setup_global_ray_cluster#

ray.util.spark.setup_global_ray_cluster(*, max_worker_nodes: int, is_blocking: bool = True, min_worker_nodes: int | None = None, num_cpus_worker_node: int | None = None, num_cpus_head_node: int | None = None, num_gpus_worker_node: int | None = None, num_gpus_head_node: int | None = None, memory_worker_node: int | None = None, memory_head_node: int | None = None, object_store_memory_worker_node: int | None = None, object_store_memory_head_node: int | None = None, head_node_options: Dict | None = None, worker_node_options: Dict | None = None, strict_mode: bool = False, collect_log_to_path: str | None = None, autoscale_upscaling_speed: float | None = 1.0, autoscale_idle_timeout_minutes: float | None = 1.0)[source]#

Set up a global mode cluster. The global Ray on spark cluster means: - You can only create one active global Ray on spark cluster at a time. On databricks cluster, the global Ray cluster can be used by all users, - as contrast, non-global Ray cluster can only be used by current notebook user. - It is up persistently without automatic shutdown. - On databricks notebook, you can connect to the global cluster by calling ray.init() without specifying its address, it will discover the global cluster automatically if it is up.

For global mode, the ray_temp_root_dir argument is not supported. Global model Ray cluster always use the default Ray temporary directory path.

All arguments are the same with setup_ray_cluster API except that: - the ray_temp_root_dir argument is not supported. Global model Ray cluster always use the default Ray temporary directory path. - A new argument “is_blocking” (default True) is added. If “is_blocking” is True, then keep the call blocking until it is interrupted. once the call is interrupted, the global Ray on spark cluster is shut down and setup_global_ray_cluster call terminates. If “is_blocking” is False, once Ray cluster setup completes, return immediately.

Parameters:
  • max_worker_nodes (int) – The maximum number of Ray worker nodes to start. Set it to ray.util.spark.MAX_NUM_WORKER_NODES to use all available resources of the Spark application.

  • is_blocking (bool) – If True, block until the call is interrupted, then shut down the cluster. If False, return once the cluster is set up.

  • min_worker_nodes (int | None) – The minimum number of worker nodes. If None or equal to max_worker_nodes, autoscaling is disabled.

  • num_cpus_worker_node (int | None) – The number of CPUs available to each Ray worker node. See setup_ray_cluster for the default.

  • num_cpus_head_node (int | None) – The number of CPUs available to the Ray head node. Defaults to the number of CPU cores on the Spark driver node.

  • num_gpus_worker_node (int | None) – The number of GPUs available to each Ray worker node. See setup_ray_cluster for the default.

  • num_gpus_head_node (int | None) – The number of GPUs available to the Ray head node. Defaults to the number of GPUs on the Spark driver node.

  • memory_worker_node (int | None) – The heap memory for each Ray worker node, passed as --memory to ray start.

  • memory_head_node (int | None) – The heap memory for the Ray head node, passed as --memory to ray start.

  • object_store_memory_worker_node (int | None) – The object store memory for each Ray worker node. See setup_ray_cluster for the default and cap.

  • object_store_memory_head_node (int | None) – The object store memory for the Ray head node. See setup_ray_cluster for the default and cap.

  • head_node_options (Dict | None) – Extra ray start options for the head node, with keys in foo_bar form and None as the value of a flag option.

  • worker_node_options (Dict | None) – Extra ray start options for the worker nodes, with keys in foo_bar form and None as the value of a flag option.

  • strict_mode (bool) – If True, raise an exception when the Spark cluster lacks the resources for the requested allocation. If False, log a warning.

  • collect_log_to_path (str | None) – If set, the path to collect the Ray node logs to after the nodes terminate.

  • autoscale_upscaling_speed (float | None) – The number of nodes allowed to be pending as a multiple of the current number of nodes, when autoscaling is enabled.

  • autoscale_idle_timeout_minutes (float | None) – The number of minutes a worker node must be idle before the autoscaler removes it, when autoscaling is enabled.

Returns:

If is_blocking is False, a tuple of (address, remote_connection_address), as setup_ray_cluster returns. Otherwise None, once the cluster shuts down.