debian-mirror-gitlab/doc/administration/high_availability

type
reference, concepts

Scaling and High Availability

GitLab supports several different types of clustering and high-availability. The solution you choose will be based on the level of scalability and availability you require. The easiest solutions are scalable, but not necessarily highly available.

GitLab provides a service that is usually essential to most organizations: it enables people to collaborate on code in a timely fashion. Any downtime should therefore be short and planned. Luckily, GitLab provides a solid setup even on a single server without special measures. Due to the distributed nature of Git, developers can still commit code locally even when GitLab is not available. However, some GitLab features such as the issue tracker and Continuous Integration are not available when GitLab is down.

Keep in mind that all highly-available solutions come with a trade-off between cost/complexity and uptime. The more uptime you want, the more complex the solution. And the more complex the solution, the more work is involved in setting up and maintaining it. High availability is not free and every HA solution should balance the costs against the benefits.

There are many options when choosing a highly-available GitLab architecture. We recommend engaging with GitLab Support to choose the best architecture for your use case. This page contains some various options and guidelines based on experience with GitLab.com and Enterprise Edition on-premises customers.

For detailed insight into how GitLab scales and configures GitLab.com, you can watch this 1 hour Q&A with John Northrup, and live questions coming in from some of our customers.

GitLab Components

The following components need to be considered for a scaled or highly-available environment. In many cases, components can be combined on the same nodes to reduce complexity.

• GitLab application nodes (Unicorn / Puma, Workhorse) - Web-requests (UI, API, Git over HTTP)
• Sidekiq - Asynchronous/Background jobs
• PostgreSQL - Database
  • Consul - Database service discovery and health checks/failover
  • PgBouncer - Database pool manager
• Redis - Key/Value store (User sessions, cache, queue for Sidekiq)
  • Sentinel - Redis health check/failover manager
• Gitaly - Provides high-level storage and RPC access to Git repositories
• S3 Object Storage service¹ and / or NFS storage servers² for entities such as Uploads, Artifacts, LFS Objects, etc...
• Load Balancer³ - Main entry point and handles load balancing for the GitLab application nodes.
• Monitor - Prometheus and Grafana monitoring with auto discovery.

Scalable Architecture Examples

When an organization reaches a certain threshold it will be necessary to scale the GitLab instance. Still, true high availability may not be necessary. There are options for scaling GitLab instances relatively easily without incurring the infrastructure and maintenance costs of full high availability.

Basic Scaling

This is the simplest form of scaling and will work for the majority of cases. Backend components such as PostgreSQL, Redis, and storage are offloaded to their own nodes while the remaining GitLab components all run on 2 or more application nodes.

This form of scaling also works well in a cloud environment when it is more cost effective to deploy several small nodes rather than a single larger one.

• 1 PostgreSQL node
• 1 Redis node
• 1 Gitaly node
• 1 or more Object Storage services¹ and / or NFS storage server²
• 2 or more GitLab application nodes (Unicorn / Puma, Workhorse, Sidekiq)
• 1 or more Load Balancer nodes³
• 1 Monitoring node (Prometheus, Grafana)

Installation Instructions

Complete the following installation steps in order. A link at the end of each section will bring you back to the Scalable Architecture Examples section so you can continue with the next step.

1. PostgreSQL with PGBouncer
2. Redis
3. Gitaly (recommended) and / or NFS²
4. GitLab application nodes
   • With Object Storage service enabled¹
5. Load Balancer(s)³
6. Monitoring node (Prometheus and Grafana)

Full Scaling

For very large installations, it might be necessary to further split components for maximum scalability. In a fully-scaled architecture, the application node is split into separate Sidekiq and Unicorn/Workhorse nodes. One indication that this architecture is required is if Sidekiq queues begin to periodically increase in size, indicating that there is contention or there are not enough resources.

• 1 or more PostgreSQL nodes
• 1 or more Redis nodes
• 1 or more Gitaly storage servers
• 1 or more Object Storage services¹ and / or NFS storage server²
• 2 or more Sidekiq nodes
• 2 or more GitLab application nodes (Unicorn / Puma, Workhorse, Sidekiq)
• 1 or more Load Balancer nodes³
• 1 Monitoring node (Prometheus, Grafana)

High Availability Architecture Examples

When organizations require scaling and high availability, the following architectures can be utilized. As the introduction section at the top of this page mentions, there is a tradeoff between cost/complexity and uptime. Be sure this complexity is absolutely required before taking the step into full high availability.

