read_namespaced_pod_log with follow=True hangs when there is a long gap between logs from pod

What happened (please include outputs or screenshots): If I try to read the logs of a pod with the following:

from kubernetes import config, client
config.load_incluster_config()
k8s = client.CoreV1Api()
namespace = "default"
pod_name="watch-this"
logs = k8s.read_namespaced_pod_log(
    name=pod_name,
    namespace=namespace,
    follow=True,
    _preload_content=False,
)
for line in logs:
  print(line)

and there is a gap (say 5 minutes) between line1 and line2 of the log being written, then the function hangs after the first line

What you expected to happen:

the function should read all the logs as they are written from the watched pod and then exit once finished

How to reproduce it (as minimally and precisely as possible):

Create the following pod to watch:

apiVersion: v1
kind: Pod
metadata:
  name: watch-this
  namespace: default
spec:
  containers:
  - command:
    - python3
    - -u
    - -c
    - |
      from time import sleep
      for i in range(3):
        print(i)
        if i != 2:
          sleep(300)
    image: python:3.7.4-slim-stretch
    imagePullPolicy: IfNotPresent
    name: python
    resources: {}
    stdin: true
    stdinOnce: true
    terminationMessagePath: /dev/termination-log
    terminationMessagePolicy: File
  priority: 0
  restartPolicy: Never
  schedulerName: default-scheduler
  serviceAccountName: default
  terminationGracePeriodSeconds: 30

then watch the pod with the following:

kubectl -n default run -i -t test-python-kube-client --image=python:3.7.4-slim-stretch --restart=Never --command -- /bin/sh

pip install kubernetes==11.0.0
python
from kubernetes import config, client
config.load_incluster_config()
k8s = client.CoreV1Api()
namespace = "default"
pod_name="watch-this"
logs = k8s.read_namespaced_pod_log(
    name=pod_name,
    namespace=namespace,
    follow=True,
    _preload_content=False,
)
for line in logs:
  print(line)

the output should get stuck like this: b'0\n'

however if I run the equivalent request using kubectl proxy and the kubernetes API, it works as expected:

kubectl proxy --port=8080 &
curl http://localhost:8080/api/v1/namespaces/default/pods/watch-this/log?follow=true
0
1
2

Anything else we need to know?:

Environment:

• Kubernetes version (Client Version: version.Info{Major:"1", Minor:"14", GitVersion:"v1.14.1", GitCommit:"b7394102d6ef778017f2ca4046abbaa23b88c290", GitTreeState:"clean", BuildDate:"2019-04-08T17:11:31Z", GoVersion:"go1.12.1", Compiler:"gc", Platform:"darwin/amd64"} Server Version: version.Info{Major:"1", Minor:"16", GitVersion:"v1.16.13", GitCommit:"1da71a35d52fa82847fd61c3db20c4f95d283977", GitTreeState:"clean", BuildDate:"2020-07-15T21:59:26Z", GoVersion:"go1.13.9", Compiler:"gc", Platform:"linux/amd64"}):
• OS ( MacOS 10.14.6):
• Python version (3.7.4)
• Python client version (kubernetes==11.0.0)

[raphael-bos]

I have faced the same problem this last week. After spending a fair amount of time investigating, I came to the conclusion the problem is with the TCP Keep Alive socket option, which urllib3 doesn’t set by default. So, while running a long http request, the server doesn’t get any tcp keep alive probes from the client and decides to close the connection without notifying the client, which then hangs waiting for the complete response from the server. These socket options worked for me (the times chosen are only for demonstration):

socket_options= [
    (socket.SOL_SOCKET, socket.SO_KEEPALIVE, 1),
    (socket.IPPROTO_TCP, socket.TCP_KEEPIDLE, 5),
    (socket.IPPROTO_TCP, socket.TCP_KEEPINTVL, 30),
    (socket.IPPROTO_TCP, socket.TCP_KEEPCNT, 3)
]

Socket options should be set in the urllib3.PoolManager class here: https://github.com/kubernetes-client/python/blob/master/kubernetes/client/rest.py#L106

Another side effect of running long tcp connections can be seen in #1158 . Basically, connections used by the PoolManager are never closed and therefore are unusable for subsequently requests. This also makes the client hangs if it tries to make another request (read_namespaced_pod, for example) after a long running one. A dirty hack is to call rest_client.pool_manager.clear() after a long running request.

[k8s-ci-robot]

@Ark-kun: You can’t reopen an issue/PR unless you authored it or you are a collaborator.

In response to this:

/reopen

Instructions for interacting with me using PR comments are available here. If you have questions or suggestions related to my behavior, please file an issue against the kubernetes/test-infra repository.