleet
   can share the same configuration.

   ```ruby
   # You can include the data dirs for all nodes in the same config, because
   # Praefect will only route requests according to the addresses provided in the
   # prior step.
   gitaly['configuration'] = {
      # ...
      storage: [
        {
          name: 'gitaly-1',
          path: '/var/opt/gitlab/git-data',
        },
        {
          name: 'gitaly-2',
          path: '/var/opt/gitlab/git-data',
        },
        {
          name: 'gitaly-3',
          path: '/var/opt/gitlab/git-data',
        },
      ],
   }
   ```

1. Save the changes to `/etc/gitlab/gitlab.rb` and
   [reconfigure Gitaly](../restart_gitlab.md#omnibus-gitlab-reconfigure):

   ```shell
   gitlab-ctl reconfigure
   ```

1. To ensure that Gitaly
   [has updated its Prometheus listen address](https://gitlab.com/gitlab-org/gitaly/-/issues/2734),
   [restart Gitaly](../restart_gitlab.md#omnibus-gitlab-restart):

   ```shell
   gitlab-ctl restart gitaly
   ```

**The steps above must be completed for each Gitaly node!**

After all Gitaly nodes are configured, run the Praefect connection
checker to verify Praefect can connect to all Gitaly servers in the Praefect
configuration.

1. SSH into each **Praefect** node and run the Praefect connection checker:

   ```shell
   sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml dial-nodes
   ```

### Load Balancer

In a fault-tolerant Gitaly configuration, a load balancer is needed to route
internal traffic from the GitLab application to the Praefect nodes. The
specifics on which load balancer to use or the exact configuration is beyond the
scope of the GitLab documentation.

NOTE:
The load balancer must be configured to accept traffic from the Gitaly nodes in
addition to the GitLab nodes.

We hope that if you're managing fault-tolerant systems like GitLab, you have a load balancer
of choice already. Some examples include [HAProxy](https://www.haproxy.org/)
(open-source), [Google Internal Load Balancer](https://cloud.google.com/load-balancing/docs/internal/),
[AWS Elastic Load Balancer](https://aws.amazon.com/elasticloadbalancing/), F5
Big-IP LTM, and Citrix Net Scaler. This documentation outlines what ports
and protocols you need configure.

NOTE:
You should use the equivalent of HAProxy `leastconn` load-balancing strategy because long-running operations (for
example, clones) keep some connections open for extended periods.

| LB Port | Backend Port | Protocol |
|:--------|:-------------|:---------|
| 2305    | 2305         | TCP      |

### GitLab

To complete this section you need:

- [Configured Praefect node](#praefect)
- [Configured Gitaly nodes](#gitaly)

The Praefect cluster needs to be exposed as a storage location to the GitLab
application, which is done by updating the `git_data_dirs`.

Particular attention should be shown to:

- the storage name added to `git_data_dirs` in this section must match the
  storage name under `praefect['configuration'][:virtual_storage]` on the Praefect nodes. This
  was set in the [Praefect](#praefect) section of this guide. This document uses
  `default` as the Praefect storage name.

1. SSH into the **GitLab** node and login as root:

   ```shell
   sudo -i
   ```

1. Configure the `external_url` so that files could be served by GitLab
   by proper endpoint access by editing `/etc/gitlab/gitlab.rb`:

   You need to replace `GITLAB_SERVER_URL` with the real external facing
   URL on which current GitLab instance is serving:

   ```ruby
   external_url 'GITLAB_SERVER_URL'
   ```

1. Disable the default Gitaly service running on the GitLab host. It isn't needed
   because GitLab connects to the configured cluster.

   WARNING:
   If you have existing data stored on the default Gitaly storage,
   you should [migrate the data to your Gitaly Cluster storage](index.md#migrate-to-gitaly-cluster)
   first.

   ```ruby
   gitaly['enable'] = false
   ```

1. Add the Praefect cluster as a storage location by editing
   `/etc/gitlab/gitlab.rb`.

   You need to replace:

   - `PRAEFECT_LOADBALANCER_HOST` with the IP address or hostname of the load
     balancer.
   - `PRAEFECT_EXTERNAL_TOKEN` with the real secret

   If you are using TLS:

   - The `gitaly_address` should begin with `tls://` instead.
   - The port should be changed to `3305`.

