Disable Transparent Huge Pages from Kubernetes - docker

I deploy Redis container via Kubernetes and get the following warning:
WARNING you have Transparent Huge Pages (THP) support enabled in your kernel. This will create latency and memory usage issues with Redis. To fix this issue run the command 'echo never > /sys/kernel/mm/transparent_hugepage/enabled' as root, and add it to your /etc/rc.local in order to retain the setting after a reboot. Redis must be restarted after THP is disabled
Is it possible to disable THP via Kubernetes? Perhaps via init-containers?

Yes, with init-containers it's quite straightforward:
apiVersion: v1
kind: Pod
metadata:
name: thp-test
spec:
restartPolicy: Never
terminationGracePeriodSeconds: 1
volumes:
- name: host-sys
hostPath:
path: /sys
initContainers:
- name: disable-thp
image: busybox
volumeMounts:
- name: host-sys
mountPath: /host-sys
command: ["sh", "-c", "echo never >/host-sys/kernel/mm/transparent_hugepage/enabled"]
containers:
- name: busybox
image: busybox
command: ["cat", "/sys/kernel/mm/transparent_hugepage/enabled"]
Demo (notice that this is a system wide setting):
$ ssh THATNODE cat /sys/kernel/mm/transparent_hugepage/enabled
always [madvise] never
$ kubectl create -f thp-test.yaml
pod "thp-test" created
$ kubectl logs thp-test
always madvise [never]
$ kubectl delete pod thp-test
pod "thp-test" deleted
$ ssh THATNODE cat /sys/kernel/mm/transparent_hugepage/enabled
always madvise [never]

Ay,
I don't know if what I did is a good idea but we needed to deactivate THP on all our K8S VMs for all our apps. So I used a DaemonSet instead of adding an init-container to all our stacks :
apiVersion: apps/v1
kind: DaemonSet
metadata:
name: thp-disable
namespace: kube-system
spec:
selector:
matchLabels:
name: thp-disable
template:
metadata:
labels:
name: thp-disable
spec:
restartPolicy: Always
terminationGracePeriodSeconds: 1
volumes:
- name: host-sys
hostPath:
path: /sys
initContainers:
- name: disable-thp
image: busybox
volumeMounts:
- name: host-sys
mountPath: /host-sys
command: ["sh", "-c", "echo never >/host-sys/kernel/mm/transparent_hugepage/enabled"]
containers:
- name: busybox
image: busybox
command: ["watch", "-n", "600", "cat", "/sys/kernel/mm/transparent_hugepage/enabled"]
I think it's a little dirty but it works.

Related

Knative service deployment fails with reason RevisionMissing

I have deployed a service on Knative. I iterated on the service code/Docker image and I try to redeploy it at the same address. I proceeded as follow:
Pushed the new Docker image on our private Docker repo
Updated the service YAML file to point to the new Docker image (see YAML below)
Delete the service with the command: kubectl -n myspacename delete -f myservicename.yaml
Recreate the service with the command: kubectl -n myspacename apply -f myservicename.yaml
During the deployment, the service shows READY = Unknown and REASON = RevisionMissing, and after a while, READY = False and REASON = ProgressDeadlineExceeded. When looking at the logs of the pod with the following command kubectl -n myspacename logs revision.serving.knative.dev/myservicename-00001, I get the message:
no kind "Revision" is registered for version "serving.knative.dev/v1" in scheme "pkg/scheme/scheme.go:28"
Here is the YAML file of the service:
---
apiVersion: serving.knative.dev/v1
kind: Service
metadata:
name: myservicename
namespace: myspacename
spec:
template:
metadata:
annotations:
autoscaling.knative.dev/class: kpa.autoscaling.knative.dev
autoscaling.knative.dev/metric: concurrency
autoscaling.knative.dev/target: '1'
autoscaling.knative.dev/minScale: '0'
autoscaling.knative.dev/maxScale: '5'
autoscaling.knative.dev/scaleDownDelay: 60s
autoscaling.knative.dev/window: 600s
spec:
tolerations:
- key: nvidia.com/gpu
operator: Exists
effect: NoSchedule
volumes:
- name: nfs-volume
persistentVolumeClaim:
claimName: myspacename-models-pvc
imagePullSecrets:
- name: myrobotaccount-pull-secret
containers:
- name: myservicename
image: quay.company.com/project/myservicename:0.4.0
ports:
- containerPort: 5000
name: user-port
protocol: TCP
resources:
limits:
cpu: "4"
memory: 36Gi
nvidia.com/gpu: 1
requests:
cpu: "2"
memory: 32Gi
volumeMounts:
- name: nfs-volume
mountPath: /tmp/static/
securityContext:
privileged: true
env:
- name: CLOUD_STORAGE_PASSWORD
valueFrom:
secretKeyRef:
name: myservicename-cloud-storage-password
key: key
envFrom:
- configMapRef:
name: myservicename-config
The protocol I followed above is correct, the problem was because of a bug in the code of the Docker image that Knative is serving. I was able to troubleshoot the issue by looking at the logs of the pods as follow:
First run the following command to get the pod name: kubectl -n myspacename get pods. Example of pod name = myservicename-00001-deployment-56595b764f-dl7x6
Then get the logs of the pod with the following command: kubectl -n myspacename logs myservicename-00001-deployment-56595b764f-dl7x6

Why I cannot read files from a shared PersistentVolumeClaim between containers in Kubernetes?

I have a docker image felipeogutierrez/tpch-dbgen that I build using docker-compose and I push it to docker-hub registry using travis-CI.
version: "3.7"
services:
other-images: ....
tpch-dbgen:
build: ../docker/tpch-dbgen
image: felipeogutierrez/tpch-dbgen
volumes:
- tpch-dbgen-data:/opt/tpch-dbgen/data/
- datarate:/tmp/
stdin_open: true
and this is the Dockerfile to build this image:
FROM gcc AS builder
RUN mkdir -p /opt
COPY ./generate-tpch-dbgen.sh /opt/generate-tpch-dbgen.sh
WORKDIR /opt
RUN chmod +x generate-tpch-dbgen.sh && ./generate-tpch-dbgen.sh
In the end, this scripts creates a directory /opt/tpch-dbgen/data/ with some files that I would like to read from another docker image that I am running on Kubernetes. Then I have a Flink image that I create to run into Kubernetes. This image starts 3 Flink Task Managers and one stream application that reads files from the image tpch-dbgen-data. I think that the right approach is to create a PersistentVolumeClaim so I can share the directory /opt/tpch-dbgen/data/ from image felipeogutierrez/tpch-dbgen to my flink image in Kubernetes. So, first I have this file to create the PersistentVolumeClaim:
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
name: tpch-dbgen-data-pvc
spec:
accessModes:
- ReadWriteOnce
resources:
requests:
storage: 200Mi
Then, I am creating an initContainers to launch the image felipeogutierrez/tpch-dbgen and after that launch my image felipeogutierrez/explore-flink:1.11.1-scala_2.12:
apiVersion: apps/v1
kind: Deployment
metadata:
name: flink-taskmanager
spec:
replicas: 3
selector:
matchLabels:
app: flink
component: taskmanager
template:
metadata:
labels:
app: flink
component: taskmanager
spec:
initContainers:
- name: tpch-dbgen
image: felipeogutierrez/tpch-dbgen
#imagePullPolicy: Always
env:
command: ["ls"]
# command: ['sh', '-c', 'for i in 1 2 3; do echo "job-1 `date`" && sleep 5s; done;', 'ls']
volumeMounts:
- name: tpch-dbgen-data
mountPath: /opt/tpch-dbgen/data
containers:
- name: taskmanager
image: felipeogutierrez/explore-flink:1.11.1-scala_2.12
#imagePullPolicy: Always
env:
args: ["taskmanager"]
ports:
- containerPort: 6122
name: rpc
- containerPort: 6125
name: query-state
livenessProbe:
tcpSocket:
port: 6122
initialDelaySeconds: 30
periodSeconds: 60
volumeMounts:
- name: flink-config-volume
mountPath: /opt/flink/conf/
- name: tpch-dbgen-data
mountPath: /opt/tpch-dbgen/data
securityContext:
runAsUser: 9999 # refers to user _flink_ from official flink image, change if necessary
volumes:
- name: flink-config-volume
configMap:
name: flink-config
items:
- key: flink-conf.yaml
path: flink-conf.yaml
- key: log4j-console.properties
path: log4j-console.properties
- name: tpch-dbgen-data
persistentVolumeClaim:
claimName: tpch-dbgen-data-pvc
The Flink stream application is starting but it cannot read the files on the directory /opt/tpch-dbgen/data of the image felipeogutierrez/tpch-dbgen. I am getting the error: java.io.FileNotFoundException: /opt/tpch-dbgen/data/orders.tbl (No such file or directory). It is strange because when I try to go into the container felipeogutierrez/tpch-dbgen I can list the files. So I suppose there is something wrong on my Kubernetes configuration. Does anyone know to point what I am missing on the Kubernetes configuration files?
$ docker run -i -t felipeogutierrez/tpch-dbgen /bin/bash
root#10c0944a95f8:/opt# pwd
/opt
root#10c0944a95f8:/opt# ls tpch-dbgen/data/
customer.tbl dbgen dists.dss lineitem.tbl nation.tbl orders.tbl part.tbl partsupp.tbl region.tbl supplier.tbl
Also, when I list the logs of the container tpch-dbgen I can see the directory tpch-dbgen that I want to read. Although I cannot execute the command command: ["ls tpch-dbgen"] inside my Kubernetes config file.
$ kubectl get pods
NAME READY STATUS RESTARTS AGE
flink-jobmanager-n9nws 1/1 Running 2 17m
flink-taskmanager-777cb5bf77-ncdl4 1/1 Running 0 4m54s
flink-taskmanager-777cb5bf77-npmrx 1/1 Running 0 4m54s
flink-taskmanager-777cb5bf77-zc2nw 1/1 Running 0 4m54s
$ kubectl logs flink-taskmanager-777cb5bf77-ncdl4 tpch-dbgen
generate-tpch-dbgen.sh
tpch-dbgen
Docker has an unusual feature where, under some specific circumstances, it will populate a newly created volume from the image. You should not rely on this functionality, since it completely ignores updates in the underlying images and it doesn't work on Kubernetes.
In your Kubernetes setup, you create a new empty PersistentVolumeClaim, and then mount this over your actual data in both the init and main containers. As with all Unix mounts, this hides the data that was previously in that directory. Nothing causes data to get copied into that volume. This works the same way as every other kind of mount, except the Docker named-volume mount: you'll see the same behavior if you change your Compose setup to do a host bind mount, or if you play around with your local development system using a USB drive as a "volume".
You need to make your init container (or something else) explicitly copy data into the directory. For example:
initContainers:
- name: tpch-dbgen
image: felipeogutierrez/tpch-dbgen
command:
- /bin/cp
- -a
- /opt/tpch-dbgen/data
- /data
volumeMounts:
- name: tpch-dbgen-data
mountPath: /data # NOT the same path as in the image
If the main process modifies these files in place, you can make the command be more intelligent, or write a script into your image that only copies the individual files in if they don't exist yet.
It could potentially make more sense to have your image generate the data files at startup time, rather than at image-build time. That could look like:
FROM gcc
COPY ./generate-tpch-dbgen.sh /usr/local/bin/
RUN chmod +x /usr/local/bin/generate-tpch-dbgen.sh
CMD ["generate-tpch-dbgen.sh"]
Then in your init container, you can run the default command (the generate script) with the working directory set to the volume directory
initContainers:
- name: tpch-dbgen
image: felipeogutierrez/tpch-dbgen
volumeMounts:
- name: tpch-dbgen-data
mountPath: /opt/tpch-dbgen/data # or anywhere really
workingDir: /opt/tpch-dbgen/data # matching mountPath
