How to make glusterfs survive cluster upgrade - google-cloud-platform

I'm trying to use glusterfs installed directly on my GCE cluster nodes.
The installation does not persist through cluster upgrades, which could be solved with a bootstrap script.
The problem is that when I did reinstall the glusterfs manually and mounted the brick, there was no volumes present, which I had to force recreate.
What happened? Does glusterfs store volume data somewhere else than on bricks? How do I prevent this?

Can I confirm you are doing this on a Kubernetes cluster? I presume you are as you mentioned cluster upgrades.
If so, when you say gluster was installed directly on your nodes, I'm not sure I understand that part of your post. My understanding of the intended use of glusterfs is that it's exists as a distributed file system, and the storage is therefore part of a separate cluster to the Kubernetes nodes.
I believe this is the recommended method to use glusterfs with Kubernetes, and this way the data in the volumes will be retained after the Kubernetes cluster upgrade.
Here are the steps I performed.
I created the glusterfs cluster using the information/script from the first three steps in this this tutorial (specially the 'Clone' 'Bootstrap your Cluster' and 'Create your first volume' steps). In terms of the YAML below, It may be useful to know my glusterfs volume was named 'glustervolume'.
Once I'd confirmed the gluster volume had been created, I created Kubernetes and service and end points that point at that volume. The IP addresses in the the end point section of the YAML below are the internal IP addresses of the instances in the glusterfs storage cluster.
---
apiVersion: v1
kind: Service
metadata:
name: glusterfs-cluster
spec:
ports:
- port: 1
---
apiVersion: v1
kind: Endpoints
metadata:
name: glusterfs-cluster
subsets:
- addresses:
- ip: 10.132.0.6
ports:
- port: 1
- addresses:
- ip: 10.132.0.7
ports:
- port: 1
- addresses:
- ip: 10.132.0.8
ports:
- port: 1
I then created a pod to make use of the gluster volume:
---
apiVersion: v1
kind: Pod
metadata:
name: glusterfs
spec:
containers:
- name: glusterfs
image: nginx
volumeMounts:
- mountPath: "/mnt/glusterfs"
name: glustervolume
volumes:
- name: glustervolume
glusterfs:
endpoints: glusterfs-cluster
path: glustervolume
readOnly: false
As the glusterfs volume exists separately to the Kubernetes cluster (i.e. on it's own cluster), Kubernetes upgrades will not affect the volume.

Related

Kubernetes Deployment Persistent Volume and VM disk size

After creating a NFS Persistent Volume for one of Deployments running in a cluster the containers are able to store and share the file data between each other. The file data is persistent between the containers life cycles too. And that's great! But I wonder where exactly is this file data stored: where is it "physically" located? Is it saved onto the container itself or is it saved somewhere onto a VM's disk - the VM that is used to run the Deployment?
The VM that is used to host the Deployment has only 20 Gb available disk space by default. Let's say I am running a Docker container inside a pod on a Node (aka VM) running some file server. What happens if I attempt to transfer a 100 Gb file to that File Server? Where will be this gigantic file saved if the VM disk itself has only 20 Gb available space?
Edited later by appending the portion of yaml file:
apiVersion: v1
kind: PersistentVolumeClaim
metadata:
name: pv-claim
labels:
app: deployment
spec:
accessModes:
- ReadWriteOnce
resources:
requests:
storage: 20Gi
# ---
apiVersion: apps/v1
kind: Deployment
metadata:
name: deployment
spec:
selector:
matchLabels:
app: app
replicas: 1
minReadySeconds: 10
strategy:
type: RollingUpdate # Recreate
rollingUpdate:
maxUnavailable: 1
maxSurge: 1
template:
metadata:
labels:
app: app
spec:
containers:
- name: container
image: 12345.dkr.ecr.us-west-2.amazonaws.com/container:v001
ports:
- containerPort: 80
imagePullPolicy: IfNotPresent
volumeMounts:
- name: volume-mount
mountPath: /data
volumes:
- name: volume-mount
persistentVolumeClaim:
claimName: pv-claim
The "physical" location of the volume is defined by the provisioner, which is defined by the storage class. Your PV claim doesn't have a storage class assigned. That means that the default storage class is used, and it can be anything. I suspect that in EKS default storage class will be EBS, but you should double check that.
First, see what storage class is actually assigned to your persistent volumes:
kubectl get pv -o wide
Then see what provisioner is assigned to that storage class:
kubectl get storageclass
Most likely you will see something like kubernetes.io/aws-ebs. Then google documentation for a specific provisioner to understand where the volume is stored "physically".
In your case the data is stored on NFS share. Connect to NFS server and browse through the shares and find the share that is mounted to the pod.

