Cluster network configuration

Network considerations

The simplest network design for clustering is to place the nodes on a single LAN. If a cluster must span subnetworks, we do not recommend configuring any firewall rules between the networks. The latency between nodes should be less than 1 millisecond.

For high availability, the latency between the network with the active nodes and the network with the passive nodes must be less than 70 milliseconds. We don't recommend configuring a firewall between the two networks.

Application ports for end users

Application ports provide web application and Git access for end users.

Port	Description	Encrypted
22/TCP	Git over SSH	Yes
25/TCP	SMTP	Requires STARTTLS
80/TCP	HTTP	No (When SSL is enabled this port redirects to HTTPS)
443/TCP	HTTPS	Yes
9418/TCP	Simple Git protocol port (Disabled in private mode)	No

Administrative ports

Administrative ports are not required for basic application use by end users.

Port	Description	Encrypted
ICMP	ICMP Ping	No
122/TCP	Administrative SSH	Yes
161/UDP	SNMP	No
8080/TCP	Management Console HTTP	No (When SSL is enabled this port redirects to HTTPS)
8443/TCP	Management Console HTTPS	Yes

Cluster communication ports

If a network level firewall is in place between nodes, these ports will need to be accessible. The communication between nodes is not encrypted. These ports should not be accessible externally.

Port	Description
1336/TCP	Internal API
3033/TCP	Internal SVN access
3037/TCP	Internal SVN access
3306/TCP	MySQL
4486/TCP	Governor access
5115/TCP	Storage backend
5208/TCP	Internal SVN access
6379/TCP	Redis
8001/TCP	Grafana
8090/TCP	Internal GPG access
8149/TCP	GitRPC file server access
8300/TCP	Consul
8301/TCP	Consul
8302/TCP	Consul
9000/TCP	Git Daemon
9102/TCP	Pages file server
9105/TCP	LFS server
9200/TCP	Elasticsearch
9203/TCP	Semantic code service
9300/TCP	Elasticsearch
11211/TCP	Memcache
161/UDP	SNMP
8125/UDP	Statsd
8301/UDP	Consul
8302/UDP	Consul
25827/UDP	Collectd

Configuring a load balancer

We recommend an external TCP-based load balancer that supports the PROXY protocol to distribute traffic across nodes. Consider these load balancer configurations:

• TCP ports (shown below) should be forwarded to nodes running the web-server service. These are the only nodes that serve external client requests.
• Sticky sessions shouldn't be enabled.

Handling client connection information

Because client connections to the cluster come from the load balancer, the client IP address can be lost. To properly capture the client connection information, additional consideration is required.

If your load balancer can support it, we strongly recommend implementing the PROXY protocol. When no PROXY support is available, it is also possible to load balance the HTTP and HTTPS ports using the X-Forwarded-For header.

Enabling PROXY support on GitHub Enterprise Server

We strongly recommend enabling PROXY support for both your instance and the load balancer.

• For your instance, use this command:

  $ ghe-config 'loadbalancer.proxy-protocol' 'true' && ghe-cluster-config-apply

• For the load balancer, use the instructions provided by your vendor.

  PROXY protocol TCP port mappings

Source port	Destination port	Service description
22	23	Git over SSH
80	81	HTTP
443	444	HTTPS
8080	8081	Management Console HTTP
8443	8444	Management Console HTTPS
9418	9419	Git

Enabling X-Forwarded-For support on GitHub Enterprise Server

Use the X-Forwarded-For protocol only when the PROXY protocol is unavailable. The X-Forwarded-For header only works with HTTP and HTTPS. The IP address reported for Git connections over SSH will show the load balancer IP.

To enable the X-Forwarded-For header, use this command:

$ ghe-config 'loadbalancer.http-forward' 'true' && ghe-cluster-config-apply

Protocol TCP port mappings for use without PROXY support

Source port	Destination port	Service description
22	22	Git over SSH
25	25	SMTP
80	80	HTTP
443	443	HTTPS
8080	8080	Management Console HTTP
8443	8443	Management Console HTTPS

Configuring Health Checks

Health checks allow a load balancer to stop sending traffic to a node that is not responding if a pre-configured check fails on that node. If a cluster node fails, health checks paired with redundant nodes provides high availability.

Configure the load balancer to check one of these URLs:

• https://HOSTNAME/status if HTTPS is enabled (default)
• http://HOSTNAME/status if HTTPS is disabled

The check will return status code 200 (OK) if the node is healthy and available to service end-user requests.

DNS Requirements

DNS lookups for the GitHub Enterprise Server hostname should resolve to the load balancer. We recommend that you enable subdomain isolation. If subdomain isolation is enabled, an additional wildcard record (*.HOSTNAME) should also resolve to the load balancer. For more information, see "Enabling subdomain isolation."