   ```ruby
   git_data_dirs({
     "default" => {
       "gitaly_address" => "tcp://PRAEFECT_LOADBALANCER_HOST:2305",
       "gitaly_token" => 'PRAEFECT_EXTERNAL_TOKEN'
     }
   })
   ```

1. Configure the GitLab Shell secret token so that callbacks from Gitaly nodes during a `git push`
   are properly authenticated. Either:

   - Method 1:

     1. Copy `/etc/gitlab/gitlab-secrets.json` from the Gitaly client to same path on the Gitaly
        servers and any other Gitaly clients.
     1. [Reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure) on Gitaly servers.

   - Method 2:

     1. Edit `/etc/gitlab/gitlab.rb`.
     1. Replace `GITLAB_SHELL_SECRET_TOKEN` with the real secret.

        ```ruby
        gitlab_shell['secret_token'] = 'GITLAB_SHELL_SECRET_TOKEN'
        ```

1. Add Prometheus monitoring settings by editing `/etc/gitlab/gitlab.rb`. If Prometheus
   is enabled on a different node, make edits on that node instead.

   You need to replace:

   - `PRAEFECT_HOST` with the IP address or hostname of the Praefect node
   - `GITALY_HOST_*` with the IP address or hostname of each Gitaly node

   ```ruby
   prometheus['scrape_configs'] = [
     {
       'job_name' => 'praefect',
       'static_configs' => [
         'targets' => [
           'PRAEFECT_HOST:9652', # praefect-1
           'PRAEFECT_HOST:9652', # praefect-2
           'PRAEFECT_HOST:9652', # praefect-3
         ]
       ]
     },
     {
       'job_name' => 'praefect-gitaly',
       'static_configs' => [
         'targets' => [
           'GITALY_HOST_1:9236', # gitaly-1
           'GITALY_HOST_2:9236', # gitaly-2
           'GITALY_HOST_3:9236', # gitaly-3
         ]
       ]
     }
   ]
   ```

1. Save the changes to `/etc/gitlab/gitlab.rb` and [reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure):

   ```shell
   gitlab-ctl reconfigure
   ```

1. Verify on each Gitaly node the Git Hooks can reach GitLab. On each Gitaly node run:
   - For GitLab 15.3 and later, run `sudo /opt/gitlab/embedded/bin/gitaly check /var/opt/gitlab/gitaly/config.toml`.
   - For GitLab 15.2 and earlier, run `sudo /opt/gitlab/embedded/bin/gitaly-hooks check /var/opt/gitlab/gitaly/config.toml`.

1. Verify that GitLab can reach Praefect:

   ```shell
   gitlab-rake gitlab:gitaly:check
   ```

1. Check that the Praefect storage is configured to store new repositories:

   1. On the top bar, select **Main menu > Admin**.
   1. On the left sidebar, select **Settings > Repository**.
   1. Expand the **Repository storage** section.

   Following this guide, the `default` storage should have weight 100 to store all new repositories.

1. Verify everything is working by creating a new project. Check the
   "Initialize repository with a README" box so that there is content in the
   repository that viewed. If the project is created, and you can see the
   README file, it works!

#### Use TCP for existing GitLab instances

When adding Gitaly Cluster to an existing Gitaly instance, the existing Gitaly storage
must be listening on TCP/TLS. If `gitaly_address` is not specified, then a Unix socket is used,
which prevents the communication with the cluster.

For example:

```ruby
git_data_dirs({
  'default' => { 'gitaly_address' => 'tcp://old-gitaly.internal:8075' },
  'cluster' => {
    'gitaly_address' => 'tls://<PRAEFECT_LOADBALANCER_HOST>:3305',
    'gitaly_token' => '<praefect_external_token>'
  }
})
```

See [Mixed Configuration](configure_gitaly.md#mixed-configuration) for further information on
running multiple Gitaly storages.

### Grafana

Grafana is included with GitLab, and can be used to monitor your Praefect
cluster. See [Grafana Dashboard Service](https://docs.gitlab.com/omnibus/settings/grafana.html)
for detailed documentation.

To get started quickly:

1. SSH into the **GitLab** node (or whichever node has Grafana enabled) and login as root:

   ```shell
   sudo -i
   ```

1. Enable the Grafana login form by editing `/etc/gitlab/gitlab.rb`.

   ```ruby
   grafana['disable_login_form'] = false
   ```

1. Save the changes to `/etc/gitlab/gitlab.rb` and
   [reconfigure GitLab](../restart_gitlab.md#omnibus-gitlab-reconfigure):

   ```shell
   gitlab-ctl reconfigure
   ```

1. Set the Grafana administrator password. This command prompts you to enter a new
   password:

   ```shell
   gitlab-ctl set-grafana-password
   ```

1. In your web browser, open `/-/grafana` (such as
   `https://gitlab.example.com/-/grafana`) on your GitLab server.

   Login using the password you set, and the username `admin`.