For all examples below, we recommend running Consul and Redis Sentinel on dedicated nodes. If Consul is running on PostgreSQL nodes or Sentinel on Redis nodes, there is a potential that high resource usage by PostgreSQL or Redis could prevent communication between the other Consul and Sentinel nodes. This may lead to the other nodes believing a failure has occurred and initiating automated failover. Isolating Redis and Consul from the services they monitor reduces the chances of a false positive that a failure has occurred.

The examples below do not address high availability of NFS for objects. We recommend a S3 Object Storage service¹ is used where possible over NFS but it's still required in certain cases². Where NFS is to be used some enterprises have access to NFS appliances that manage availability and this would be best case scenario.

There are many options in between each of these examples. Work with GitLab Support to understand the best starting point for your workload and adapt from there.

Horizontal

This is the simplest form of high availability and scaling. It requires the fewest number of individual servers (virtual or physical) but does have some trade-offs and limits.

This architecture will work well for many GitLab customers. Larger customers may begin to notice certain events cause contention/high load - for example, cloning many large repositories with binary files, high API usage, a large number of enqueued Sidekiq jobs, and so on. If this happens, you should consider moving to a hybrid or fully distributed architecture depending on what is causing the contention.

• 3 PostgreSQL nodes
• 2 Redis nodes
• 3 Consul/Sentinel nodes
• 2 or more GitLab application nodes (Unicorn / Puma, Workhorse, Sidekiq)
• 1 Gitaly storage servers
• 1 Object Storage service¹ and / or NFS storage server²
• 1 or more Load Balancer nodes³
• 1 Monitoring node (Prometheus, Grafana)

Hybrid

In this architecture, certain components are split on dedicated nodes so high resource usage of one component does not interfere with others. In larger environments this is a good architecture to consider if you foresee or do have contention due to certain workloads.

• 3 PostgreSQL nodes
• 1 PgBouncer node
• 2 Redis nodes
• 3 Consul/Sentinel nodes
• 2 or more Sidekiq nodes
• 2 or more GitLab application nodes (Unicorn / Puma, Workhorse, Sidekiq)
• 1 Gitaly storage servers
• 1 Object Storage service¹ and / or NFS storage server²
• 1 or more Load Balancer nodes³
• 1 Monitoring node (Prometheus, Grafana)

Fully Distributed

This architecture scales to hundreds of thousands of users and projects and is the basis of the GitLab.com architecture. While this scales well it also comes with the added complexity of many more nodes to configure, manage, and monitor.

• 3 PostgreSQL nodes
• 1 or more PgBouncer nodes (with associated internal load balancers)
• 4 or more Redis nodes (2 separate clusters for persistent and cache data)
• 3 Consul nodes
• 3 Sentinel nodes
• Multiple dedicated Sidekiq nodes (Split into real-time, best effort, ASAP, CI Pipeline and Pull Mirror sets)
• 2 or more Git nodes (Git over SSH/Git over HTTP)
• 2 or more API nodes (All requests to /api)
• 2 or more Web nodes (All other web requests)
• 2 or more Gitaly storage servers
• 1 or more Object Storage services¹ and / or NFS storage servers²
• 1 or more Load Balancer nodes³
• 1 Monitoring node (Prometheus, Grafana)

Reference Architecture Examples

The Support and Quality teams build, performance test, and validate Reference Architectures that support set large numbers of users. The specifications below are a representation of this work so far and may be adjusted in the future based on additional testing and iteration.

The architectures have been tested with specific coded workloads. The throughputs used for testing are calculated based on sample customer data. We test each endpoint type with the following number of requests per second (RPS) per 1000 users:

• API: 20 RPS
• Web: 2 RPS
• Git: 2 RPS

Note that your exact needs may be more, depending on your workload. Your workload is influenced by factors such as - but not limited to - how active your users are, how much automation you use, mirroring, and repo/change size.

10,000 User Configuration

• Supported Users (approximate): 10,000
• Test RPS Rates: API: 200 RPS, Web: 20 RPS, Git: 20 RPS
• Known Issues: While validating the reference architecture, slow API endpoints were discovered. For details, see the related issues list in this issue.

Service	Nodes	Configuration	GCP type
GitLab Rails - Puma workers on each node set to 90% of available CPUs with 16 threads	3	32 vCPU, 28.8GB Memory	n1-highcpu-32
PostgreSQL	3	4 vCPU, 15GB Memory	n1-standard-4
PgBouncer	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
Gitaly - Gitaly Ruby workers on each node set to 20% of available CPUs	X4 .	16 vCPU, 60GB Memory	n1-standard-16
Redis Cache + Sentinel - Cache maxmemory set to 90% of available memory	3	4 vCPU, 15GB Memory	n1-standard-4
Redis Persistent + Sentinel	3	4 vCPU, 15GB Memory	n1-standard-4
Sidekiq	4	4 vCPU, 15GB Memory	n1-standard-4
Consul	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
NFS Server2 .	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
S3 Object Storage1 .	-	-	-
Monitoring node	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
External load balancing node3 .	1	2 vCPU, 1.8GB Memory	n1-highcpu-2
Internal load balancing node3 .	1	2 vCPU, 1.8GB Memory	n1-highcpu-2

NOTE: Note: Memory values are given directly by GCP machine sizes. On different cloud vendors a best effort like for like can be used.

