GCS fault tolerance#

The Global Control Service, or GCS, manages cluster-level metadata. It also handles cluster-level operations, including actor, placement group, and node management. By default, the GCS isn’t fault tolerant because it stores all data in memory. If it fails, the entire Ray cluster fails. To enable GCS fault tolerance, back the GCS with durable storage so it can reload cluster metadata after a restart. Ray offers two backends:

  • External Redis: The GCS persists its state to a highly available Redis instance, known as HA Redis. This backend is officially supported.

  • Embedded RocksDB: The GCS persists its state to a local RocksDB database on a persistent volume, so you don’t run an external data store. This backend is alpha. See Embedded RocksDB backend (alpha).

With either backend, when the GCS restarts, it reloads all its data from the backing store and resumes its regular functions.

During the recovery period, the following functions aren’t available:

  • Actor creation, deletion, and reconstruction

  • Placement group creation, deletion, and reconstruction

  • Resource management

  • Worker node registration

  • Worker process creation

However, running Ray tasks and actors remain alive, and any existing objects stay available.

Setting up Redis#

If you use KubeRay, see the KubeRay docs on GCS fault tolerance.

If you use ray start to start the Ray head node, set the RAY_REDIS_ADDRESS OS environment variable to the Redis address, and pass the password with the --redis-password flag when you call ray start:

RAY_REDIS_ADDRESS=redis_ip:port ray start --head --redis-password PASSWORD --redis-username default

If you use ray up to start the Ray cluster, add RAY_REDIS_ADDRESS and --redis-password to the ray start command in the head_start_ray_commands field:

head_start_ray_commands:
  - ray stop
  - ulimit -n 65536; RAY_REDIS_ADDRESS=redis_ip:port ray start --head --redis-password PASSWORD --redis-username default --port=6379 --object-manager-port=8076 --autoscaling-config=~/ray_bootstrap_config.yaml --dashboard-host=0.0.0.0

After you back the GCS with Redis, it recovers its state from Redis when it restarts. While the GCS recovers, each raylet tries to reconnect to it. If a raylet can’t reconnect for more than 60 seconds, that raylet exits and the corresponding node fails. Set this timeout with the RAY_gcs_rpc_server_reconnect_timeout_s OS environment variable.

If the GCS IP address might change after restarts, use a qualified domain name and pass it to all raylets at start time. Each raylet resolves the domain name and connects to the correct GCS. Make sure that only one GCS is alive at any time.

Note

GCS fault tolerance with external Redis is officially supported only if you use KubeRay for Ray Serve fault tolerance. In other cases, you can use it at your own risk, but you need to implement additional mechanisms that detect a failure of the GCS or the head node and restart it.

Note

You can also enable GCS fault tolerance when you run Ray on Anyscale. For instructions, see the Anyscale documentation.

Embedded RocksDB backend (alpha)#

Note

The embedded RocksDB backend is alpha and might change before it becomes stable. The Ray project welcomes feedback. Share your experience in a GitHub issue.

The Redis backend makes the GCS fault tolerant, but it also adds an external, highly available Redis instance that you have to deploy, secure, and operate. The embedded RocksDB backend removes that dependency. The GCS persists its state to a local RocksDB database on a persistent volume instead of to Redis. Instead of running a separate data store, you provide a directory on durable storage.

The recovery model is identical to Redis-backed fault tolerance. When the GCS restarts, it reads its state back from disk and resumes, and each raylet reconnects while the GCS recovers. The only difference is where the state persists, in a local RocksDB database instead of an external Redis instance.

Redis or RocksDB?#

External Redis

Embedded RocksDB

Extra process to operate

Yes, HA Redis

No

Where state lives

External Redis instance

Local RocksDB database on a persistent volume

Survives head node or Pod loss

Yes, if Redis survives

Yes, if the persistent volume survives and reattaches to the new head

Platform support

All platforms

Linux only

Maturity

Officially supported with KubeRay for Ray Serve

Alpha

Choose the embedded RocksDB backend when you want GCS fault tolerance without running Redis, and you can attach a durable volume, such as a Kubernetes PersistentVolume, that you can reattach to whichever node runs the GCS.

Enabling it#

Set two environment variables before you start the head node:

  • RAY_gcs_storage=rocksdb selects the backend.

  • RAY_gcs_storage_path=<dir> points at a directory on a persistent volume where RocksDB stores its files. This variable is required. If it’s unset, Ray fails fast at startup.

RAY_gcs_storage=rocksdb RAY_gcs_storage_path=/mnt/ray-gcs ray start --head

The directory must live on storage that survives a restart of the GCS or the head node, and that you can reattach to the node running the recovered GCS. HA Redis satisfies the same durability requirement for the Redis backend.

Note

The GCS process embeds the RocksDB database, which is single-writer. Make sure that only one GCS opens the storage path at a time. Point every restart of a cluster’s head node at the same path, and never share a path between clusters.

For a step-by-step Kubernetes walkthrough, see GCS fault tolerance with embedded RocksDB (alpha).

Advanced tuning#

RocksDB I/O runs on a dedicated thread pool so it never stalls the GCS event loop. This I/O includes the write-ahead-log fsync, which dominates write latency. The defaults suit the GCS metadata workload, so you usually don’t need to change them. Tune the pool with the following two environment variables:

  • RAY_gcs_rocksdb_io_pool_size: The number of worker threads in the RocksDB I/O offload pool. Defaults to 4.

  • RAY_gcs_rocksdb_strand_buckets: The number of per-key ordering buckets. Single-key operations are hashed into a bucket and serialized within it, while different buckets run concurrently. Defaults to 64.