1. Go to **Explore** and query `gitlab_build_info` to verify that you are
   getting metrics from all your machines.

Congratulations! You've configured an observable fault-tolerant Praefect
cluster.

## Configure replication factor

WARNING:
Configurable replication factors require [repository-specific primary nodes](#repository-specific-primary-nodes) to be used.

Praefect supports configuring a replication factor on a per-repository basis, by assigning
specific storage nodes to host a repository.

Praefect does not store the actual replication factor, but assigns enough storages to host the repository
so the desired replication factor is met. If a storage node is later removed from the virtual storage,
the replication factor of repositories assigned to the storage is decreased accordingly.

You can configure:

- A default replication factor for each virtual storage that is applied to newly-created repositories.
  The configuration is added to the `/etc/gitlab/gitlab.rb` file:

  ```ruby
   praefect['configuration'] = {
      # ...
      virtual_storage: [
         {
            # ...
            name: 'default',
            default_replication_factor: 1,
         },
      ],
   }
  ```

- A replication factor for an existing repository using the `set-replication-factor` sub-command.
  `set-replication-factor` automatically assigns or unassigns random storage nodes as
  necessary to reach the desired replication factor. The repository's primary node is
  always assigned first and is never unassigned.

  ```shell
  sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml set-replication-factor -virtual-storage <virtual-storage> -repository <relative-path> -replication-factor <replication-factor>
  ```

  - `-virtual-storage` is the virtual storage the repository is located in.
  - `-repository` is the repository's relative path in the storage.
  - `-replication-factor` is the desired replication factor of the repository. The minimum value is
    `1`, as the primary needs a copy of the repository. The maximum replication factor is the number of
    storages in the virtual storage.

  On success, the assigned host storages are printed. For example:

  ```shell
  $ sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml set-replication-factor -virtual-storage default -repository @hashed/3f/db/3fdba35f04dc8c462986c992bcf875546257113072a909c162f7e470e581e278.git -replication-factor 2

  current assignments: gitaly-1, gitaly-2
  ```

If `default_replication_factor` is unset, the repositories are always replicated on every node defined in `virtual_storages`. If a new
node is introduced to the virtual storage, both new and existing repositories are replicated to the node automatically.

### Repository storage recommendations

The size of the required storage can vary between instances and depends on the set
[replication factor](index.md#replication-factor). You might want to include implementing
repository storage redundancy.

For a replication factor:

- Of `1`: NFS, Gitaly, and Gitaly Cluster have roughly the same storage requirements.
- More than `1`: The amount of required storage is `used space * replication factor`. `used space`
  should include any planned future growth.

## Repository verification

> [Introduced](https://gitlab.com/gitlab-org/gitaly/-/issues/4080) in GitLab 15.0.

Praefect stores metadata about the repositories in a database. If the repositories are modified on disk
without going through Praefect, the metadata can become inaccurate. Because the metadata is used for replication
and routing decisions, any inaccuracies may cause problems. Praefect contains a background worker that
periodically verifies the metadata against the actual state on the disks. The worker:

1. Picks up a batch of replicas to verify on healthy storages. The replicas are either unverified or have exceeded
   the configured verification interval. Replicas that have never been verified are prioritized, followed by
   the other replicas ordered by longest time since the last successful verification.
1. Checks whether the replicas exist on their respective storages. If the:
   - Replica exists, update its last successful verification time.
   - Replica doesn't exist, remove its metadata record.
   - Check failed, the replica is picked up for verification again when the next worker dequeues more work.

The worker acquires an exclusive verification lease on each of the replicas it is about to verify. This avoids multiple
workers from verifying the same replica concurrently. The worker releases the leases when it has completed its check.
Praefect contains a background goroutine that releases stale leases every 10 seconds when workers are terminated for
some reason without releasing the lease.

The worker logs each of the metadata removals prior to executing them. The `perform_deletions` key
indicates whether the invalid metadata records are actually deleted or not. For example:

```json
{
  "level": "info",
  "msg": "removing metadata records of non-existent replicas",
  "perform_deletions": false,
  "replicas": {
    "default": {
      "@hashed/6b/86/6b86b273ff34fce19d6b804eff5a3f5747ada4eaa22f1d49c01e52ddb7875b4b.git": [
        "praefect-internal-0"
      ]
    }
  }
}
```

### Configure the verification worker

The worker is enabled by default and verifies the metadata records every seven days. The verification
interval is configurable with any valid [Go duration string](https://pkg.go.dev/time#ParseDuration).

