Scaling MongoDB on Kubernetes
Kubernetes was primarily used for Stateless Applications. However, PetSets were introduced in the version 1.3 and later they evolved to Stateful Sets. The official documentation describes Stateful sets as
StatefulSets are intended to be used with stateful applications and distributed systems.
One of the best use cases for this is to orchaestrate data-store services such as MongoDB, ElasticSearch, Redis, Zookeeper and so on.
Some of the features that can be ascribed to StatefulSets are:
- Pods with Ordinal Indexes
- Stable Network Identities
- Ordered and Parallel Pod Management
- Rolling Updates
Details for these can be found here.
One very distinct feature of Stateful Sets is to provide Stable
Network Identities which when used with Headless Services , can be even more powerful.
Without spending much time on information readily available in Kubernetes documentation, let us focus on running and scaling a MongoDB cluster.
You need to have a running Kubernetes Cluster with RBAC enabled (recommended). In this tutorial I will be using a GKE cluster, however, AWS EKS or Microsoft’s AKS or a Kops Managed K8’s K8’s are also viable alternatives.
We will deploy the following components for our MongoDB cluster
- Daemon Set to configure HostVM
- Service Account and ClusterRole Binding for Mongo Pods
- Storage Class to provision persistent SSDs for the Pods
- Headless Service to access to Mongo Containers
- Mongo Pods Stateful Set
- GCP Internal LB to access MongoDB from outside the kuberntes cluster (Optional)
- Access to pods using Ingress (Optional)
It is important to note that each MongoDB Pod will have a sidecar running, in order to configure the replica set, on the fly. The sidecar checks for new members every 5 seconds.
Daemon Set for HostVM Configuration
Configuration for ServiceAccount, Storage Class, Headless SVC and StatefulSet
- The Sidecar for Mongo should be configured carefully with proper environment variables, stating the labels given to the pod, namespace for the deployment and service. Details about the sidecar container can be found here.
- The guidance around default cache size is: “50% of RAM minus 1 GB, or 256 MB”. Given that the amount of memory requested is 2GB, the WiredTiger cache size here, has been set to 256MB
- Inter-Pod Anti-Affinity ensures that no 2 Mongo Pods are scheduled on the same worker node, thus, making it resilient to node failures. Also, it is recommended to keep the nodes in different AZs so that the cluster is resilient to Zone failures.
- The Sevice Account currently deployed has admin priviledges. However, it should be restricted to the DB’s namespace.
Both configurations mentioned above can also be found here.
Deploying the MongoDB cluster
kubectl apply -f configure-node.yml
kubectl apply -f mongo.yml
You can view the components by
root$ kubectl -n mongo get all
NAME DESIRED CURRENT AGE
statefulsets/mongo 3 3 3m
NAME READY STATUS RESTARTS AGE
po/mongo-0 2/2 Running 0 3m
po/mongo-1 2/2 Running 0 2m
po/mongo-2 2/2 Running 0 1m
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
svc/mongo ClusterIP None <none> 27017/TCP 3m
As you can see, that the service has no Cluster-IP and neither an External-IP, it is a headless service. This service will directly resolve to Pod-IPs for our Stateful Sets.
To verify the DNS resolution. We launch an interactive shell within our cluster
kubectl run my-shell --rm -i --tty --image ubuntu -- bash
root@my-shell-68974bb7f7-cs4l9:/# dig mongo.mongo +search +noall +answer
; <<>> DiG 9.11.3-1ubuntu1.1-Ubuntu <<>> mongo.mongo +search +noall +answer
;; global options: +cmd
mongo.mongo.svc.cluster.local. 30 IN A 10.56.7.10
mongo.mongo.svc.cluster.local. 30 IN A 10.56.8.11
mongo.mongo.svc.cluster.local. 30 IN A 10.56.1.4
The DNS for service will be <name of service>.<namespace of service>, therefore, in our case it will be mongo.mongo .
The IPS( 10.56.6.17, 10.56.7.10, 10.56.8.11 ) are our Mongo Stateful Set’s Pod IPs. This can be tested by running a nslookup over these, from inside the cluster.
root@my-shell-68974bb7f7-cs4l9:/# nslookup 10.56.6.17
126.96.36.199.in-addr.arpa name = mongo-0.mongo.mongo.svc.cluster.local.
root@my-shell-68974bb7f7-cs4l9:/# nslookup 10.56.7.10
10.7.56.10.in-addr.arpa name = mongo-1.mongo.mongo.svc.cluster.local.
root@my-shell-68974bb7f7-cs4l9:/# nslookup 10.56.8.11
188.8.131.52.in-addr.arpa name = mongo-2.mongo.mongo.svc.cluster.local.
