How do you scale applications in Kubernetes?

Kubernetes provides manual and automatic scaling methods. Manual scaling uses kubectl scale for direct control, ideal for predictable workloads, but requires manual adjustment back down. Automatic scaling uses the Horizontal Pod Autoscaler (HPA) based on metrics like CPU usage, adjusting replicas between defined min and max values, requiring resource limits and a metrics server. Cluster-level scaling uses the Cluster Autoscaler to add or remove nodes based on scheduling needs, best suited for unpredictable or large-scale workloads. Each method has trade-offs in setup complexity, cost, and responsiveness.

Scaling applications in Kubernetes boils down to two main methods: manual and automatic scaling. Both rely on Kubernetes’ built-in controllers like Deployments and ReplicaSets, which manage how many copies (replicas) of your app are running at any given time.

Manual Scaling with kubectl scale

If you want direct control over how many replicas of your application are running, you can use the kubectl scale command. It's straightforward and good for predictable workloads or one-off adjustments.

For example:

kubectl scale deployment my-app --replicas=5

This tells Kubernetes to run 5 instances of your app. You can check the current number of replicas by looking at the output of:

kubectl get deployments

• This method is fast and reliable.
• Great for scheduled traffic spikes you know about in advance.
• No need to set up extra resources like metrics servers.

Just keep in mind that once you scale manually, you’ll have to scale back down when the load decreases — unless you want to keep paying for unused capacity.

Automatic Scaling Based on Metrics

When your workload fluctuates unpredictably, it’s better to let Kubernetes handle the scaling automatically. That’s where the Horizontal Pod Autoscaler (HPA) comes in.

To use HPA, you first need to define what metric it should watch — usually CPU usage or memory consumption. Here's a basic example using CPU:

kubectl autoscale deployment my-app --cpu-percent=50 --min=2 --max=10

This means Kubernetes will try to keep CPU around 50% by adjusting the number of replicas between 2 and 10.

• Make sure your pods have resource limits defined; otherwise, HPA won’t have a baseline to work from.
• You'll also need a metrics server running in your cluster.
• You can create more advanced HPAs using custom metrics like request latency or queue size if you have the right tools (e.g., Prometheus adapter).

It takes a little setup, but once it’s working, it adjusts quietly in the background without you needing to lift a finger.

Cluster-Level Scaling with Cluster Autoscaler

Sometimes your nodes themselves become the bottleneck — maybe all your nodes are full and there’s no space to schedule new pods. That’s where the Cluster Autoscaler comes in.

The Cluster Autoscaler watches for pending pods due to lack of resources and responds by asking your cloud provider to add more nodes. When those nodes aren't needed anymore, it removes them to save costs.

• Works best when used with auto-scaling groups (like AWS ASGs or GCP MIGs).
• Needs proper tagging and IAM permissions depending on your cloud provider.
• Be aware of cooldown periods and delays — sometimes it takes a few minutes before new nodes show up.

This level of scaling isn’t always necessary, especially if you're running a small or predictable workload. But for larger, variable apps, it helps prevent downtime during unexpected traffic surges.

That’s basically it. Whether you’re doing it manually or letting Kubernetes take the wheel, the tools are there to make your life easier — as long as you set them up right.

The above is the detailed content of How do you scale applications in Kubernetes?. For more information, please follow other related articles on the PHP Chinese website!