Dashboard loading slowly and hang #805

dbalabka · 2023-08-24T14:45:21Z

Describe the issue:

The dashboard loads very slowly. Port forwarding reports timeouts and broken pipe errors during dashboard loads:

❯ kubectl port-forward 'cluser-scheduler-d99f64dc8-hwkt5' 8787:8787
Forwarding from 127.0.0.1:8787 -> 8787
Forwarding from [::1]:8787 -> 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
E0824 17:23:18.023748   11638 portforward.go:378] error copying from remote stream to local connection: readfrom tcp6 [::1]:8787->[::1]:35270: write tcp6 [::1]:8787->[::1]:35270: write: broken pipe
Handling connection for 8787

❯ kubectl port-forward 'cluser-scheduler-d99f64dc8-hwkt5' 8787:8787
Forwarding from 127.0.0.1:8787 -> 8787
Forwarding from [::1]:8787 -> 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
E0824 17:15:27.312003   11363 portforward.go:368] error creating forwarding stream for port 8787 -> 8787: Timeout occurred
E0824 17:15:27.312033   11363 portforward.go:368] error creating forwarding stream for port 8787 -> 8787: Timeout occurred
E0824 17:15:27.312060   11363 portforward.go:368] error creating forwarding stream for port 8787 -> 8787: Timeout occurred
E0824 17:15:27.312063   11363 portforward.go:368] error creating forwarding stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
Handling connection for 8787
Handling connection for 8787
E0824 17:15:57.375975   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
E0824 17:15:57.392403   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
E0824 17:16:20.522829   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
E0824 17:16:27.573642   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
Handling connection for 8787
E0824 17:16:58.658054   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
E0824 17:16:58.658175   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
E0824 17:17:28.665025   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
Handling connection for 8787
E0824 17:18:03.699222   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
E0824 17:18:03.699310   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
E0824 17:18:33.708104   11363 portforward.go:346] error creating error stream for port 8787 -> 8787: Timeout occurred
Handling connection for 8787
Handling connection for 8787

Disabling JS solve the issue. It looks like it is related to WebSockets. Adding Ingress with an HTTPS certificate solves the problem.

Minimal Complete Verifiable Example:

from dask_kubernetes.operator import KubeCluster, make_cluster_spec

spec = make_cluster_spec(
    name="dask-cluser", 
    image='ghcr.io/dask/dask:2023.8.1-py3.10',
    n_workers=1, 
    env={
        "EXTRA_PIP_PACKAGES": "s3fs",
    },
)
 
cluster = KubeCluster(
    namespace="dask-operator", 
    custom_cluster_spec=spec,
)
cluster = KubeCluster.from_name('dmitryb-cluser')

Anything else we need to know?:

Environment:

Required poetry packages:

[tool.poetry]
name = "test-bug"
version = "0.1.0"
description = ""
authors = ["Your Name <[email protected]>"]
readme = "README.md"

[tool.poetry.dependencies]
python = "^3.10.10,<3.11"
pandas = "^2.0.0"
dask-kubernetes = "~2023.8.0"
dask = "~2023.8.0"
s3fs = "^2023.6.0"

[build-system]
requires = ["poetry-core>=1.0.0"]
build-backend = "poetry.core.masonry.api"

K8S version:

❯ kubectl version
WARNING: This version information is deprecated and will be replaced with the output from kubectl version --short.  Use --output=yaml|json to get the full version.
Client Version: version.Info{Major:"1", Minor:"24", GitVersion:"v1.24.16", GitCommit:"51e33fadff13065ae5518db94e84598293965939", GitTreeState:"clean", BuildDate:"2023-07-19T12:26:21Z", GoVersion:"go1.20.6", Compiler:"gc", Platform:"linux/amd64"}
Kustomize Version: v4.5.4
Server Version: version.Info{Major:"1", Minor:"24", GitVersion:"v1.24.14+rke2r1", GitCommit:"0018fa8af8ffaff22f36d3bd86289753ca61da81", GitTreeState:"clean", BuildDate:"2023-05-18T17:22:39Z", GoVersion:"go1.19.9 X:boringcrypto", Compiler:"gc", Platform:"linux/amd64"}

The text was updated successfully, but these errors were encountered:

dbalabka · 2023-08-24T14:48:52Z

I assume kubectl port-forward does not work properly with WebSockets. The solution might be providing the option to create Ingress attached to existing service. make_cluster_spec function can generate the following YAML:

apiVersion: networking.k8s.io/v1
kind: Ingress
metadata:
  name: cluser-scheduler
  namespace: dask-operator
spec:
  rules:
  - host: <Ask user to provide a host because it is Kubernetes setup specific value>
    http:
      paths:
      - backend:
          service:
            name: cluser-scheduler
            port:
              number: 8787
        pathType: ImplementationSpecific

dbalabka · 2023-08-24T14:55:59Z

It seems that kubectl port-forward does not support WS connections
kubernetes/kubernetes#74884

jacobtomlinson · 2023-08-24T15:13:23Z

We have used kubectl port-forward successfully with the Dask Dashboard for many years, so I don't think it's that. Websockets work just fine in my experience, however, the connection is known to close after some period of time (see #663).

It looks to me like the network connection between your machine and the Kubernetes API is not very stable given all the timeout errors in your output. Or the Kubernetes API nodes are too busy.

I agree that you might want to expose an Ingress instead of using port-forwards if your k8s cluster is set up for that. However, that is out of scope for our Python API because there are just so many different ways to do it. I would recommend defining your cluster in YAML along with an appropriate Ingress and then optionally from Python use KubeCluster.from_name to get an object pointer to it for convenience.

