Scale Horizontally
Horizontal scaling means adding more machines or workers into your pool of resources. Horizontally scaling |RCE| gives a huge performance increase, especially under large traffic scenarios with a high number of requests. This is very beneficial when |RCE| is serving many users simultaneously, or if continuous integration servers are automatically pulling and pushing code.
To horizontally scale |RCE| you should use the following steps:
- In the :file:`/home/{user}/.rccontrol/{instance-id}/rhodecode.ini` file, set instance_id = *. This enables |RCE| to use multiple nodes.
- Define the number of worker threads using the formula (2*Cores) + 1. For example 4 CPU cores would lead to (2*4) + 1 = 9 workers. In some cases it's ok to increase number of workers even beyond this formula. Generally the more workers, the more simultaneous connections the system can handle.
It is recommended to create another dedicated |RCE| instance to handle traffic from build farms or continuous integration servers.
Note
You should configure your load balancing accordingly. We recommend writing load balancing rules that will separate regular user traffic from automated process traffic like continuous servers or build bots.
If you scale across different machines, each |RCE| instance needs to store its data on a shared disk, preferably together with your repositories. This data directory contains template caches, a whoosh index, and is used for task locking to ensure safety across multiple instances. To do this, set the following properties in the :file:`/home/{user}/.rccontrol/{instance-id}/rhodecode.ini` file to set the shared location across all |RCE| instances.
cache_dir = /file/path # set to shared directory location search.location = /file/path # set to shared directory location beaker.cache.data_dir = /file/path # set to shared directory location beaker.cache.lock_dir = /file/path # set to shared directory location
Note
If Celery is used on each instance then you should run separate Celery instances, but the message broker should be the same for all of them. This excludes one RabbitMQ shared server.