To verify the metadata every three days:

```ruby
praefect['configuration'] = {
   # ...
   background_verification: {
      # ...
      verification_interval: '72h',
   },
}
```

Values of 0 and below disable the background verifier.

```ruby
praefect['configuration'] = {
   # ...
   background_verification: {
      # ...
      verification_interval: '0',
   },
}
```

#### Enable deletions

> - [Introduced](https://gitlab.com/gitlab-org/gitaly/-/issues/4080) and disabled by default in GitLab 15.0
> - [Default enabled](https://gitlab.com/gitlab-org/gitaly/-/merge_requests/5321) in GitLab 15.9.

WARNING:
Deletions were disabled by default prior to GitLab 15.9 due to a race condition with repository renames
that can cause incorrect deletions, which is especially prominent in Geo instances as Geo performs more renames
than instances without Geo. In GitLab 15.0 to 15.5, you should enable deletions only if the [`gitaly_praefect_generated_replica_paths` feature flag](index.md#praefect-generated-replica-paths-gitlab-150-and-later) is enabled. The feature flag was removed in GitLab 15.6 making deletions always safe to enable.

By default, the worker deletes invalid metadata records. It also logs the deleted records and outputs Prometheus
metrics.

You can disable deleting invalid metadata records with:

```ruby
praefect['configuration'] = {
   # ...
   background_verification: {
      # ...
      delete_invalid_records: false,
   },
}
```

### Prioritize verification manually

You can prioritize verification of some replicas ahead of their next scheduled verification time.
This might be needed after a disk failure, for example, when the administrator knows that the disk contents may have
changed. Praefect would eventually verify the replicas again, but users may encounter errors in the meantime.

To manually prioritize reverification of some replicas, use the `praefect verify` subcommand. The subcommand marks
replicas as unverified. Unverified replicas are prioritized by the background verification worker. The verification
worker must be enabled for the replicas to be verified.

Prioritize verifying the replicas of a specific repository:

```shell
sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml verify -repository-id=<repository-id>
```

Prioritize verifying all replicas stored on a virtual storage:

```shell
sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml verify -virtual-storage=<virtual-storage>
```

Prioritize verifying all replicas stored on a storage:

```shell
sudo /opt/gitlab/embedded/bin/praefect -config /var/opt/gitlab/praefect/config.toml verify -virtual-storage=<virtual-storage> -storage=<storage>
```

The output includes the number of replicas that were marked unverified.

## Automatic failover and primary election strategies

Praefect regularly checks the health of each Gitaly node, which is used to automatically fail over
to a newly-elected primary Gitaly node if the current primary node is found to be unhealthy.

You should use [repository-specific primary nodes](#repository-specific-primary-nodes). This is
[the only available election strategy](https://gitlab.com/gitlab-org/gitaly/-/issues/3574) from GitLab 14.0.

### Repository-specific primary nodes

> - [Introduced](https://gitlab.com/gitlab-org/gitaly/-/issues/3492) in GitLab 13.12, with primary elections run when Praefect starts or the cluster's consensus of a Gitaly node's health changes.
> - [Changed](https://gitlab.com/gitlab-org/gitaly/-/merge_requests/3543) in GitLab 14.1, primary elections are run lazily.

Gitaly Cluster supports electing repository-specific primary Gitaly nodes. Repository-specific
Gitaly primary nodes are enabled in `/etc/gitlab/gitlab.rb` by setting
`praefect['failover_election_strategy'] = 'per_repository'`.

Praefect's [deprecated election strategies](#deprecated-election-strategies):

- Elected a primary Gitaly node for each virtual storage, which was used as the primary node for
  each repository in the virtual storage.
- Prevented horizontal scaling of a virtual storage. The primary Gitaly node needed a replica of
  each repository and thus became the bottleneck.

The `per_repository` election strategy solves this problem by electing a primary Gitaly node separately for each
repository. Combined with [configurable replication factors](#configure-replication-factor), you can
horizontally scale storage capacity and distribute write load across Gitaly nodes.

Primary elections are run lazily. Praefect doesn't immediately elect a new primary node if the current
one is unhealthy. A new primary is elected if a request must be served while the current primary is unavailable.

A valid primary node candidate is a Gitaly node that:

- Is healthy. A Gitaly node is considered healthy if `>=50%` Praefect nodes have
  successfully health checked the Gitaly node in the previous ten seconds.
- Has a fully up to date copy of the repository.

