Charmed PostgreSQL K8s
- Canonical
- Databases
Channel | Revision | Published | Runs on |
---|---|---|---|
latest/stable | 20 | 20 Sep 2022 | |
14/stable | 445 | 12 Nov 2024 | |
14/stable | 444 | 12 Nov 2024 | |
14/candidate | 463 | 19 Nov 2024 | |
14/candidate | 462 | 19 Nov 2024 | |
14/beta | 463 | 19 Nov 2024 | |
14/beta | 462 | 19 Nov 2024 | |
14/edge | 465 | 20 Nov 2024 | |
14/edge | 464 | 20 Nov 2024 |
juju deploy postgresql-k8s --channel 14/edge
Deploy Kubernetes operators easily with Juju, the Universal Operator Lifecycle Manager. Need a Kubernetes cluster? Install MicroK8s to create a full CNCF-certified Kubernetes system in under 60 seconds.
Platform:
Deploy on multiple availability zones (AZ)
During the deployment to Kubernetes, it is important to spread all the database copies (K8s pods/Juju units) to different hardware servers, or even better, to the different cloud availability zones (AZ). This will guarantee no shared service-critical components across the DB cluster (eliminate the case with all eggs in the same basket).
This guide will take you through deploying a PostgreSQL cluster on GKE using 3 available zones. All pods will be set up to sit in their dedicated zones only, which effectively guarantees database copy survival across all available AZs.
This documentation assumes that your cloud supports and provides availability zones concepts such as the K8s label topology.kubernetes.io/zone
. This is enabled by default on EKS/GKE/AKS and supported by MicroK8s/Charmed Kubernetes.
See the Additional resources section for more details about AZ on specific clouds.
Summary
Set up Kubernetes on Google Cloud
Let’s deploy the PostgreSQL Cluster on GKE (us-east4) using all 3 zones there (us-east4-a
, us-east4-b
, us-east4-c
) and make sure all pods always sits in the dedicated zones only.
Warning: Creating the following GKE resources may cost you money - be sure to monitor your GCloud costs.
Log into Google Cloud and bootstrap nine nodes of managed K8s on GCloud:
gcloud auth login
gcloud container clusters create --zone us-east4 $USER-$RANDOM --cluster-version 1.29.8 --machine-type n1-standard-4 --num-nodes=2 --total-max-nodes=2 --no-enable-autoupgrade
kubectl config get-contexts
/snap/juju/current/bin/juju add-k8s gke --client --context-name=$(kubectl config get-contexts | grep gke | awk '{print $2}')
juju bootstrap gke gke
juju add-model mymodel
Each node is equally grouped across availability zones (three K8s nodes in each AZ). You can check them with the following kubectl
commands:
kubectl get nodes
> NAME STATUS ROLES AGE VERSION
> gke-default-pool-5c034921-fmtj Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-5c034921-t1sx Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-5c034921-zkzm Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-b33634ac-l0c9 Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-b33634ac-phjx Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-b33634ac-w2jv Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-d196956a-0zfc Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-d196956a-591j Ready <none> 3h56m v1.29.8-gke.1278000
> gke-default-pool-d196956a-zm6h Ready <none> 3h56m v1.29.8-gke.1278000
kubectl get nodes --show-labels | awk 'NR == 1 {next} {print $1,$2,$6}' | awk -F "[ /]" '{print $1" \t"$NF" \t"$2}'
gke-default-pool-5c034921-fmtj zone=us-east4-b Ready
gke-default-pool-5c034921-t1sx zone=us-east4-b Ready
gke-default-pool-5c034921-zkzm zone=us-east4-b Ready
gke-default-pool-b33634ac-l0c9 zone=us-east4-c Ready
gke-default-pool-b33634ac-phjx zone=us-east4-c Ready
gke-default-pool-b33634ac-w2jv zone=us-east4-c Ready
gke-default-pool-d196956a-0zfc zone=us-east4-a Ready
gke-default-pool-d196956a-591j zone=us-east4-a Ready
gke-default-pool-d196956a-zm6h zone=us-east4-a Ready
Deploy PostgreSQL with anti-affinity rules
Juju provides the support for affinity/anti-affinity rules using constraints. Read more about it in this forum post.
The command below demonstrates how to deploy Charmed PostgreSQL K8s with Juju constraints that create a pod anti-affinity rule:
export MYAPP="mydatabase" ; \
juju deploy postgresql-k8s ${MYAPP} --trust -n 3 \
--constraints="tags=anti-pod.app.kubernetes.io/name=${MYAPP},anti-pod.topology-key=topology.kubernetes.io/zone"
This will effectively create a K8s pod anti-affinity rule. Check with the following command:
kubectl get pod mydatabase-0 -o yaml -n mymodel
...
spec:
affinity:
podAntiAffinity:
requiredDuringSchedulingIgnoredDuringExecution:
- labelSelector:
matchExpressions:
- key: app.kubernetes.io/name
operator: In
values:
- mydatabase
topologyKey: topology.kubernetes.io/zone
...
