Kubernetes is the most popular orchestrator for container deployment and management. It equips you with powerful tools to reliably run containerized apps in production.
With flexibility comes complexity, however. Kubernetes includes its own concepts, terms, and object types for modeling your application. Choosing when to use different components can be confusing to newcomers and experienced users alike because the effects of your decisions aren’t always easy to anticipate.
In this article, we’ll explore 15 common Kubernetes pitfalls which many teams encounter. Being able to recognize and avoid these challenges will improve your app’s scalability, reliability, and security while giving you more control over your cluster and its deployments.
- Deploying Containers With the “Latest” Tag
- Not Using Liveness and Readiness Probes
- Broken Pod Affinity/Anti-Affinity Rules
- Forgetting Network Policies
- No Monitoring/Logging
- Label Selector Mismatches
- Service Port Mismatches
- Using Multiple Load Balancers
- Accidentally Deploying to the Wrong Namespace
- Pods Without Resource Requests and Limits
- Not Budgeting For Failure With PodDisruptionBudgets
- Incorrect Cluster Size and Faulty Auto-Scaling
- Inefficient Scheduling Due to Missing Node Selectors
- Relying on the Standard Tools
- Not Using Pod Security Admission Standards
Arguably one of the most frequently violated Kubernetes best practices is using the
latest tag when you deploy containers. This puts you at risk of unintentionally receiving major changes which could break your deployments.
latest tag is used in different ways by individual authors, but most will point
latest to the newest release of their project. Using
helm:latest today will deliver Helm v3, for example, but it’ll immediately update to v4 after that release is launched.
When you use
latest, the actual versions of the images in your cluster are unpredictable and subject to change. Kubernetes will always pull the image when a new Pod is started, even if a version is already available on the host Node. This differs from other tags, where the existing image on the Node will be reused when it exists.
Probes make your applications more resilient. They inform Kubernetes of the health of your Pods.
Liveness probes instruct Kubernetes when it should restart a container because a problem has occurred. It allows malfunctioning containers to be replaced in circumstances where they’re broken but haven’t stopped themselves. Readiness probes indicate when a container is ready to begin accepting traffic from a service, preventing failures that occur during application startup.
Probes are easy to configure, but they’re often forgotten.
Here’s a simple Pod with both liveness and readiness probes:
apiVersion: v1 kind: Pod metadata: name: probes-demo spec: containers: - name: probes-demo image: nginx:latest livenessProbe: httpGet: path: / port: 80 readinessProbe: httpGet: path: / port: 80
Several different probe types are supported, including HTTP (shown here), TCP, gRPC, and command execution.
Pod affinity and anti-affinity rules allow you to instruct Kubernetes which Node is the best match for new Pods. Rules can be conditioned on Node-level characteristics such as labels, or characteristics of the other Pods already running on the Node.
Affinity rules attract Pods to Nodes, making it more likely that a Pod will schedule to a particular Node, whereas anti-affinity has a repelling effect which reduces the probability of scheduling. Kubernetes evaluates the Pod’s affinity rules for each of the possible Nodes that could be used for scheduling, then selects the most suitable one.
The affinity system is capable of supporting complex scheduling behavior, but it’s also easy to misconfigure affinity rules. When this happens, Pods will unexpectedly schedule to incorrect Nodes, or refuse to schedule or all. Inspect affinity rules for contradictions and impossible selectors, such as labels which no Nodes possess.
Network policies control the permissible traffic flows to Pods in your cluster. Each
NetworkPolicy object targets a set of Pods and defines the IP address ranges, Kubernetes namespaces, and other Pods that the set can communicate with.
Pods that aren’t covered by a policy have no networking restrictions imposed. This is a security issue because it unnecessarily increases your attack surface. A compromised neighboring container could direct malicious traffic to sensitive Pods without being subject to any filtering.
Including all Pods in at least one
NetworkPolicy is a simple but effective layer of extra protection. Policies are easy to create, too – here’s an example where only Pods labeled
app-component: api can communicate with those labeled
apiVersion: networking.k8s.io/v1 kind: NetworkPolicy metadata: name: database-policy spec: podSelector: matchLabels: app-component: database policyTypes: - Ingress - Egress ingress: - from: - podSelector: matchLabels: app-component: api egress: - to: - podSelector: matchLabels: app-component: api
Accurate visibility into cluster utilization, application errors, and real-time performance data is essential as you scale your apps in Kubernetes. Spiking memory consumption, Pod evictions, and container crashes are all problems you should know about, but standard Kubernetes doesn’t come with any observability features to alert you when problems occur.
To enable monitoring for your cluster, you should deploy an observability stack such as Prometheus. This collects metrics from Kubernetes, ready for you to query and visualize on dashboards. It includes an alerting system to notify you of important events.
Kubernetes without good observability can create a false sense of security. You won’t know what’s working or be able to detect emerging faults. Failures will be harder to resolve without easy access to the logs that preceded them.
Objects such as Deployments and Services rely on correct label selectors to identify the Pods and other objects they manage. Mismatches between selectors and the labels actually assigned to your objects will cause your deployment to fail.
The following example demonstrates this problem:
apiVersion: apps/v1 kind: Deployment metadata: name: demo-deployment spec: replicas: 3 selector: matchLabels: app: demo-app template: metadata: labels: # Label does not match the deployment's selector! app: demo-application spec: containers: name: demo-app image: nginx:latest
When this happens, Kubectl will display a
selector does not match template labels error. To fix the problem, adjust your manifest’s
spec.template.metadata.labels fields so they have the same key-value pairs.
Similarly, it’s important to make sure your services route traffic to the correct port on your Pods. Incorrect service port definitions can make it look like a Pod has failed, when in fact your traffic simply isn’t reaching it.
The following manifest contains an example of this problem. The service listens on port
9000 and forwards traffic to port
8080 on its Pods, but the container actually expects traffic to hit port
apiVersion: v1 kind: Pod metadata: name: demo-pod labels: app: demo-app spec: image: nginx:latest ports: - containerPort: 80 --- apiVersion: v1 kind: Service metadata: name: demo-service spec: ports: - port: 9000 protocol: TCP targetPort: 8080 selector: app: demo-app
LoadBalancer services in your cluster can be useful but is often unintentionally wasteful. Each
LoadBalancer service you create will provision a new load balancer and external IP address from your cloud provider, increasing your costs.
Ingress is a better way to publicly expose multiple services using HTTP routes. Installing an Ingress controller such as Ingress-NGINX lets you direct traffic between your services based on characteristics of incoming HTTP requests, such as URL and hostname.
With Ingress, you can use a single load balancer to serve all your applications. Only add another load balancer when your application requires an additional external IP address, or manual control over routing behavior.
Read more about Kubernetes load balancers.
Kubernetes namespaces logically group objects together, providing a degree of isolation in your cluster. Creating a namespace for each team, app, and environment prevents name collisions and simplifies the management experience.
When using namespaces, remember to specify the target namespace for each of your objects and Kubectl commands. Otherwise, the
default namespace will be used. This can be a debugging headache if objects don’t appear where you expected them.
metadata.namespace field on all your namespaced objects so they’re added to the correct namespace:
apiVersion: v1 kind: Pod metadata: name: demo-pod namespace: demo-app spec: # ...
--namespace flag with your Kubectl commands to scope an operation to a namespace:
# Get the Pods in the demo-app namespace $ kubectl get pods -n demo-app
This flag is also supported by Kubernetes ecosystem tools such as Helm. For a simpler namespace-switching experience, try kubens to quickly change namespaces and persist your selection between consecutive commands.
Correct resource management is essential to preserve your cluster’s stability. Pods don’t apply any resource limits unless you configure them, which can permit CPU and memory exhaustion to occur.
Set proper resource requests and limits on all your Pods to reduce resource contention. A request instructs Kubernetes to reserve a particular amount of a resource for your Pod, preventing it from scheduling onto Nodes that can’t provide enough capacity. Limits set the maximum amount of the resource which the Pod can use; Pods that exceed a CPU limit will be throttled, while reaching a memory limit prompts the out-of-memory (OOM) killer to terminate the process running in the Pod.
Requests and limits are defined in the
spec.container.resources field of a Pod’s manifest:
apiVersion: v1 kind: Pod metadata: name: demo-pod spec: containers: - name: demo-container image: nginx:latest resources: requests: cpu: 100m memory: 1Gi limits: memory: 1Gi
This Pod requests 100m (100 millicores) of CPU time and 1Gi of memory. It will only schedule onto Nodes that can provide sufficient resources. The Pod also has a memory limit set which prevents it using more than the requested 1Gi. It is best practice to set a Pod’s memory limit equal to its request. CPU limits aren’t usually required because Kubernetes proportionally throttles Pods that exceed their request.