If there are multiple primary node candidates, Praefect:

- Picks one of them randomly.
- Prioritizes promoting a Gitaly node that is assigned to host the repository. If
  there are no assigned Gitaly nodes to elect as the primary, Praefect may temporarily
  elect an unassigned one. The unassigned primary is demoted in favor of an assigned
  one when one becomes available.

If there are no valid primary candidates for a repository:

- The unhealthy primary node is demoted and the repository is left without a primary node.
- Operations that require a primary node fail until a primary is successfully elected.

#### Migrate to repository-specific primary Gitaly nodes

New Gitaly Clusters can start using the `per_repository` election strategy immediately.

To migrate existing clusters:

1. Praefect nodes didn't historically keep database records of every repository stored on the cluster. When
   the `per_repository` election strategy is configured, Praefect expects to have database records of
   each repository. A [background database migration](https://gitlab.com/gitlab-org/gitaly/-/merge_requests/2749)
   creates any missing database records for repositories. Before migrating, check Praefect's logs to verify
   that the database migration ran.

   Check Praefect's logs for `repository importer finished` message. The `virtual_storages` field contains
   the names of virtual storages and whether they've had any missing database records created.

   For example, the `default` virtual storage has been successfully migrated:

   ```json
   {"level":"info","msg":"repository importer finished","pid":19752,"time":"2021-04-28T11:41:36.743Z","virtual_storages":{"default":true}}
   ```

   If a virtual storage has not been successfully migrated, it would have `false` next to it:

   ```json
   {"level":"info","msg":"repository importer finished","pid":19752,"time":"2021-04-28T11:41:36.743Z","virtual_storages":{"default":false}}
   ```

   The database migration runs when Praefect starts. If the database migration is unsuccessful, you can restart
   a Praefect node to reattempt it.

1. Running two different election strategies side by side can cause a split brain, where different
   Praefect nodes consider repositories to have different primaries. This can be avoided either:

   - If a short downtime is acceptable:

      1. Shut down all Praefect nodes before changing the election strategy. Do this by running `gitlab-ctl stop praefect` on the Praefect nodes.

      1. On the Praefect nodes, configure the election strategy in `/etc/gitlab/gitlab.rb` with `praefect['failover_election_strategy'] = 'per_repository'`.

      1. Run `gitlab-ctl reconfigure && gitlab-ctl start` to reconfigure and start the Praefect nodes.

   - If downtime is unacceptable:

      1. Determine which Gitaly node is [the current primary](troubleshooting.md#determine-primary-gitaly-node).

      1. Comment out the secondary Gitaly nodes from the virtual storage's configuration in `/etc/gitlab/gitlab.rb`
      on all Praefect nodes. This ensures there's only one Gitaly node configured, causing both of the election
      strategies to elect the same Gitaly node as the primary.

      1. Run `gitlab-ctl reconfigure` on all Praefect nodes. Wait until all Praefect processes have restarted and
      the old processes have exited. This can take up to one minute.

      1. On all Praefect nodes, configure the election strategy in `/etc/gitlab/gitlab.rb` with
      `praefect['failover_election_strategy'] = 'per_repository'`.

      1. Run `gitlab-ctl reconfigure` on all Praefect nodes. Wait until all of the Praefect processes have restarted and
      the old processes have exited. This can take up to one minute.

      1. Uncomment the secondary Gitaly node configuration commented out in the earlier step on all Praefect nodes.

      1. Run `gitlab-ctl reconfigure` on all Praefect nodes to reconfigure and restart the Praefect processes.

### Deprecated election strategies

WARNING:
The below election strategies are deprecated and were removed in GitLab 14.0.
Migrate to [repository-specific primary nodes](#repository-specific-primary-nodes).

- **PostgreSQL:** Enabled by default until GitLab 14.0, and equivalent to:
  `praefect['failover_election_strategy'] = 'sql'`.

  This configuration option:

  - Allows multiple Praefect nodes to coordinate via the PostgreSQL database to elect a primary
    Gitaly node.
  - Causes Praefect nodes to elect a new primary Gitaly node, monitor its health, and elect a new primary
    Gitaly node if the current one is not reached within 10 seconds by a majority of the Praefect
    nodes.
- **Memory:** Enabled by setting `praefect['failover_election_strategy'] = 'local'`
  in `/etc/gitlab/gitlab.rb` on the Praefect node.

  If a sufficient number of health checks fail for the current primary Gitaly node, a new primary is
  elected. **Do not use with multiple Praefect nodes!** Using with multiple Praefect nodes is
  likely to result in a split brain.