Tune PostgreSQL

You should tune PostgreSQL when:

• Other GitLab components are reconfigured or scaled up in a way that affects the database.
• The performance of your GitLab environment is impaired.
• GitLab uses an external PostgreSQL service.

Required settings for external instances

The following settings are required for externally managed PostgreSQL instances.

You can configure some PostgreSQL settings for the specific database, rather than for all databases on the server.

• You might limit configuration to specific databases when hosting multiple databases on the same server.
• For guidance on where to apply configuration, consult your database administrator or vendor.
• For GCP Cloud SQL, you can set statement_timeout on a specific database or user, but not as a database flag. For example: ALTER DATABASE gitlab SET statement_timeout = '60s';

Plan your database connections

GitLab uses database connections from multiple components. Proper connection planning prevents database connection exhaustion and performance issues.

Each GitLab component uses database connections based on its configuration. Sidekiq and Puma establish a pool of connections to PostgreSQL at initialization. The number of connections in the pool can increase later if there are connection spikes or temporary increases in demand:

• Configure database pool headroom with the environment variable DB_POOL_HEADROOM.
• When you tune PostgreSQL, plan for pool headroom but do not change it. GitLab deployments respond better to higher demand if more capacity is available: deploy more Sidekiq or Puma workers.

Puma

Puma connections = puma['worker_processes'] × (puma['max_threads'] + DB_POOL_HEADROOM)

By default:

• puma['worker_processes'] is based on CPU core count.
• puma['max_threads'] is 4.
• DB_POOL_HEADROOM is 10.

Per-worker calculation: Each Puma worker uses 4 threads + 10 headroom, for a total of 14 connections.

Default calculation, assuming 8 vCPU: 8 workers × 14 connections per worker, for a total of 112 Puma connections.

Sidekiq

Sidekiq connections = Number of Sidekiq processes × (sidekiq['concurrency'] + 1 + DB_POOL_HEADROOM)

By default:

• The number of Sidekiq processes is 1.
• sidekiq['concurrency'] is 20.
• DB_POOL_HEADROOM is 10.

Default calculation: 1 Sidekiq process × (20 concurrency + 1 + 10 headroom), for a total of 31 total Sidekiq connections.

Geo Log Cursor (Geo installations only)

The Geo Log Cursor daemon runs on all GitLab Rails nodes in a secondary site.

Geo log cursor connections = 1 + DB_POOL_HEADROOM

Default calculation: 1 + 10 headroom, for a total of 11 Geo connections.

Total connection requirements

For single node installations:

Total connections = 2 × (Puma + Sidekiq + Geo)

For multi-node installations, multiply by the number of nodes running each component:

Total connections = 2 × ((Puma × Rails nodes) + (Sidekiq × Sidekiq nodes) + (Geo × secondary Rails nodes))

Multiplying by 2 accounts for the dual database connections in GitLab 16.0 and later.

For Geo installations:

• Primary site: Use Geo = 0. Geo Log Cursor doesn’t run on primary sites.
• Secondary sites: Calculate the Geo Log Cursor database connections for one secondary site, and apply that same calculation to all secondary sites.
• Each Geo site connects to its own database, so you don’t need to sum connections across multiple Geo sites.
• Set max_connections to the same value on both the primary PostgreSQL database and all replica databases, using the highest connection requirement across all Geo sites.

Examples

Single node installation

This example is based on the GitLab reference architecture for 20 RPS (requests per second) or 1000 users:

Multi-node installation

This example is based on the GitLab reference architecture for 40 RPS (requests per second) or 2000 users:

Single node installation with Geo

This example is based on the GitLab reference architecture for 20 RPS (requests per second) or 1000 users.

Multi-node installation with Geo

This example is based on the GitLab reference architecture for 40 RPS (requests per second) or 2000 users: