Over the past two years at Magalix, we have focused on building our system, introducing new features, and scaling our infrastructure and microservices. During this time, we had a look at our Kubernetes clusters utilization and found it to be very low. We were paying for resources we didn’t use, so we started a cost-saving practice to increase cluster utilization, use the resources we already had and pay less to run our cluster. In this article, I will discuss the top five techniques we used to better utilize our Kubernetes clusters on the cloud and eliminate wasted resources, thus saving money. In the end, we were able to cut our monthly bill by more than 50%! 1. Applying Workload Right-Sizing Kubernetes manages and schedules pods are based on container resource specs: : Kubernetes scheduler is used to place containers on the right node which has enough capacity Resource Requests : Containers are NOT allowed to use more than their resource limit Resource Limits Resources requests and limits are container-scooped specs, while multi-container pods define separate resource specs for each container: : apps/v1 : Deployment : : nginx : magalix : : : : nginx : : : : : nginx : : - : nginx : : . : - : : : : m : Mi : : : Gi apiVersion kind metadata name namespace spec selector matchLabels app replicas 2 template metadata labels app spec containers name image nginx 1.14 2 ports containerPort 80 resources requests cpu 100 memory 100 limits cpu 1 memory 1 Kubernetes schedules pods based on resource requests and other restrictions without impairing availability. The scheduler uses CPU and memory resource requests to schedule the workloads in the right nodes, control which pod works on which node and if multiple pods can schedule together on a single node. Every node type has its own allocatable CPU and memory capacities. Assigning high/unneeded CPU or memory resource requests can end up running underutilized pods on each node, which leads to underutilized nodes. In this section, we compared resource requests, limited against actual usage and changed the resource request to something closer to the actual utilization while adding a little safety margin. 2. Choosing The Right Worker Nodes Every Kubernetes cluster has its own special workload utilization. Some clusters use memory more than CPU (e.g: database and caching workloads), while others use CPU more than memory (e.g: user-interactive and batch-processing workloads) Cloud providers such as GCP and AWS offer various node types that you can choose from. Choosing the wrong node size for your cluster can end up costing you. For instance, choosing high CPU-to-memory ratio nodes for workloads that use memory extensively can starve for memory easily and trigger auto node scale-up, wasting more CPUs that we don’t need. Calculating the right ratio of CPU-to-memory isn’t easy; you will need to monitor and know your workloads well. For example, GCP offers general purpose, compute-optimized, memory-optimized with various CPU and memory count and ratios: Just keep in mind that 1 vCPU is way more expensive than 1GB memory. I have enough memory in the clusters I manage so I try to make sure that when there is a pending pod, this pod is pending on CPUs (which is expensive) so the autoscaler triggers a scale-up for the new node. To see the cost difference between CPU and memory, let us look at the GCP N2 machine price. GCP gives you the freedom to choose a custom machine type: (# vCPU x vCPU price) + (# GB memory x GB memory price) 1 1 It’s clear here that the 1vCPU costs 7.44 times more than the cost of 1GB. 3. Autoscaling Workloads Autoscaling is great because it helps to scale up/down your workloads and shut down nodes to save you money while you sleep. Many cases can benefit from autoscaling: 1. Variable load web applications: A good example would be a web application which receives variable traffic through the day: traffic increases during certain hours of the day and decreases during the night. Kubernetes comes with the Horizontal Pod Autoscaler (HPA) that can scale workload replicas based on their CPU and memory utilization ratio to the resource request. Kubernetes will keep monitoring the target resource in scaleTargetRef and will scale up/down the replica count to keep targetCPUUtilizationPercentage around 75%. In this example, the deployment frontend will be scaled up to 20 replicas during the high load, and 4 replicas during the low load. apiVersion: autoscaling/v1 kind: HorizontalPodAutoscaler metadata: name: frontend namespace: magalix spec: maxReplicas: minReplicas: scaleTargetRef: apiVersion: apps/v1 kind: Deployment Name: frontend targetCPUUtilizationPercentage: 20 4 75 2. Event-driven workers: background workers that need to be started in multiple replicas when there are messages in a Kafka topic or a message queue. It can be scaled to zero when there are no messages. Compared to HPA, there is a more advanced Kubernetes Event-driven Autoscaling ( ) that can integrate with Prometheus/PosqreSQL/Kafka/Redis and many more to scale based on more advanced metrics from multiple data sources. In this example, we installed this KEDA ScaledObject custom resource definition to scale the worker deployment eventer replicas. When the Kafka consumer lag changes, it can scale to 0 when there are no messages to consume and can scale up 1 replica for every 10,000 lagged messages up to 8 when the lag is more than 80,000: KEDA apiVersion: keda.k8s.io/v1alpha1 kind: ScaledObject metadata: labels: deploymentName: eventer name: eventer namespace: magalix spec: cooldownPeriod: maxReplicaCount: minReplicaCount: pollingInterval: scaleTargetRef: deploymentName: eventer triggers: - metadata: metricName: eventer query: sum(kafka_consumergroup_lag{consumergroup= }) serverAddress: http: threshold: type: prometheus 10 8 0 15 "eventer-group" //prometheus-server.monitoring.svc.cluster.local "10000" 4. Autoscaling Worker Nodes After scaling workloads, you will notice the number of running pods is low, but this won’t save you money unless we configure auto-scaling the worker nodes Some Could providers provide node autoscaling out of the box on some node pools. can help you manage worker node autoscaling. Cluster Autoscaler GCP: Kubernetes Engine → Cluster → Node Pool AWS: EKS → Cluster → Node Group Azure: Kubernetes services → Node pools → Scale → Automatic The result of scaling workload and worker nodes together can end with the node count trends: 5. Purchasing Commitment/Saving Plans Running a Kubernetes service is relatively cheap. What’s most expensive is the worker nodes compute cost. offers “Commitment Plans” on a certain number of vCPUs, memory, GPUs, and local SSDs for 1 or 3 years. This can save up to 57% of the compute cost GCP offers “Compute Saving Plans” to commit to using a certain amount of money on compute every month, as well as “Reserved Instances” to commit to using a certain type of machine, both for 1 or 3 years with a possible savings of up to 60% AWS offers “Azure Reserved VM Instances” such as AWS with a possible savings of up to 60% Azure Conclusion As we read in this article, there are multiple factors and considerations when trying to reduce your cluster cost. Going through the whole process can give you huge savings. - and you can do the same! We have managed to reduce our cluster daily cost by 56% Also published here.