This example instructs the K8s scheduler to run K8s pods that match the Juju application name app.kubernetes.io/name: my database
label on K8s nodes that have different values of topology.kubernetes.io/zone
label. In other words, we asked to run all PostgreSQL instances in different Availability Zones, which is the common recommendation for any database deployment.
The selector and affinity/anti-affinity rules are extremely flexible and often require cloud-specific fine-tuning for your specific needs. The rule of thumb is to always check K8s node labels actually available in your environment and choose the appropriate label on your infrastructure.
The example below shows all available labels on GKE. You may want to run database instances on different virtual machines (kubernetes.io/hostname
label) or in different Availability Zones (topology.kubernetes.io/zone
label) :
kubectl get node gke-default-pool-b33634ac-l0c9 -o yaml
...
labels:
beta.kubernetes.io/arch: amd64
beta.kubernetes.io/instance-type: n1-standard-4
beta.kubernetes.io/os: linux
cloud.google.com/gke-boot-disk: pd-balanced
cloud.google.com/gke-container-runtime: containerd
cloud.google.com/gke-cpu-scaling-level: "4"
cloud.google.com/gke-logging-variant: DEFAULT
cloud.google.com/gke-max-pods-per-node: "110"
cloud.google.com/gke-nodepool: default-pool
cloud.google.com/gke-os-distribution: cos
cloud.google.com/gke-provisioning: standard
cloud.google.com/gke-stack-type: IPV4
cloud.google.com/machine-family: n1
cloud.google.com/private-node: "false"
failure-domain.beta.kubernetes.io/region: us-east4
failure-domain.beta.kubernetes.io/zone: us-east4-c
kubernetes.io/arch: amd64
kubernetes.io/hostname: gke-default-pool-b33634ac-l0c9
kubernetes.io/os: linux
node.kubernetes.io/instance-type: n1-standard-4
topology.gke.io/zone: us-east4-c
topology.kubernetes.io/region: us-east4
topology.kubernetes.io/zone: us-east4-c
...
After a successful deployment, juju status
will show an active application:
Model Controller Cloud/Region Version SLA Timestamp
mymodel gke gke/us-east4 3.5.3 unsupported 22:02:32+02:00
App Version Status Scale Charm Channel Rev Address Exposed Message
mydatabase 14.13 active 3 postgresql-k8s 14/stable 382 34.118.235.169 no
Unit Workload Agent Address Ports Message
mydatabase/0 active idle 10.80.5.9
mydatabase/1* active idle 10.80.6.7 Primary
mydatabase/2 active idle 10.80.1.6
and each pod will sit in the separate AZ out of the box:
kubectl get pods -n mymodel -o wide
> NAME READY STATUS RESTARTS AGE IP NODE
> mydatabase-0 2/2 Running 0 16m 10.80.5.7 gke-default-pool-b33634ac-l0c9 ... # us-east4-c
> mydatabase-1 2/2 Running 0 24m 10.80.6.7 gke-default-pool-d196956a-0zfc ... # us-east4-a
> mydatabase-2 2/2 Running 0 24m 10.80.0.3 gke-default-pool-5c034921-zkzm ... # us-east4-b
> ...
Simulation: A node gets drained
Let’s drain a node and make sure the rescheduled pod will stay the same AZ:
kubectl drain --ignore-daemonsets --delete-emptydir-data gke-default-pool-b33634ac-l0c9
> node/gke-default-pool-b33634ac-l0c9 cordoned
> ...
> evicting pod mymodel/mydatabase-0
> ...
> pod/mydatabase-0 evicted
> node/gke-default-pool-b33634ac-l0c9 drained
As we can see the newly rescheduled pod landed the new node in the same AZ us-east4-c
:
kubectl get pods -n mymodel -o wide
> NAME READY STATUS RESTARTS AGE IP NODE
> mydatabase-0 2/2 Running 0 1m 10.80.5.7 gke-default-pool-b33634ac-phjx ... # us-east4-c
> mydatabase-1 2/2 Running 0 35m 10.80.6.7 gke-default-pool-d196956a-0zfc ... # us-east4-a
> mydatabase-2 2/2 Running 0 35m 10.80.0.3 gke-default-pool-5c034921-zkzm ... # us-east4-b
> ...
