ray: [Bug] Resource demand backlog caps out at 10+ in multi-node K8s cluster

Search before asking

• I searched the issues and found no similar issues.

Ray Component

Ray Clusters

What happened + What you expected to happen

My Setup:

I have deployed a Ray Cluster on Kubernetes with EKS. The cluster has a worker type with 15 CPUs that is always on (min_workers: 1). Other worker types include 7 CPUs, 15 CPUs, 30 CPUs.

I’ve also configured: upscaling_speed=9999, AUTOSCALER_MAX_LAUNCH_BATCH=9999, AUTOSCALER_MAX_CONCURRENT_LAUNCHES=9999 as recommended here: https://github.com/ray-project/ray/issues/21683#issuecomment-1018015823

Main Issue:

Whenever I launch 200 tasks, the autoscaler only launches a single 7-cpu worker. After that worker is ready, it launches another single 7-cpu worker. I expect it to launch at least enough workers for 185 CPUs in the first iteration.

In the Ray Operator Pod Logs, I see {'CPU': 1.0}: 1+ pending tasks/actors. I imagine this should be something like “185+ pending tasks/actors”?

However, I can successfully use ray.autoscaler.sdk.request_resources to request 200 cpus all at once.

---

Ray Operator Pod Logs: kubectl logs -f po/ray-operator-987dc99b9-47v92 | grep Resources -A 10

Resources
---------------------------------------------------------------
Usage:
 0.0/15.0 CPU
 0.00/25.900 GiB memory
 0.00/10.249 GiB object_store_memory

Demands:
 (no resource demands)
py38-cu112,karpenter:2022-02-04 10:47:45,337    DEBUG gcs_utils.py:245 -- internal_kv_put b'__autoscaling_status_legacy' b"Cluster status: 1 nodes\n - MostDelayedHeartbeats: {'10.16.115.16': 0.6050498485565186, '10.16.126.217': 0.6049869060516357}\n - NodeIdleSeconds: Min=0 Mean=38219 Max=76437\n - ResourceUsage: 0.0/15.0 CPU, 0.0 GiB/25.9 GiB memory, 0.0 GiB/10.25 GiB object_store_memory\n - TimeSinceLastHeartbeat: Min=0 Mean=0 Max=0\nWorker node types:\n - wkr-15cpu30g-ondemand: 1" True None
py38-cu112,karpenter:2022-02-04 10:47:45,338    DEBUG legacy_info_string.py:26 -- Cluster status: 1 nodes
--
Resources
---------------------------------------------------------------
Usage:
 4.0/15.0 CPU
 0.00/25.900 GiB memory
 0.00/10.249 GiB object_store_memory

Demands:
 {'CPU': 1.0}: 1+ pending tasks/actors
py38-cu112,karpenter:2022-02-04 10:47:51,278    DEBUG gcs_utils.py:245 -- internal_kv_put b'__autoscaling_status_legacy' b"Cluster status: 1 nodes\n - MostDelayedHeartbeats: {'10.16.115.16': 0.5087563991546631, '10.16.126.217': 0.5086848735809326}\n - NodeIdleSeconds: Min=0 Mean=0 Max=0\n - ResourceUsage: 4.0/15.0 CPU, 0.0 GiB/25.9 GiB memory, 0.0 GiB/10.25 GiB object_store_memory\n - TimeSinceLastHeartbeat: Min=0 Mean=0 Max=0\nWorker node types:\n - wkr-15cpu30g-ondemand: 1" True None
py38-cu112,karpenter:2022-02-04 10:47:51,289    DEBUG legacy_info_string.py:26 -- Cluster status: 1 nodes
--
Resources
---------------------------------------------------------------
Usage:
 10.0/15.0 CPU
 0.00/25.900 GiB memory
 0.00/10.249 GiB object_store_memory

Demands:
 {'CPU': 1.0}: 1+ pending tasks/actors
py38-cu112,karpenter:2022-02-04 10:47:57,323    DEBUG gcs_utils.py:245 -- internal_kv_put b'__autoscaling_status_legacy' b"Cluster status: 1 nodes\n - MostDelayedHeartbeats: {'10.16.115.16': 0.4549856185913086, '10.16.126.217': 0.4549136161804199}\n - NodeIdleSeconds: Min=0 Mean=0 Max=0\n - ResourceUsage: 10.0/15.0 CPU, 0.0 GiB/25.9 GiB memory, 0.0 GiB/10.25 GiB object_store_memory\n - TimeSinceLastHeartbeat: Min=0 Mean=0 Max=0\nWorker node types:\n - wkr-15cpu30g-ondemand: 1" True None
py38-cu112,karpenter:2022-02-04 10:47:57,324    DEBUG legacy_info_string.py:26 -- Cluster status: 1 nodes
--
Resources
---------------------------------------------------------------
Usage:
 15.0/15.0 CPU
 0.00/25.900 GiB memory
 0.00/10.249 GiB object_store_memory