Deploying Django Application with PostgreSQL to Kubernetes Google Cloud cluster

I am having trouble trying to deploy my Django Application and PostgreSQL database to Kubernetes Google Cloud cluster that I've already configured.
I have successfully created Docker containers for my Django Application and PostgreSQL database. Here is what my docker-compose.yml file looks like:
version: '3'
services:
db:
image: postgres
environment:
- POSTGRES_USER=stefan_radonjic
- POSTGRES_PASSWORD=cepajecar995
- POSTGRES_DB=agent_technologies_db
web:
build: .
command: python manage.py runserver 0.0.0.0:8000 --settings=agents.config.docker-settings
volumes:
- .:/agent-technologies
ports:
- "8000:8000"
links:
- db
depends_on:
- db
I have already build the images, and tried sudo docker-compose up command, and the application works perfectly fine.
After successfully dockerizing Django Application and PostgreSQL, I have tried to configure Deployment / Service YML files required by Kubernetes, but I am having trouble doing so. For example:
deployment-definition.yml - File for deploying Django application:
apiVersion: apps/v1
kind: Deployment
metadata:
name: agent-technologies-deployment
labels:
app: agent-technologies
tier: backend
spec:
template:
metadata:
name: agent-technologies-pod
labels:
app: agent-technologies
tier: backend
spec:
containers:
- name:
image:
ports:
- containerPort: 8000
replicas:
selector:
matchLabels:
tier: backend
Inside container list of dictionaries, I know that my container name should be web, but I am not sure where the image of that container is located so I do not know what should i specify as container image.
Another problem lies in postgres/deployment-definition.yml :
apiVersion: apps/v1
kind: Deployment
metadata:
name: postgres
spec:
replicas: 1
selector:
matchLabels:
app: postgres-container
template:
metadata:
labels:
app: postgres-container
tier: backend
spec:
containers:
- name: postgres-container
image: postgres:9.6.6
env:
- name: POSTGRES_USER
valueFrom:
secretKeyRef:
name: postgres-credentials
key: user
- name: POSTGRES_PASSWORD
valueFrom:
secretKeyRef:
name: postgres-credentials
key: password
- name: POSTGRES_DB
value: agent_technologies_db
ports:
- containerPort: 5432
volumeMounts:
- name: postgres-volume-mount
mountPath: /var/lib/postgresql/data
volumes:
- name: postgres-volume-mount
persistentVolumeClaim:
claimName: postgres-pvc
I do not understand what volumeMounts and volumes are for, and if i even specified them correctly.
Here is my secret-definition.yml file:
apiVersion: v1
kind: Secret
metadata:
name: postgres-credentials
type: Opaque
data:
user: stefan_radonjic
passowrd: cepajecar995
My postgres/service-definition.yml file:
apiVersion: v1
kind: Service
metadata:
name: postgres-service
spec:
selector:
app: postgres-container
ports:
- protocol: TCP
port: 5432
targetPort: 5432
My postgres/volume-definition.yml file:
apiVersion: v1
kind: PersistentVolume
metadata:
name: postgres-pv
labels:
type: local
spec:
capacity:
storage: 2Gi
storageClassName: standard
accessModes:
- ReadWriteMany
hostPath:
path: /data/postgres-pv
And my postgres/volume-claim-definitono.yml file:
apiVersion: v1
kind: PersistentVolume
metadata:
name: postgres-pv
labels:
type: local
spec:
capacity:
storage: 2Gi
storageClassName: standard
accessModes:
- ReadWriteMany
hostPath:
path: /data/postgres-pv
Last but not least, my service-definition.yml file - for Django application
apiVersion: v1
kind: Service
metadata:
name: agent-technologies-service
spec:
selector:
app: agent-technologies
ports:
- protocol: TCP
port: 8000
targetPort: 8000
type: NodePort
Besides the questions I have already asked above, I also want to ask am I doing this right? If not, what can I do to fix this.
Inside container list of dictionaries, I know that my container name should be web, but I am not sure where the image of that container is located so I do not know what should i specify as container image.
Name for container is local to the pod (you can have several containers sharing same pod). Container name (web in your case) is for your files given under deployment:
# setting name of first container within pod to web
spec:
containers:
- name: web
Image for container has to be in some available docker container registry. There are multiple options from hosting own docker registry to use publicly available ones. In any case you have to be able to push in your build phase to that docker container registry (be it amazon ECR, Docker, Gitlab, self hosted...) and to pull from that registry from within kubernetes (security settings, pull secrets etc...). In your docker-compose file you use two containers. For db you use public postgres image, and for web you use build command and image is stored to local docker registry on that host only (you have to push it to public repository for k8s to be able to pull from it during deployment).
I do not understand what volumeMounts and volumes are for, and if i even specified them correctly.
In a nutshell, volumes are for attaching volumes to containers. Depending on your use case and decided architecture there are several approaches to volumes, but all in all they boil down to ephemeral, constant and persistent. Ephemeral will be lost on container termination or restart, constant (such as from configMaps) are used for passing configuration files to containers and persistent are most interesting for stateful applications (databases among other things). You can specify volumes in several ways, all volume have to have name (to be referenced by volumeMount) and either direct volume specification or volume claim specification (latter is advised for persistent volume since you can benefit from automatic provisioning that way).
VolumeMounts are for defining where on container file system predefined volume should be mounted. They reference volume to be mounted by name, provide mount point on container filesystem by mountPath and can have subpaths to specific files in some cases.
In your example you tied persistent volume claim obtained volume to data path of postgres (/var/lib/postgresql/data). Althought you use storage class that you didn't specify, interesting part is that your Persistent volume is defined as localpath on host. That means that on each node you have this database pod started you will end up pointing /var/lib/postgresql/data of that pod's db container to /data/postgres-pv on that specific node. This opens up you to following issue: say you have 3 nodes (A, B and C) and your database pod is started on A, uses A's /data/postgres-pv folder as own /var/lib/postrgresql/data. And then you restart it, it gets terminated and rescheduled to node B. All of the sudden, it uses B's /data/postgres-pv local folder (empty) and you end up with empty database. If you use host's local filesystem for persisntence you need to tie such pods with node (or better yet with affinity) selectors. It is advisable for performance reasons to run database volumes of local filesystem, but hose pods lose ability to be rescheduled easily. Another approach is to have some truly persistent volume that can be mounted independently of node (Amazon EBS for example) and they require different PVC (or provisioner to be used).
Besides the questions I have already asked above, I also want to ask am I doing this right? If not, what can I do to fix this.
As stated above, define storage class and either lock db pod to specific node or apply some kind of dynamic provisioning so volume will follow pod's placement on nods.
Oppiniated preference: don't place everything in default namespace, use separate namespace for handling k8s manifests, later on it is much harder to move everything, and harder to accidentally delete wrong thingie...
Also personal preference: database is stateful application and as such it is advised to use statefulset instead of deployment.
There are tools to help you out when you start from docker-compose files and want to convert to kubernetes manifests, worth checking.
Documentation on kubernetes is a bit outdated but quite good and you can have some nice read on volumes and volumeClaims there, there is also active slack channel.
Oh, and mock user/pass when posting files here, we know now about cepa...
Lastly, you are doing great job!

