How to control the total memory of ray.serve?

I use ray[serve] as the server and send requests by http . But after sending hours of requests, I found the memory of ray’s actors is increasing continuously (over 90G for 20 actors). How can I control the total memory used by ray ?

Memory usage is increasing linearly.

  • OS: MacOS 10.15.7 / Linux Ubuntu 18 LTS
  • Ray version: 1.7.0/1.4.1
  • Python version: 3.7.11
  • Installation:
cat requirements
> ray[serve]==1.7.0
> psutil
> requests

pip install -r requirements

The code for reproduction is as follows. I have tried object_store_memory , but it doesn’t work for ray serve.

import json
import os
import numpy as np
import psutil
import ray
import requests
from ray import serve

@serve.deployment(name="test_qps", route_prefix="/test_qps", ray_actor_options={"num_cpus": 1}, num_replicas=1)
class QpsTest:
    def __init__(self):

    async def __call__(self, request):
        step_cnt = await request.json()
        state = np.random.randint(0, 255, (1000, 1000), np.uint64)

        process = psutil.Process(os.getpid())
        proc_mem = process.memory_info().rss / (1024 ** 2)
        print(f'actor_pid={} \t mem={proc_mem:6.1f} MB.')
        return state

if __name__ == '__main__':
    ray.init(num_cpus=1, dashboard_host="", object_store_memory=150_000_000)

    client = serve.start(http_options={"host": ""})

    step_no = 5000000

    step_cnt = 0
    url = ""
    while step_cnt < step_no:
        req =, data=json.dumps(step_cnt))
        step_cnt += 1
        process = psutil.Process(os.getpid())
        proc_mem = process.memory_info().rss / (1024 ** 2)
        print(f'main_pid={} \t mem={proc_mem:6.1f} MB.')
        print('-' * 30)

Hi @Rong thanks for posting. I created a tracking issue here and I’m looking into it. [Bug] Serve memory grow unbounded after hours of requests. · Issue #19891 · ray-project/ray · GitHub

1 Like

Thank you simon-mo, I will trace the issue and discuss more on that.

We found the memory leaking is not due to ray, but the third-party package pytorch_forecasting. The relative Github issue is Memory leak in TimeSeriesDataSet · Issue #648 · jdb78/pytorch-forecasting · GitHub

Thanks. ray-serve works very well on Debian.