• Requirements to enable Database Load Balancing
• Configuring Database Load Balancing
  • Hosts
  • Service Discovery
  • Handling stale reads
• Logging
• Implementation Details
  • Balancing queries
  • Prepared statements
  • Primary sticking
  • Failover handling

Database Load Balancing

With Database Load Balancing, read-only queries can be distributed across multiple PostgreSQL nodes to increase performance.

This functionality is provided natively in GitLab Rails and Sidekiq where they can be configured to balance their database read queries in a round-robin approach, without any external dependencies:

Requirements to enable Database Load Balancing

To enable Database Load Balancing, make sure that:

• The HA PostgreSQL setup has one or more secondary nodes replicating the primary.
• Each PostgreSQL node is connected with the same credentials and on the same port.

For Omnibus GitLab, you also need PgBouncer configured on each PostgreSQL node to pool all load-balanced connections when configuring a multi-node setup.

Configuring Database Load Balancing

Database Load Balancing can be configured in one of two ways:

• (Recommended) Hosts: a list of PostgreSQL hosts.
• Service Discovery: a DNS record that returns a list of PostgreSQL hosts.

Hosts

To configure a list of hosts, perform these steps on all GitLab Rails (Sidekiq) nodes for each environment you want to balance:

1. Edit the /etc/gitlab/gitlab.rb file.

2. In gitlab_rails['db_load_balancing'], create an array of the read-only replicas you want to balance. Do not add the primary host. For example, on an environment with PostgreSQL running on the hosts primary.example.com, host1.example.com, host2.example.com, and host3.example.com:

   gitlab_rails['db_load_balancing'] = { 'hosts' => ['host1.example.com', 'host2.example.com', `host3.example.com`] }
   

   These replicas must be reachable on the same port configured with gitlab_rails['db_port'].

3. Save the file and reconfigure GitLab.

Service Discovery

Service discovery allows GitLab to automatically retrieve a list of PostgreSQL hosts to use. It periodically checks a DNS A record, using the IPs returned by this record as the addresses for the secondaries. For service discovery to work, all you need is a DNS server and an A record containing the IP addresses of your secondaries.

When using Omnibus GitLab the provided Consul service works as a DNS server and returns PostgreSQL addresses via the postgresql-ha.service.consul record. For example:

1. On each GitLab Rails / Sidekiq node, edit /etc/gitlab/gitlab.rb and add the following:

  gitlab_rails['db_load_balancing'] = { 'discover' => {
      'nameserver' => 'localhost'
      'record' => 'postgresql-ha.service.consul'
      'record_type' => 'A'
      'port' => '8600'
      'interval' => '60'
      'disconnect_timeout' => '120'
    }
  }

1. Save the file and reconfigure GitLab for the changes to take effect.

If record_type is set to SRV, then GitLab continues to use round-robin algorithm and ignores the weight and priority in the record. Since SRV records usually return hostnames instead of IPs, GitLab needs to look for the IPs of returned hostnames in the additional section of the SRV response. If no IP is found for a hostname, GitLab needs to query the configured nameserver for ANY record for each such hostname looking for A or AAAA records, eventually dropping this hostname from rotation if it can’t resolve its IP.

The interval value specifies the minimum time between checks. If the A record has a TTL greater than this value, then service discovery honors said TTL. For example, if the TTL of the A record is 90 seconds, then service discovery waits at least 90 seconds before checking the A record again.

When the list of hosts is updated, it might take a while for the old connections to be terminated. The disconnect_timeout setting can be used to enforce an upper limit on the time it takes to terminate all old database connections.

Handling stale reads

To prevent reading from an outdated secondary the load balancer checks if it is in sync with the primary. If the data is recent enough, the secondary is used, otherwise it is ignored. To reduce the overhead of these checks we only perform them at certain intervals.

There are three configuration options that influence this behavior:

The defaults should be sufficient for most users.

To configure these options with a hosts list, use the following example:

gitlab_rails['db_load_balancing'] = {
  'hosts' => ['host1.example.com', 'host2.example.com', `host3.example.com`]
  'max_replication_difference' => 16777216 # 16 MB
  'max_replication_lag_time' => 30
  'replica_check_interval' => 30
}

Logging

The load balancer logs various events in database_load_balancing.log, such as

• When a host is marked as offline
• When a host comes back online
• When all secondaries are offline
• When a read is retried on a different host due to a query conflict

The log is structured with each entry a JSON object containing at least:

• An event field useful for filtering.
• A human-readable message field.
• Some event-specific metadata. For example, db_host
• Contextual information that is always logged. For example, severity and time.

For example:

{"severity":"INFO","time":"2019-09-02T12:12:01.728Z","correlation_id":"abcdefg","event":"host_online","message":"Host came back online","db_host":"111.222.333.444","db_port":null,"tag":"rails.database_load_balancing","environment":"production","hostname":"web-example-1","fqdn":"gitlab.example.com","path":null,"params":null}

Implementation Details

Balancing queries

Read-only SELECT queries balance among all the given hosts. Everything else (including transactions) executes on the primary. Queries such as SELECT ... FOR UPDATE are also executed on the primary.

Prepared statements

Prepared statements don’t work well with load balancing and are disabled automatically when load balancing is enabled. This shouldn’t impact response timings.

Primary sticking

After a write has been performed, GitLab sticks to using the primary for a certain period of time, scoped to the user that performed the write. GitLab reverts back to using secondaries when they have either caught up, or after 30 seconds.

Failover handling

In the event of a failover or an unresponsive database, the load balancer tries to use the next available host. If no secondaries are available the operation is performed on the primary instead.

If a connection error occurs while writing data, the operation retries up to 3 times using an exponential back-off.

When using load balancing, you should be able to safely restart a database server without it immediately leading to errors being presented to the users.