Create specific A record entry to Kubernetes local DNS

I have a Kubernetes v1.4 cluster running in AWS with auto-scaling nodes.
I also have a Mongo Replica Set cluster with SSL-only connections (FQDN common-name) and public DNS entries:
node1.mongo.example.com -> 1.1.1.1
node2.mongo.example.com -> 1.1.1.2
node3.mongo.example.com -> 1.1.1.3
The Kubernetes nodes are part of a security group that allows access to the mongo cluster, but only via their private IPs.
Is there a way of creating A records in the Kubernetes DNS with the private IPs when the public FQDN is queried?
The first thing I tried was a script & ConfigMap combination to update /etc/hosts on startup (ref. Is it a way to add arbitrary record to kube-dns?), but that is problematic as other Kubernetes services may also update the hosts file at different times.
I also tried a Services & Enpoints configuration:
---
apiVersion: v1
kind: Service
metadata:
name: node1.mongo.example.com
spec:
ports:
- protocol: TCP
port: 27017
targetPort: 27017
---
apiVersion: v1
kind: Endpoints
metadata:
name: node1.mongo.example.com
subsets:
- addresses:
- ip: 192.168.0.1
ports:
- port: 27017
But this fails as the Service name cannot be a FQDN...
While not so obvious at first, the solution is quite simple. kube-dns image in recent versions includes dnsmasq as one of it's components. If you look into its man page, you will see some usefull options. Following that lecture you can choose a path similar to this :
Create a ConfigMap to store your dns mappings :
apiVersion: v1
kind: ConfigMap
metadata:
name: kube-dns
namespace: kube-system
data:
myhosts: |
10.0.0.1 foo.bar.baz
Having that ConfigMap applied in your cluster you can now make some changes to kube-dns-vXX deployment you use in your kubernetes.
Define volume that will expose your CM to dnsmasq
volumes:
- name: hosts
configMap:
name: kube-dns
and mount is in your dnsmasq container of kube-dns deployment/rc template
volumeMounts:
- name: hosts
mountPath: /etc/hosts.d
and finally, add a small config flag to your dnsmasq arguments :
args:
- --hostsdir=/etc/hosts.d
now, as you apply these changes to the kube-dns-vXX deployment in your cluster it will mount the configmap and use files mounted in /etc/hosts.d/ (with typical hosts file format) as a source of knowledge for dnsmasq. Hence if you now query for foo.bar.baz in your pods, they will resolve to respective IP. These entries take precedence over public DNS, so it should perfectly fit your case.
Mind that dnsmasq is not watching for changes in ConfigMap so it has to be restarted manually if it changes.
Tested and validated this on a live cluster just few minutes ago.