25,000 User Configuration

• Supported Users (approximate): 25,000
• Test RPS Rates: API: 500 RPS, Web: 50 RPS, Git: 50 RPS
• Known Issues: The slow API endpoints that were discovered during testing the 10,000 user architecture also affect the 25,000 user architecture. For details, see the related issues list in this issue.

Service	Nodes	Configuration	GCP type
GitLab Rails - Puma workers on each node set to 90% of available CPUs with 16 threads	7	32 vCPU, 28.8GB Memory	n1-highcpu-32
PostgreSQL	3	8 vCPU, 30GB Memory	n1-standard-8
PgBouncer	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
Gitaly - Gitaly Ruby workers on each node set to 20% of available CPUs	X4 .	32 vCPU, 120GB Memory	n1-standard-32
Redis Cache + Sentinel - Cache maxmemory set to 90% of available memory	3	4 vCPU, 15GB Memory	n1-standard-4
Redis Persistent + Sentinel	3	4 vCPU, 15GB Memory	n1-standard-4
Sidekiq	4	4 vCPU, 15GB Memory	n1-standard-4
Consul	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
NFS Server2 .	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
S3 Object Storage1 .	-	-	-
Monitoring node	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
External load balancing node3 .	1	2 vCPU, 1.8GB Memory	n1-highcpu-2
Internal load balancing node3 .	1	4 vCPU, 3.6GB Memory	n1-highcpu-4

NOTE: Note: Memory values are given directly by GCP machine sizes. On different cloud vendors a best effort like for like can be used.

50,000 User Configuration

• Supported Users (approximate): 50,000
• Test RPS Rates: API: 1000 RPS, Web: 100 RPS, Git: 100 RPS
• Status: Work-in-progress
• Related Issue: See the related issue for more information.

NOTE: Note: This architecture is a work-in-progress of the work so far. The Quality team will be certifying this environment in late 2019. The specifications may be adjusted prior to certification based on performance testing.

Service	Nodes	Configuration	GCP type
GitLab Rails - Puma workers on each node set to 90% of available CPUs with 16 threads	15	32 vCPU, 28.8GB Memory	n1-highcpu-32
PostgreSQL	3	8 vCPU, 30GB Memory	n1-standard-8
PgBouncer	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
Gitaly - Gitaly Ruby workers on each node set to 20% of available CPUs	X4 .	64 vCPU, 240GB Memory	n1-standard-64
Redis Cache + Sentinel - Cache maxmemory set to 90% of available memory	3	4 vCPU, 15GB Memory	n1-standard-4
Redis Persistent + Sentinel	3	4 vCPU, 15GB Memory	n1-standard-4
Sidekiq	4	4 vCPU, 15GB Memory	n1-standard-4
Consul	3	2 vCPU, 1.8GB Memory	n1-highcpu-2
NFS Server2 .	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
S3 Object Storage1 .	-	-	-
Monitoring node	1	4 vCPU, 3.6GB Memory	n1-highcpu-4
External load balancing node3 .	1	2 vCPU, 1.8GB Memory	n1-highcpu-2
Internal load balancing node3 .	1	8 vCPU, 7.2GB Memory	n1-highcpu-8

NOTE: Note: Memory values are given directly by GCP machine sizes. On different cloud vendors a best effort like for like can be used.

---

1. For data objects such as LFS, Uploads, Artifacts, etc... We recommend a S3 Object Storage where possible over NFS due to better performance and availability. Several types of objects are supported for S3 storage - Job artifacts, LFS, Uploads, Merge Request Diffs, Packages (Optional Feature), Dependency Proxy (Optional Feature). ↩︎

2. NFS storage server is still required for GitLab Pages and optionally for CI Job Incremental Logging (can be switched to use Redis instead). ↩︎

3. Our architectures have been tested and validated with HAProxy as the load balancer. However other reputable load balancers with similar feature sets should also work here but be aware these aren't validated. ↩︎

4. Gitaly node requirements are dependent on customer data. We recommend 2 nodes as an absolute minimum for performance at the 10,000 and 25,000 user scale and 4 nodes as an absolute minimum at the 50,000 user scale, but additional nodes should be considered in conjunction with a review of project counts and sizes. ↩︎