I got to run the PersistentVolumeClaim and share it between pods. Basically I had to use a subPath property which I learned from this answer https://stackoverflow.com/a/43404857/2096986 and I am using a simple Job that I learned from this answer https://stackoverflow.com/a/64023672/2096986. The final results is below:
The Dockerfile:
FROM gcc AS builder
RUN mkdir -p /opt
COPY ./generate-tpch-dbgen.sh /opt/generate-tpch-dbgen.sh
WORKDIR /opt
RUN chmod +x /opt/generate-tpch-dbgen.sh
ENTRYPOINT ["/bin/sh","/opt/generate-tpch-dbgen.sh"]
and the script generate-tpch-dbgen.sh has to have this line in the end sleep infinity & wait to not finalize. The PersistentVolumeClaim is the same of the question. Then I create a Job with the subPath property.
apiVersion: batch/v1
kind: Job
metadata:
name: tpch-dbgen-job
spec:
template:
metadata:
labels:
app: flink
component: tpch-dbgen
spec:
restartPolicy: OnFailure
volumes:
- name: tpch-dbgen-data
persistentVolumeClaim:
claimName: tpch-dbgen-data-pvc
containers:
- name: tpch-dbgen
image: felipeogutierrez/tpch-dbgen
imagePullPolicy: Always
volumeMounts:
- mountPath: /opt/tpch-dbgen/data
name: tpch-dbgen-data
subPath: data
and I use it on the other deployment also with the subPath property.
apiVersion: apps/v1
kind: Deployment
metadata:
name: flink-taskmanager
spec:
replicas: 3
selector:
matchLabels:
app: flink
component: taskmanager
template:
metadata:
labels:
app: flink
component: taskmanager
spec:
volumes:
- name: flink-config-volume
configMap:
name: flink-config
items:
- key: flink-conf.yaml
path: flink-conf.yaml
- key: log4j-console.properties
path: log4j-console.properties
- name: tpch-dbgen-data
persistentVolumeClaim:
claimName: tpch-dbgen-data-pvc
containers:
- name: taskmanager
image: felipeogutierrez/explore-flink:1.11.1-scala_2.12
imagePullPolicy: Always
env:
args: ["taskmanager"]
ports:
- containerPort: 6122
name: rpc
- containerPort: 6125
name: query-state
livenessProbe:
tcpSocket:
port: 6122
initialDelaySeconds: 30
periodSeconds: 60
volumeMounts:
- name: flink-config-volume
mountPath: /opt/flink/conf/
- name: tpch-dbgen-data
mountPath: /opt/tpch-dbgen/data
subPath: data
securityContext:
runAsUser: 9999 # refers to user _flink_ from official flink image, change if necessary
Maybe the issue is the accessMode you set on your PVC. ReadWriteOnce means it can only be mounted by one POD.
See here for Details.
You could try to use ReadWriteMany.
Your generate-tpch-dbgen.sh script is executed while building the docker image resulting those files in /opt/tpch-dbgen/data directory. So, when you run the image, you can see those files.
But the problem with k8s pvc, when you mount the volume (initially empty) to your containers, it replaces the /opt/tpch-dbgen/data directory along with the files in it.
Solution:
Don't execute the generate-tpch-dbgen.sh while building the docker image, rather execute it in the runtime. Then, the files will be created in the shared pv from the init container.
Something like below:
FROM gcc AS builder
RUN mkdir -p /opt
COPY ./generate-tpch-dbgen.sh /opt/generate-tpch-dbgen.sh
RUN chmod +x /opt/generate-tpch-dbgen.sh
ENTRYPOINT ["/bin/sh","/opt/generate-tpch-dbgen.sh"]

