Running a self-hosted GitLab Omnibus instance provides engineering teams with full governance over their codebase, automated CI/CD pipelines, and project management tools. However, GitLab is an enterprise-grade suite composed of multiple microservices, including the Puma application server, Sidekiq background workers, Gitaly storage daemon, Redis cache, and a PostgreSQL database. Sizing physical and virtual server hardware accurately ensures that concurrent code pushes, automated pipeline triggers, and web interface navigation execute without resource contention.
Official Hardware Baselines: CPU, Memory, and Storage
According to official GitLab self-managed installation documentation, hardware baselines scale according to active registered users and continuous integration workload intensity:
- CPU Sizing: GitLab officially recommends a baseline of 8 vCPU cores for a single-node deployment supporting up to 1,000 users or approximately 20 requests per second. While smaller development teams with under 50 users can run on a 4-core instance during evaluation, production environments with active webhook integrations and background jobs require at least 8 vCPUs. Official documentation specifically advises against using burstable CPU instances (such as shared cloud instances with CPU credit models), as CPU throttling during peak compilation or pipeline triggers leads to severe service timeouts.
- Memory (RAM) Requirements: The recommended production baseline is 16 GB of RAM. In resource-constrained environments or small teams with minimal pipeline volume, GitLab can operate on a minimum of 8 GB of RAM. Running below 8 GB causes the Linux Out-Of-Memory (OOM) killer to terminate Puma worker processes or PostgreSQL connections, causing intermittent 502 Bad Gateway errors.
- Swap Memory Guidance: Official GitLab documentation advises against relying on active swap space. If an instance actively swaps memory pages to disk under load, overall system performance degrades exponentially. Provision sufficient physical RAM so that system services and memory caches operate purely in memory.
- Storage Architecture: GitLab heavily stresses disk I/O due to concurrent git repository packfile writes, PostgreSQL transaction logging, and build artifact storage. Enterprise RAID NVMe storage is strongly recommended over standard mechanical or network-attached SATA disks.
Component Resource Distribution Inside GitLab Omnibus
Understanding which internal components consume system resources enables precise performance tuning inside /etc/gitlab/gitlab.rb:
- Puma Web Server: Puma handles web interface navigation and API calls. By default, Omnibus GitLab tunes Puma worker count based on available CPU cores. Each Puma worker process consumes roughly 500 MB to 1 GB of memory. On an 8-core instance, configure 4 to 6 workers to preserve memory for other daemons.
- Sidekiq Background Processing: Sidekiq processes asynchronous queues, such as email dispatch, repository integrity checks, webhook delivery, and pipeline creation. Under heavy push volume, Sidekiq memory usage scales with concurrent job execution.
- Gitaly: Gitaly is the git RPC service that executes git commands directly against disk repositories. High-concurrency operations, such as multi-user pulls or large automated build checkouts, require high disk IOPS to avoid I/O wait spikes.
- PostgreSQL & Redis: GitLab includes an embedded PostgreSQL engine that requires dedicated shared memory buffers, alongside Redis for fast queue management and session storage.
Isolating GitLab CI/CD Runners: The Dedicated Host Rule
A frequent infrastructure mistake is deploying GitLab CI/CD Runners on the exact same virtual server hosting the primary GitLab Omnibus instance. When developers trigger intensive build pipelines, such as compiling Go binaries, executing Docker-in-Docker builds, or running browser test suites, the runner consumes 100% of available CPU and memory. This starves the GitLab Puma server and Gitaly daemon, causing connection dropouts and rendering the web interface unresponsive.
To maintain infrastructure stability, deploy GitLab Runners on separate, dedicated KVM VPS instances. Use the GitLab Runner registration token to link remote runner hosts back to your primary instance via HTTPS, ensuring that compute spikes during automated builds remain isolated from your code repository.
GitLab Infrastructure Sizing Matrix
| Team Size / Workload | Recommended Configuration | Storage Specification | Architecture Strategy |
|---|---|---|---|
| Small Team (Up to 25 Users) | 4 vCPU, 8 GB RAM (KVM VPS) | 80 GB RAID NVMe | Single-node Omnibus with external runner node. |
| Growing Team (25 to 100 Users) | 8 vCPU, 16 GB RAM (KVM VPS) | 160 GB RAID NVMe | Dedicated GitLab server + 2x isolated CI/CD runner nodes. |
| Enterprise (100 to 1,000 Users) | AMD Ryzen 9 9950X / Dedicated, 32 GB+ DDR5 | 500 GB+ Enterprise NVMe | High-frequency compute instance for fast Gitaly and Puma execution. |
StreamData Networks provides high-performance KVM Cloud VPS instances and high-frequency AMD Ryzen 9 9950X servers with RAID NVMe storage pools and 1 Gbps or 10 Gbps port options, engineered for developer workloads and self-hosted DevOps stacks. Review available configurations on KVM VPS, configure bare-metal hosting on Dedicated Servers, or learn about infrastructure capabilities via StreamData Company Overview.