[Core] Worker nodes dying with 1000 tasks

[psarka]

What happened + What you expected to happen

I am unable to run 1000 tasks in 1 job with ray - worker nodes start dying with Expected termination: received SIGTERM. My repro program is a simplified version of the example from the docs :

import time

import ray


@ray.remote(num_cpus=1, max_retries=0)
def process(task):
    print(f'Starting {task}')
    time.sleep(100)


if __name__ == '__main__':
    ray.init(log_to_driver=False)
    unfinished = [process.remote(i) for i in range(1000)]

    while unfinished:
        finished, unfinished = ray.wait(unfinished, num_returns=1, fetch_local=False)

When submitting this job with ray job submit, autoscaler starts spinning up worker nodes, which then proceed to die (at a slower rate, but still). For example, 12 minutes in, I see 18 alive nodes and 5 dead. With 4 tasks per node, 20 tasks have already failed.

Versions / Dependencies

ray 2.44.1, python 3.12, gcloud instances

Reproduction script

These are the commands that I'm running:

uv run ray up ray-cluster-repro.yaml -y
uv run ray attach ray-cluster-repro.yaml -p 10001
uv run ray dashboard ray-cluster-repro.yaml
RAY_ADDRESS="ray://localhost:10001" uv run ray job submit --no-wait --working-dir=./ --runtime-env=driver-env-repro.yaml -- python cluster_test.py

driver-env-repro.yaml:

env_vars:
  RAY_RUNTIME_ENV_HOOK: "ray._private.runtime_env.uv_runtime_env_hook.hook"

py_executable: "uv run"

ray-cluster-repro.yaml:

cluster_name: repro
max_workers: 1024
upscaling_speed: 1.0
docker:
  image: rayproject/ray:2.44.1-py312-cpu
  container_name: "ray_container"
  pull_before_run: True
  run_options:  # Extra options to pass into "docker run"
    - --ulimit nofile=65536:65536
  worker_image: "rayproject/ray:2.44.1-py312-cpu"

idle_timeout_minutes: 3

# Cloud-provider specific configuration.
provider:
    type: gcp
    region: europe-west4
    availability_zone: europe-west4-a
    project_id: axial-matter-417704

auth:
    ssh_user: ubuntu

available_node_types:
    ray_head_default:
        resources: {"CPU": 0}
        node_config:
            machineType: n4-highmem-8
            disks:
              - boot: true
                autoDelete: true
                type: PERSISTENT
                initializeParams:
                  diskSizeGb: 1000
                  # See https://cloud.google.com/compute/docs/images for more images
                  sourceImage: projects/deeplearning-platform-release/global/images/common-cpu-v20240922

    ray_worker_n4_standard_4:
      min_workers: 1
      max_workers: 100
      resources: {"CPU": 4}
      node_config:
        machineType: n4-standard-4
        disks:
          - boot: true
            autoDelete: true
            type: PERSISTENT
            initializeParams:
              diskSizeGb: 50
              sourceImage: projects/deeplearning-platform-release/global/images/common-cpu-v20240922
        scheduling:
          - preemptible: false
        serviceAccounts:
          - email: ray-autoscaler-sa-v1@axial-matter-417704.iam.gserviceaccount.com
            scopes:
              - https://www.googleapis.com/auth/cloud-platform


head_node_type: ray_head_default

file_mounts: {}

cluster_synced_files: []

file_mounts_sync_continuously: False

rsync_exclude:
    - "**/.git"
    - "**/.git/**"

rsync_filter:
    - ".gitignore"

initialization_commands: []

setup_commands: [
  pip install uv
]

head_setup_commands: []

worker_setup_commands: []

head_start_ray_commands:
    - ray stop
    - >-
      ray start
      --head
      --port=6379
      --object-manager-port=8076
      --autoscaling-config=~/ray_bootstrap_config.yaml

worker_start_ray_commands:
    - ray stop
    - >-
      ray start
      --address=$RAY_HEAD_IP:6379
      --object-manager-port=8076

Issue Severity

None