How to add "-v /var/run/docker.sock:/var/run/docker.sock" when running container from kubernetes deployment yaml

I'm setting up a kubernetes deployment with an image that will execute docker commands (docker ps etc.).
My yaml looks as the following:
kind: Deployment
apiVersion: apps/v1
metadata:
name: discovery
namespace: kube-system
labels:
discovery-app: kubernetes-discovery
spec:
selector:
matchLabels:
discovery-app: kubernetes-discovery
strategy:
type: Recreate
template:
metadata:
labels:
discovery-app: kubernetes-discovery
spec:
containers:
- image: docker:dind
name: discover
ports:
- containerPort: 8080
name: my-awesome-port
imagePullSecrets:
- name: regcred3
volumes:
- name: some-volume
emptyDir: {}
serviceAccountName: kubernetes-discovery
Normally I will run a docker container as following:
docker run -v /var/run/docker.sock:/var/run/docker.sock docker:dind
Now, kubernetes yaml supports commands and args but for some reason does not support options.
What is the right thing to do?
Perhaps I should configure a volume, but then, is it volumeMount or just a volume?
I am new with kubernetes so it is important for me to do it the right way.
Thank you
You want to add the volume to the container.
spec:
containers:
- name: discover
image: docker:dind
volumeMounts:
- name: dockersock
mountPath: "/var/run/docker.sock"
volumes:
- name: dockersock
hostPath:
path: /var/run/docker.sock
It seems like a bad idea to interact directly with containers on any nodes in Kubernetes. The whole point of Kubernetes is to orchestrate. If you add containers outside of the Pod construct, then Kubernetes will not be aware the processes running on the nodes. This will affect resource allocation.
It also needs to be said that directly working with containers bypasses security.

Store/Share data with a container in Kubernetes

I've dockerized a python project that requires the use of several CSVs (~2gb). In order to keep image size down I didn't include the CSVs in the build, instead opting to give the running container the data from a directory outside the container through a volume. Locally, when running through docker, I can just do
docker run -v ~/local/path/:/container/path my-image:latest
This works, but I'm not sure how to go about doing this in Kubernetes. I've been reading the documentation and am confused by the number of volume types, where the actual CSVs should be stored, etc.
Based on the information about the project that I've provided, is there an obvious solution?
If you'd like to replicate that exact behavior from Docker the most common way to do it is to use hostPath. Something like this:
apiVersion: v1
kind: Pod
metadata:
name: test-pd
spec:
containers:
- image: my-image:latest
name: my-container
volumeMounts:
- mountPath: /container/path
name: test-volume
volumes:
- name: test-volume
hostPath:
path: /usr/local/path
type: Directory
Here is a typical example of sharing between containers. You can keep your data in a separate container and code in a different container.
https://kubernetes.io/docs/tasks/access-application-cluster/communicate-containers-same-pod-shared-volume/
apiVersion: v1
kind: Pod
metadata:
name: two-containers
spec:
restartPolicy: Never
volumes:
- name: shared-data
emptyDir: {}
containers:
- name: nginx-container
image: nginx
volumeMounts:
- name: shared-data
mountPath: /usr/share/nginx/html
- name: debian-container
image: debian
volumeMounts:
- name: shared-data
mountPath: /pod-data
command: ["/bin/sh"]
args: ["-c", "echo Hello from the debian container > /pod-data/index.html"]
Hope it helps.

Is there any definitive guide on how to pass all the arguments to Docker containers while starting a container through kubernetes?

I want to start a docker container with Kubernetes with the parameter --oom-score-adj .
My kubernetes deployment script looks like this:
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
name: xxx
spec:
template:
metadata:
labels:
app: xxx
spec:
volumes:
- name: some-name
hostPath:
path: /some-path
containers:
- name: xxx-container
image: xxx-image
imagePullPolicy: "IfNotPresent"
securityContext:
privileged: true
command:
- /bin/sh
- -c
args:
- ./rsome-command.sh
volumeMounts:
- name: some-name
mountPath: /some-path
When I inspect the created container, I find --oom-score-adj is set to 1000. I want to set it to 0. Can anyone shed any line on how can I do it? Is there any definitive guide to pass such arguments?
You can't do this yet, it's one of the frustrating things still unresolved with Kubernetes.
There's a similar issue here around logging drivers. Unfortunately, you'll have to set the value on the docker daemon

Resources