Mount kubernetes' volume in development machine

I have a Kubernetes cluster running on Google Cloud Platform. I have 3 nodes and several pods running on these nodes.
One of the pods runs Ghost blog platform and has mounted a gcePersistentDisk volume. The manifest file to create the pod:
apiVersion: extensions/v1beta1
kind: Deployment
metadata:
labels:
name: ghost
name: ghost
spec:
replicas: 1
template:
metadata:
labels:
name: ghost
spec:
containers:
- image: ghost:0.7
name: ghost
env:
- name: NODE_ENV
value: production
ports:
- containerPort: 2368
name: http-server
volumeMounts:
- name: ghost
mountPath: /var/lib/ghost
volumes:
- name: ghost
gcePersistentDisk:
pdName: ghost
fsType: ext4
I'd like someway to access this volume from my development machine. Is there any way to mount this disk in my machine?
If your development machine is not part of the GCE cluster (i.e. a GCE VM), then you will not be able to directly mount it. Your best bet in that case would be to SSH to it via a machine it is mounted it (i.e the node your pod is scheduled to).

Kubernetes: How can I set the number of replicas more than 1 using awsElasticBlockStore?

I'm trying to create a Cassandra cluster in Kubernetes. I want to use awsElasticBlockStore to make the data persistent. As a result, I've written a YAML file like following for the corresponding Replication Controller:
apiVersion: v1
kind: ReplicationController
metadata:
name: cassandra-rc
spec:
# Question: How can I do this?
replicas: 2
selector:
name: cassandra
template:
metadata:
labels:
name: cassandra
spec:
containers:
- resources:
limits :
cpu: 1.0
image: cassandra:2.2.6
name: cassandra
ports:
- containerPort: 7000
name: comm
- containerPort: 9042
name: cql
- containerPort: 9160
name: thrift
volumeMounts:
- name: cassandra-persistent-storage
mountPath: /cassandra_data
volumes:
- name: cassandra-persistent-storage
awsElasticBlockStore:
volumeID: aws://ap-northeast-1c/vol-xxxxxxxx
fsType: ext4
However, only one pod can be properly launched with this configuration.
$ kubectl get pods
NAME READY STATUS RESTARTS AGE
cassandra-rc-xxxxx 0/1 ContainerCreating 0 5m
cassandra-rc-yyyyy 1/1 Running 0 5m
When I run $ kubectl describe pod cassandra-rc-xxxxx, I see an error like following:
Error syncing pod, skipping: Could not attach EBS Disk "aws://ap-northeast-1c/vol-xxxxxxxx": Error attaching EBS volume: VolumeInUse: vol-xxxxxxxx is already attached to an instance
It's understandable because an ELB Volume can be mounted from only one node. So only one pod can successfully mount the volume and bootup, while others just fail.
Is there any good solution for this? Do I need to create multiple Replication Controllers for each pod?
You are correct, one EBS volume can only be mounted on a single EC2 at a given time. To solve you have the following options:
Use multiple EBS volumes with multiple Replication Controllers
Use a distributed file system (e.g. Gluster) and avoid EBS issue
Follow along with PetSet (https://github.com/kubernetes/kubernetes/issues/260)