If you app is deployed in the K8’s cluster itself, then it can access the nodes by
If you would like to access the mongo nodes from outside the cluster you can deploy internal load balancers for each of these pods or create an internal ingress, using an Ingress Controller such as NGINX or Traefik.
GCP Internal LB SVC Configuration (Optional)
Deploy 2 more such services for mongo-1 and mongo-2.
You can provide the IPs of the Internal Load Balancer to the MongoClient URI.
root$ kubectl -n mongo get svc
NAME TYPE CLUSTER-IP EXTERNAL-IP PORT(S) AGE
mongo ClusterIP None <none> 27017/TCP 15m
mongo-0 LoadBalancer 10.59.252.157 10.20.20.2 27017:30184/TCP 9m
mongo-1 LoadBalancer 10.59.252.235 10.20.20.3 27017:30343/TCP 9m
mongo-2 LoadBalancer 10.59.254.199 10.20.20.4 27017:31298/TCP 9m
The external IPs for mongo-0/1/2 are the IPs of the newly created TCP loadbalancers. These are local to your Subnetwork or peered networks, if any.
Access Pods using Ingress (Optional)
Traffic to Mongo Stateful set pods can also be directed using an Ingress Controller such as Nginx. Make sure the ingress service is internal and not exposed over public ip. The ingress object will look something like this:
- host: mongo.example.com
- path: '/mongo-0'
serviceName: mongo # There is no extra service. This is
servicePort: '27017' # the headless service.
It is important to note that your application is aware of atleast one mongo node which is currently up so that it can discover all the others.
I use Robo 3T as a mongo client on my local mac. After connecting to one of the nodes. and running
rs.status() , you can view the details of the replica set and check if the other 2 pods were configured and connected to the Replica Set automatically.
Now we scale the Stateful Set for mongo Pods to check if the new mongo containers get added to the ReplicaSet or not.
root$ kubectl -n mongo scale statefulsets mongo --replicas=4
statefulset "mongo" scaled
root$ kubectl -n mongo get pods -o wide
NAME READY STATUS RESTARTS AGE IP NODE
mongo-0 2/2 Running 0 25m 10.56.6.17 gke-k8-demo-demo-k8-pool-1-45712bb7-vfqs
mongo-1 2/2 Running 0 24m 10.56.7.10 gke-k8-demo-demo-k8-pool-1-c6901f2e-trv5
mongo-2 2/2 Running 0 23m 10.56.8.11 gke-k8-demo-demo-k8-pool-1-c7622fba-qayt
mongo-3 2/2 Running 0 3m 10.56.1.4 gke-k8-demo-demo-k8-pool-1-85308bb7-89a4
It can be seen that all four pods are deployed to different GKE nodes and thus our Pod-Anti Affinity policies are working correctly.
The scaling action will also automatically provision a persistent volume, which will act as the data directory for the new pod.
root$ kubectl -n mongo get pvc
NAME STATUS VOLUME CAPACITY ACCESS MODES STORAGECLASS AGE
mongo-persistent-storage-mongo-0 Bound pvc-337fb7d6-9f8f-11e8-bcd6-42010a940024 11G RWO fast 49m
mongo-persistent-storage-mongo-1 Bound pvc-53375e31-9f8f-11e8-bcd6-42010a940024 11G RWO fast 49m
mongo-persistent-storage-mongo-2 Bound pvc-6cee0f97-9f8f-11e8-bcd6-42010a940024 11G RWO fast 48m
mongo-persistent-storage-mongo-3 Bound pvc-3e89573f-9f92-11e8-bcd6-42010a940024 11G RWO fast 28m
To check whether the pod named mongo-3 gets added to the replica set or not, we run
rs.status() once again on the same node and observe the difference.
- It can be helpful to label the Node Pool which will be used for Mongo Pods and ensure that appropriate Node Affinity is mentioned in the Spec for the Stateful Set and HostVM configurer Daemon Set . This is because the Daemon set will tweak some parameters of the host OS and those settings should be restricted for MongoDB Pods only. Other applications might work better without those settings.
- Labelling a node pool is extremely easy in GKE, can be directly from the GCP console.
- Although we have specified CPU and Memory limits in the Pod Spec, we can also consider deploying a VPA (Vertical Pod Autoscaler).
- Traffic to our DB from inside the cluster can be controlled by implementing network policies or a service mesh such as Istio.
If you have reached this far, I believe you have gone through the entire blogpost. I have tried to collate a lot of scattered information and present it as one. My aim is to give you enough information in order to get started with Stateful Sets on Kubernetes and hope many of you find it useful. Feedback,
comments or suggestions are more than welcome. :)