GCS Fault Tolerance
GCS Fault Tolerance#
Global Control Service (GCS) is a server that manages cluster-level metadata. It also provides a handful of cluster-level operations including actor, placement groups and node management. By default, the GCS is not fault tolerant since all the data is stored in-memory and its failure means that the entire Ray cluster fails. To make the GCS fault tolerant, HA Redis is required. Then, when the GCS restarts, it loads all the data from the Redis instance and resumes regular functions.
During the recovery period, the following functions are not available:
Actor creation, deletion and reconstruction.
Placement group creation, deletion and reconstruction.
Worker node registration.
Worker process creation.
However, running Ray tasks and actors remain alive and any existing objects will continue to be available.
Setting up Redis#
If you are using ray start to start the Ray head node,
set the OS environment
the Redis address, and supply the
--redis-password flag with the password when calling
RAY_REDIS_ADDRESS=redis_ip:port ray start --head --redis-password PASSWORD
head_start_ray_commands: - ray stop - ulimit -n 65536; RAY_REDIS_ADDRESS=redis_ip:port ray start --head --redis-password PASSWORD --port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --dashboard-host=0.0.0.0
Once the GCS is backed by Redis, when it restarts, it’ll recover the
state by reading from Redis. When the GCS is recovering from its failed state, the raylet
will try to reconnect to the GCS.
If the raylet fails to reconnect to the GCS for more than 60 seconds,
the raylet will exit and the corresponding node fails.
This timeout threshold can be tuned by the OS environment variable
If the IP address of GCS will change after restarts, it’s better to use a qualified domain name and pass it to all raylets at start time. Raylet will resolve the domain name and connect to the correct GCS. You need to ensure that at any time, only one GCS is alive.