Reference architecture: up to 5,000 users

This page describes GitLab reference architecture for up to 5,000 users. For a full list of reference architectures, see Available reference architectures.

  • Supported users (approximate): 5,000
  • High Availability: True
  • Test RPS rates: API: 100 RPS, Web: 10 RPS, Git: 10 RPS
Service Nodes Configuration (8) GCP AWS (9) Azure(9)
GitLab Rails (1) 3 16 vCPU, 14.4GB Memory n1-highcpu-16 c5.4xlarge F16s v2
PostgreSQL 3 2 vCPU, 7.5GB Memory n1-standard-2 m5.large D2s v3
PgBouncer 3 2 vCPU, 1.8GB Memory n1-highcpu-2 c5.large F2s v2
Gitaly (2) (5) (7) X 8 vCPU, 30GB Memory n1-standard-8 m5.2xlarge D8s v3
Redis (3) 3 2 vCPU, 7.5GB Memory n1-standard-2 m5.large D2s v3
Consul + Sentinel (3) 3 2 vCPU, 1.8GB Memory n1-highcpu-2 c5.large F2s v2
Sidekiq 4 2 vCPU, 7.5GB Memory n1-standard-2 m5.large D2s v3
Object Storage (4) - - - - -
NFS Server (5) (7) 1 4 vCPU, 3.6GB Memory n1-highcpu-4 c5.xlarge F4s v2
Monitoring node 1 2 vCPU, 1.8GB Memory n1-highcpu-2 c5.large F2s v2
External load balancing node (6) 1 2 vCPU, 1.8GB Memory n1-highcpu-2 c5.large F2s v2
Internal load balancing node (6) 1 2 vCPU, 1.8GB Memory n1-highcpu-2 c5.large F2s v2

Footnotes

  1. In our architectures we run each GitLab Rails node using the Puma webserver and have its number of workers set to 90% of available CPUs along with four threads. For nodes that are running Rails with other components the worker value should be reduced accordingly where we've found 50% achieves a good balance but this is dependent on workload.

  2. Gitaly node requirements are dependent on customer data, specifically the number of projects and their sizes. We recommend two nodes as an absolute minimum for HA environments and at least four nodes should be used when supporting 50,000 or more users. We also recommend that each Gitaly node should store no more than 5TB of data and have the number of gitaly-ruby workers set to 20% of available CPUs. Additional nodes should be considered in conjunction with a review of expected data size and spread based on the recommendations above.

  3. Recommended Redis setup differs depending on the size of the architecture. For smaller architectures (less than 3,000 users) a single instance should suffice. For medium sized installs (3,000 - 5,000) we suggest one Redis cluster for all classes and that Redis Sentinel is hosted alongside Consul. For larger architectures (10,000 users or more) we suggest running a separate Redis Cluster for the Cache class and another for the Queues and Shared State classes respectively. We also recommend that you run the Redis Sentinel clusters separately for each Redis Cluster.

  4. For data objects such as LFS, Uploads, Artifacts, etc. We recommend an Object Storage service over NFS where possible, due to better performance and availability.

  5. NFS can be used as an alternative for both repository data (replacing Gitaly) and object storage but this isn't typically recommended for performance reasons. Note however it is required for GitLab Pages.

  6. Our architectures have been tested and validated with HAProxy as the load balancer. Although other load balancers with similar feature sets could also be used, those load balancers have not been validated.

  7. We strongly recommend that any Gitaly or NFS nodes be set up with SSD disks over HDD with a throughput of at least 8,000 IOPS for read operations and 2,000 IOPS for write as these components have heavy I/O. These IOPS values are recommended only as a starter as with time they may be adjusted higher or lower depending on the scale of your environment's workload. If you're running the environment on a Cloud provider you may need to refer to their documentation on how configure IOPS correctly.

  8. The architectures were built and tested with the Intel Xeon E5 v3 (Haswell) CPU platform on GCP. On different hardware you may find that adjustments, either lower or higher, are required for your CPU or Node counts accordingly. For more information, a Sysbench benchmark of the CPU can be found here.

  9. AWS-equivalent and Azure-equivalent configurations are rough suggestions and may change in the future. They have not yet been tested and validated.