Pod disruption budgets inform Kubernetes how much disruption your app can tolerate. They’re used during periods of restricted cluster availability, such as when Nodes are offline for an upgrade.
Disruption budgets tell Kubernetes that a specified number of Pods must be kept available when disruption occurs. The following
PodDisruptionBudget object preserves at least three replicas of Pods with the
app: demo-app label:
apiVersion: policy/v1 kind: PodDisruptionBudget metadata: name: demo-pdp spec: minAvailable: 2 selector: matchLabels: app: demo-app
Alternatively, you can specify the maximum number of Pods that Kubernetes can evict when disruption is encountered:
apiVersion: policy/v1 kind: PodDisruptionBudget metadata: name: demo-pdp spec: minAvailable: 2 selector: matchLabels: app: demo-app
Using this mechanism makes your application more responsive to times of reduced cluster capacity. It guarantees that a minimum service level is maintained.
Kubernetes is often seen as a route to easy scalability. A correctly configured cluster lets you dynamically scale both horizontally and vertically by automatically adding new Pods and Nodes when demand spikes. Unfortunately, many teams scale their clusters incorrectly or find their auto-scaling is unpredictable.
Regularly review your cluster’s utilization to check whether it’s still suitable for your workloads. Test autoscaling rules by using a load-testing tool like Locust to direct excess traffic to your cluster. This lets you spot problems earlier, ensuring your Pods will scale seamlessly when real traffic arrives.
Overall cluster performance depends on Pods being correctly scheduled to suitable Nodes. Many clusters combine several types of Node, such as small 2 CPU/4 GB machines for standard applications and larger 8 CPU/16GB Nodes for intensive backend services.
Cluster utilization will be inefficient if your Pods don’t reliably schedule to the Node pool you’d intended. This can increase your cluster’s costs by forcing new larger Nodes to be created unnecessarily, even though underused smaller ones are available. Avoid this problem by setting labels on your Nodes, then using node selectors to assign each Pod to a compatible Node:
apiVersion: v1 kind: Pod metadata: name: pod-node-selector-demo spec: containers: - name: nginx image: nginx:latest nodeSelector: node-class: 2vcpu4gb
This Pod will only schedule to Nodes that have the
node-class: 2vcpu4gb label set.
kubectl label command to set labels on your matching Nodes:
$ kubectl get nodes NAME STATUS ROLES AGE VERSION minikube Ready control-plane 10d v1.26.1 $ kubectl label node minikube node-class=2vcpu4gb node/minikube labelled
Setting proper scheduling constraints will maximize Node usage and maintain stable cluster performance.
Standard tools, including Kubectl and the Kubernetes API become inefficient when you’re managing larger clusters. Automating your infrastructure with an IaC solution lets you reliably provision clusters and collaborate on changes.
Avoid manually administering Kubernetes by integrating cluster operations into your CI/CD and GitOps workflows using a management platform such as Spacelift. Spacelift helps you reduce Kubernetes complexity, synchronize changes between environments, and enforce compliance policies. It works with other infrastructure components, including Terraform and Pulumi, too.
Pod security admission standards allow you to enforce security best practices for your cluster. Admission controllers are able to reject Pods which don’t meet specified security criteria, such as when privileged capabilities or direct host port bindings are used.
Kubernetes ships with three different security standards: Privileged, Baseline, and Restricted. The Restricted policy gives you the best protection by enforcing that all Pods adhere to current hardening best practices. Baseline is suitable for less critical scenarios, while Privileged removes the restrictions to support workloads that require privilege escalation.
Read more about Container security best practices and solutions.
Kubernetes is the industry-standard orchestrator for cloud-native systems, but popularity doesn’t mean perfection. To get the most from Kubernetes, your developers, and operators need to correctly configure your cluster and its objects to avoid errors, sub-par scaling, and security vulnerabilities.
This guide has covered 15 challenges to look for each time you use Kubernetes. While these will solve the most commonly encountered issues, you should also review Kubernetes best practices to get even more out of your cluster.
When operating Kubernetes becomes too demanding, try using an IaC platform to provision and manage your clusters. Spacelift is a collaborative solution for visualizing your infrastructure, enforcing policies, and preventing drift.
The Most Flexible CI/CD Automation Tool
Spacelift is an alternative to using homegrown solutions on top of a generic CI. It helps overcome common state management issues and adds several must-have capabilities for infrastructure management.