Alternatively you could change the service type to LoadBalancer if you have LB provisioning configured on your cluster, in that case KubeCluster will wait for the LB to go into service and connect to that instead of performing a port-forward.

dbalabka · 2023-08-25T08:59:39Z

It looks to me like the network connection between your machine and the Kubernetes API is not very stable given all the timeout errors in your output. Or the Kubernetes API nodes are too busy.

Taking into account that Ingress working fine, it proves that the network connection is good. Also, nodes do not perform any work at this time.

I agree that you might want to expose an Ingress instead of using port-forwards if your k8s cluster is set up for that. However, that is out of scope for our Python API because there are just so many different ways to do it.

I agree that it is too specific for each cluster. I'm looking for an elegant solution for a team and avoiding the boilerplate code. Probably, creating an extra function that will create cluster-specific YAML should solve the issue.

Alternatively you could change the service type to LoadBalancer if you have LB provisioning configured on your cluster, in that case KubeCluster will wait for the LB to go into service and connect to that instead of performing a port-forward.

Thanks, I will take a look into it.

jacobtomlinson · 2023-08-25T11:38:37Z

Taking into account that Ingress working fine, it proves that the network connection is good. Also, nodes do not perform any work at this time.

I was talking about the control plane, not the nodes. Port forwarding requires the control plane to proxy the traffic, an Ingress goes to the nodes directly.

Probably, creating an extra function that will create cluster-specific YAML should solve the issue.

If you want to do it in Python you could do something like this. But it's still a lot of boilerplate.

import kr8s

ingress = kr8s.objects.Ingress(
    {
        "apiVersion": "networking.k8s.io/v1",
        "kind": "Ingress",
        "metadata": {"name": "cluser-scheduler", "namespace": "dask-operator"},
        "spec": {
            "rules": [
                {
                    "host": "<Ask user to provide a host because it is Kubernetes setup specific value>",
                    "http": {
                        "paths": [
                            {
                                "backend": {
                                    "service": {
                                        "name": "cluser-scheduler",
                                        "port": {"number": 8787},
                                    }
                                },
                                "pathType": "ImplementationSpecific",
                            }
                        ]
                    },
                }
            ]
        },
    }
)
ingress.create()

dbalabka · 2023-08-25T11:44:09Z

@jacobtomlinson, so it seems the problem is connected with containerd runtime. Recently, we migrated from docker to containerd. Probably, you are not noticing the problem because of the docker runtime.
kubernetes/kubernetes#74551 (comment)

jacobtomlinson · 2023-08-25T12:25:09Z

@dbalabka I'm not sure that's the problem. We regularly use GKE for testing which uses containerd by default.

dbalabka · 2023-08-25T12:43:16Z

I tried to deploy it on the old cluster. So, my observations are following:

docker engine 20.10.7 - dashboard works
containerd://1.7.1-k3s1 - doesn't work

I probably have to use Ingress or ask to upgrade containerd version that contains:
containerd/containerd#8418

dbalabka · 2023-08-25T13:06:09Z

Actually, I was wrong. The same issue with the docker engine. The dashboard worked for some short period of time and started to hang. Probably, Ingress is not such a bad option. I suspect that the Upgrade of containerd will not help. @jacobtomlinson , you are right it is something environment-specific.

dbalabka · 2023-08-29T07:57:02Z

As suggested @jacobtomlinson, the problem workaround is using NodePort instead of ClusterIP:

spec = make_cluster_spec(
    ...
    scheduler_service_type="NodePort",
    ...
)
 
cluster = KubeCluster(
    ...
    custom_cluster_spec=spec,
)

dbalabka · 2024-01-15T10:20:24Z

@jacobtomlinson, it seems that in the latest version of dask-kubernetes or the newer version of k8s, the problem is gone. I will check it again right after #851 will be fixed.

jacobtomlinson added the question label Aug 24, 2023

jacobtomlinson mentioned this issue Aug 25, 2023

Utility functions to generate common resources kr8s-org/kr8s#142

Open

dbalabka closed this as completed Aug 25, 2023

dbalabka mentioned this issue Feb 14, 2024

Update dask-kubernetes to a newer kr8s #853

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Dashboard loading slowly and hang #805

Dashboard loading slowly and hang #805

dbalabka commented Aug 24, 2023 •

edited

Loading

dbalabka commented Aug 24, 2023 •

edited

Loading

dbalabka commented Aug 24, 2023

jacobtomlinson commented Aug 24, 2023 •

edited

Loading

dbalabka commented Aug 25, 2023

jacobtomlinson commented Aug 25, 2023

dbalabka commented Aug 25, 2023

jacobtomlinson commented Aug 25, 2023

dbalabka commented Aug 25, 2023

dbalabka commented Aug 25, 2023

dbalabka commented Aug 29, 2023

dbalabka commented Jan 15, 2024

Dashboard loading slowly and hang #805

Dashboard loading slowly and hang #805

Comments

dbalabka commented Aug 24, 2023 • edited Loading

dbalabka commented Aug 24, 2023 • edited Loading

dbalabka commented Aug 24, 2023

jacobtomlinson commented Aug 24, 2023 • edited Loading

dbalabka commented Aug 25, 2023

jacobtomlinson commented Aug 25, 2023

dbalabka commented Aug 25, 2023

jacobtomlinson commented Aug 25, 2023

dbalabka commented Aug 25, 2023

dbalabka commented Aug 25, 2023

dbalabka commented Aug 29, 2023

dbalabka commented Jan 15, 2024

dbalabka commented Aug 24, 2023 •

edited

Loading

dbalabka commented Aug 24, 2023 •

edited

Loading

jacobtomlinson commented Aug 24, 2023 •

edited

Loading