Simulation: All nodes get cordoned
In case we lose (cordon) all nodes in AZ, the pod will stay pending as K8s scheduler cannot find the proper node. Let’s simulate it:
kubectl drain --ignore-daemonsets --delete-emptydir-data gke-default-pool-b33634ac-phjx
kubectl drain --ignore-daemonsets --delete-emptydir-data gke-default-pool-b33634ac-w2jv
kubectl get nodes --show-labels | awk 'NR == 1 {next} {print $1,$2,$6}' | awk -F "[ /]" '{print $1" \t"$NF" \t"$2}'
> gke-default-pool-5c034921-fmtj zone=us-east4-b Ready
> gke-default-pool-5c034921-t1sx zone=us-east4-b Ready
> gke-default-pool-5c034921-zkzm zone=us-east4-b Ready
> gke-default-pool-b33634ac-l0c9 zone=us-east4-c Ready,SchedulingDisabled
> gke-default-pool-b33634ac-phjx zone=us-east4-c Ready,SchedulingDisabled
> gke-default-pool-b33634ac-w2jv zone=us-east4-c Ready,SchedulingDisabled
> gke-default-pool-d196956a-0zfc zone=us-east4-a Ready
> gke-default-pool-d196956a-591j zone=us-east4-a Ready
> gke-default-pool-d196956a-zm6h zone=us-east4-a Ready
kubectl get pods -n mymodel
> NAME READY STATUS RESTARTS AGE
> mydatabase-0 0/2 Pending 0 2m9s # Pending!!!
> mydatabase-1 2/2 Running 0 96m
> mydatabase-2 2/2 Running 0 51m
kubectl describe pod mydatabase-0 -n mymodel | tail -10
> Events:
> Type Reason Age From Message
> ---- ------ ---- ---- -------
> Warning FailedScheduling 3m32s default-scheduler 0/9 nodes are available: 3 node(s) were unschedulable, 6 node(s) had volume node affinity conflict. preemption: 0/9 nodes are available: 9 Preemption is not helpful for scheduling.
> Warning FailedScheduling 3m30s default-scheduler 0/9 nodes are available: 3 node(s) were unschedulable, 6 node(s) had volume node affinity conflict. preemption: 0/9 nodes are available: 9 Preemption is not helpful for scheduling.
> Warning FailedScheduling 3m27s default-scheduler 0/9 nodes are available: 3 node(s) were unschedulable, 6 node(s) had volume node affinity conflict. preemption: 0/9 nodes are available: 9 Preemption is not helpful for scheduling.
> Normal NotTriggerScaleUp 3m33s cluster-autoscaler pod didn't trigger scale-up:
The juju status
output will indicate this problem as well:
Model Controller Cloud/Region Version SLA Timestamp
mymodel gke gke/us-east4 3.5.3 unsupported 22:31:00+02:00
App Version Status Scale Charm Channel Rev Address Exposed Message
mydatabase 14.13 waiting 2/3 postgresql-k8s 14/stable 382 34.118.235.169 no installing agent
Unit Workload Agent Address Ports Message
mydatabase/0 unknown lost agent lost, see 'juju show-status-log mydatabase/0'
mydatabase/1* active idle 10.80.6.7 Primary
mydatabase/2 active idle 10.80.1.6
Let’s uncordon all nodes to keep the house clean:
kubectl uncordon gke-default-pool-b33634ac-l0c9
kubectl uncordon gke-default-pool-b33634ac-phjx
kubectl uncordon gke-default-pool-b33634ac-w2jv
The K8s scheduler will return the pod back to AZ us-east4-c
and Juju will automatically rejoin the database unit back to the cluster:
Model Controller Cloud/Region Version SLA Timestamp
mymodel gke gke/us-east4 3.5.3 unsupported 22:38:23+02:00
App Version Status Scale Charm Channel Rev Address Exposed Message
mydatabase 14.13 active 3 postgresql-k8s 14/stable 382 34.118.235.169 no
Unit Workload Agent Address Ports Message
mydatabase/0 active idle 10.80.5.10
mydatabase/1* active idle 10.80.6.7 Primary
mydatabase/2 active idle 10.80.1.6
At this point we can relax and enjoy the protection from Cloud Availability zones!
To survive acomplete cloud outage, we recommend setting up cluster-cluster asynchronous replication.
Remove GKE setup
Warning: Do not forget to remove your GKE test setup - it can be costly!
gcloud container clusters list
gcloud container clusters delete <gke_name> --location <gke_location>
juju unregister gke --no-prompt
juju remove-cloud gke
Additional resources
Below you will find specific information about AZs on specific clouds and more about node selection on Kubernetes.