Demands:
 {'CPU': 1.0}: 1+ pending tasks/actors
py38-cu112,karpenter:2022-02-04 10:48:03,348    DEBUG gcs_utils.py:245 -- internal_kv_put b'__autoscaling_status_legacy' b"Cluster status: 1 nodes\n - MostDelayedHeartbeats: {'10.16.115.16': 0.54921555519104, '10.16.126.217': 0.5491447448730469}\n - NodeIdleSeconds: Min=0 Mean=0 Max=0\n - ResourceUsage: 15.0/15.0 CPU, 0.0 GiB/25.9 GiB memory, 0.0 GiB/10.25 GiB object_store_memory\n - TimeSinceLastHeartbeat: Min=0 Mean=0 Max=0\nWorker node types:\n - wkr-15cpu30g-ondemand: 1" True None
py38-cu112,karpenter:2022-02-04 10:48:03,350    DEBUG legacy_info_string.py:26 -- Cluster status: 1 nodes
--
Resources
---------------------------------------------------------------
Usage:
 15.0/15.0 CPU
 0.00/25.900 GiB memory
 0.00/10.249 GiB object_store_memory

Demands:
 {'CPU': 1.0}: 1+ pending tasks/actors

Ray Pod Description:

Name:         ray-operator-987dc99b9-47v92
Namespace:    karpenter
Priority:     0
Node:         ip-10-16-65-175.us-west-2.compute.internal/10.16.65.175
Start Time:   Thu, 03 Feb 2022 13:05:01 -0600
Labels:       cluster.ray.io/component=operator
              pod-template-hash=987dc99b9
Annotations:  kubernetes.io/psp: eks.privileged
Status:       Running
IP:           10.16.68.93
IPs:
  IP:           10.16.68.93
Controlled By:  ReplicaSet/ray-operator-987dc99b9
Containers:
  ray:
    Container ID:  docker://f78dc8efee6c1f8483438e325c275a281c310341a1f304f4811aa8c80d263903
    Image:         rayproject/ray:bbc64e
    Image ID:      docker-pullable://rayproject/ray@sha256:6a0520c449555a31418ed79188827553e43276db6abc1bc4341edd16dd832eb3
    Port:          <none>
    Host Port:     <none>
    Command:
      ray-operator
    State:          Running
      Started:      Thu, 03 Feb 2022 13:05:03 -0600
    Ready:          True
    Restart Count:  0
    Limits:
      cpu:     1
      memory:  2Gi
    Requests:
      cpu:                1
      ephemeral-storage:  1Gi
      memory:             1Gi
    Environment:
      AUTOSCALER_MAX_NUM_FAILURES:         inf
      AUTOSCALER_MAX_LAUNCH_BATCH:         9999
      AUTOSCALER_MAX_CONCURRENT_LAUNCHES:  9999
    Mounts:
      /var/run/secrets/kubernetes.io/serviceaccount from kube-api-access-wq2kl (ro)
Conditions:
  Type              Status
  Initialized       True
  Ready             True
  ContainersReady   True
  PodScheduled      True
Volumes:
  kube-api-access-wq2kl:
    Type:                    Projected (a volume that contains injected data from multiple sources)
    TokenExpirationSeconds:  3607
    ConfigMapName:           kube-root-ca.crt
    ConfigMapOptional:       <nil>
    DownwardAPI:             true
QoS Class:                   Burstable
Node-Selectors:              <none>
Tolerations:                 node.kubernetes.io/not-ready:NoExecute op=Exists for 300s
                             node.kubernetes.io/unreachable:NoExecute op=Exists for 300s
Events:                      <none>

Versions / Dependencies

Ray Operator Image: rayproject/ray:bbc64e Ray Client Version: 1.9.2 Python 3.8 Kubernetes Version: 1.21 with EKS

Reproduction script

from pprint import pprint

import ray
ray.init("ray://mycluster.internal:10001")

@ray.remote
def task():
    import time
    time.sleep(30)

pprint(ray.cluster_resources())
results = ray.get([task.remote() for _ in range(200)])

Anything else

This issue occurs every time.

Are you willing to submit a PR?

• Yes I am willing to submit a PR!