| Supporting services such as NGINX, Prometheus, etc. | 2 | 2 vCPU, 7.5 GB memory | `n1-standard-2` | 3.9 vCPU, 11.8 GB memory |
<!-- Disable ordered list rule https://github.com/DavidAnson/markdownlint/blob/main/doc/Rules.md#md029---ordered-list-item-prefix -->
<!-- markdownlint-disable MD029 -->
1. Nodes configuration is shown as it is forced to ensure pod vcpu / memory ratios and avoid scaling during **performance testing**.
In production deployments there is no need to assign pods to nodes. A minimum of three nodes in three different availability zones is strongly recommended to align with resilient cloud architecture practices.
<!-- markdownlint-enable MD029 -->
Next are the backend components that run on static compute VMs via Omnibus (or External PaaS
<!-- Disable ordered list rule https://github.com/DavidAnson/markdownlint/blob/main/doc/Rules.md#md029---ordered-list-item-prefix -->
<!-- markdownlint-disable MD029 -->
1. Can be optionally run on reputable third-party external PaaS PostgreSQL solutions. Google Cloud SQL and AWS RDS are known to work, however Azure Database for PostgreSQL is [not recommended](https://gitlab.com/gitlab-org/quality/reference-architectures/-/issues/61) due to performance issues. Consul is primarily used for PostgreSQL high availability so can be ignored when using a PostgreSQL PaaS setup. However it is also used optionally by Prometheus for Omnibus auto host discovery.
2. Can be optionally run on reputable third-party external PaaS Redis solutions. Google Memorystore and AWS Elasticache are known to work.
3. Can be optionally run on reputable third-party load balancing services (LB PaaS). AWS ELB is known to work.
4. Should be run on reputable third party object storage (storage PaaS) for cloud implementations. Google Cloud Storage and AWS S3 are known to work.
<!-- markdownlint-enable MD029 -->
NOTE:
For all PaaS solutions that involve configuring instances, it is strongly recommended to implement a minimum of three nodes in three different availability zones to align with resilient cloud architecture practices.
```plantuml
@startuml 5k
card "Kubernetes via Helm Charts" as kubernetes {
card "**External Load Balancer**" as elb #6a9be7
together {
collections "**Webservice** x5" as gitlab #32CD32
collections "**Sidekiq** x3" as sidekiq #ff8dd1
}
card "**Prometheus + Grafana**" as monitor #7FFFD4
card "**Supporting Services**" as support
}
card "**Internal Load Balancer**" as ilb #9370DB
node "**Consul + Sentinel** x3" as consul_sentinel {
component Consul as consul #e76a9b
component Sentinel as sentinel #e6e727
}
card "Gitaly Cluster" as gitaly_cluster {
collections "**Praefect** x3" as praefect #FF8C00
collections "**Gitaly** x3" as gitaly #FF8C00
card "**Praefect PostgreSQL***\n//Non fault-tolerant//" as praefect_postgres #FF8C00
praefect -[#FF8C00]-> gitaly
praefect -[#FF8C00]> praefect_postgres
}
card "Database" as database {
collections "**PGBouncer** x3" as pgbouncer #4EA7FF
card "**PostgreSQL** (Primary)" as postgres_primary #4EA7FF
collections "**PostgreSQL** (Secondary) x2" as postgres_secondary #4EA7FF
pgbouncer -[#4EA7FF]-> postgres_primary
postgres_primary .[#4EA7FF]> postgres_secondary
}
card "Redis" as redis {
collections "**Redis** x3" as redis_nodes #FF6347
redis_nodes <.[#FF6347]- sentinel
}
cloud "**Object Storage**" as object_storage #white
elb -[#6a9be7]-> gitlab
elb -[#6a9be7]-> monitor
elb -[hidden]-> support
gitlab -[#32CD32]> sidekiq
gitlab -[#32CD32]--> ilb
gitlab -[#32CD32]-> object_storage
gitlab -[#32CD32]---> redis
gitlab -[hidden]--> consul
sidekiq -[#ff8dd1]--> ilb
sidekiq -[#ff8dd1]-> object_storage
sidekiq -[#ff8dd1]---> redis
sidekiq -[hidden]--> consul
ilb -[#9370DB]-> gitaly_cluster
ilb -[#9370DB]-> database
consul .[#e76a9b]-> database
consul .[#e76a9b]-> gitaly_cluster
consul .[#e76a9b,norank]--> redis
monitor .[#7FFFD4]> consul
monitor .[#7FFFD4]-> database
monitor .[#7FFFD4]-> gitaly_cluster
monitor .[#7FFFD4,norank]--> redis
monitor .[#7FFFD4]> ilb
monitor .[#7FFFD4,norank]u--> elb
@enduml
```
### Resource usage settings
The following formulas help when calculating how many pods may be deployed within resource constraints.
The [5k reference architecture example values file](https://gitlab.com/gitlab-org/charts/gitlab/-/blob/master/examples/ref/5k.yaml)
documents how to apply the calculated configuration to the Helm Chart.
#### Webservice
Webservice pods typically need about 1 vCPU and 1.25 GB of memory _per worker_.
Each Webservice pod will consume roughly 4 vCPUs and 5 GB of memory using
the [recommended topology](#cluster-topology) because four worker processes
are created by default and each pod has other small processes running.
For 5k users we recommend a total Puma worker count of around 40.
With the [provided recommendations](#cluster-topology) this allows the deployment of up to 10
Webservice pods with 4 workers per pod and 2 pods per node. Expand available resources using
the ratio of 1 vCPU to 1.25 GB of memory _per each worker process_ for each additional
Webservice pod.
For further information on resource usage, see the [Webservice resources](https://docs.gitlab.com/charts/charts/gitlab/webservice/#resources).
#### Sidekiq
Sidekiq pods should generally have 1 vCPU and 2 GB of memory.
[The provided starting point](#cluster-topology) allows the deployment of up to
8 Sidekiq pods. Expand available resources using the 1 vCPU to 2GB memory
ratio for each additional pod.
For further information on resource usage, see the [Sidekiq resources](https://docs.gitlab.com/charts/charts/gitlab/sidekiq/#resources).
