What Is the Vertical Pod Autoscaler? Effectively Optimizing Pod Resources in Kubernetes

In Kubernetes, manually setting CPU and memory resources for Pods can easily lead to either resource shortages or infrastructure waste. Improper configuration can cause applications to slow down, experience OOMKilled errors, or prevent the cluster from fully utilizing its available capacity. The Vertical Pod Autoscaler provides a smarter approach by automatically recommending and adjusting resources based on actual usage. In this article, Viettel IDC will help you gain a better understanding of VPA.

What Is the Vertical Pod Autoscaler?

The Vertical Pod Autoscaler (VPA) is an autoscaling mechanism in Kubernetes that automatically adjusts the CPU and memory resources allocated to Pods vertically. Unlike the Horizontal Pod Autoscaler (HPA), which increases or decreases the number of Pods, VPA focuses on adjusting the resource requests and limits of individual Pods to better match actual workload requirements.

VPA monitors Pod resource consumption over time and analyzes CPU and memory usage trends to provide recommendations or directly update resource configurations. The primary goal of VPA is not to handle immediate traffic spikes, but rather to ensure that Pods receive the right amount of resources over the long term.

In practice, many Kubernetes applications are deployed with resource requests based on rough estimates or copied from another environment. VPA helps eliminate this guesswork by replacing it with data-driven decisions, thereby improving both cluster stability and resource efficiency.

How Does the Vertical Pod Autoscaler Work?

The Vertical Pod Autoscaler operates by collecting Pod resource utilization metrics, analyzing historical data, and making decisions about resource adjustments. These metrics are typically obtained from monitoring systems such as Metrics Server, Prometheus, or other sources compatible with the Kubernetes API.

The overall VPA process begins by monitoring CPU and memory usage of containers within Pods. Instead of looking only at peak values, VPA analyzes the distribution of usage data over time to determine an appropriate resource level, including safety margins to prevent sudden resource shortages.

Once recommendations are available, VPA can operate in several different modes. In passive mode, VPA only provides recommendations for human operators to review. In automatic mode, VPA actively updates resource requests and may restart Pods to apply the new configuration. For this reason, VPA is better suited to long-term resource optimization than real-time autoscaling.

Key Components of the Vertical Pod Autoscaler

VPA Recommender

The VPA Recommender is the central component responsible for analyzing usage data and generating resource recommendations. It collects CPU and memory utilization metrics from Pods over a sufficiently long period to ensure accuracy and avoid being overly influenced by short-term spikes.

The Recommender uses statistical algorithms to determine optimal resource requests, typically including target, lower bound, and upper bound values. These values help the system balance performance and safety, preventing resource allocations that are too low and may cause failures, or too high and result in unnecessary waste.

VPA Updater

The VPA Updater is responsible for applying recommendations to running Pods. When it detects a significant difference between the current resource requests and the recommendations, the Updater may decide to evict the Pod so that a new Pod can be created with a more appropriate resource configuration.

Pod eviction does not happen randomly; it follows safety policies such as the Pod Disruption Budget (PDB). This helps minimize the risk of downtime, particularly in production environments.

VPA Admission Controller

The VPA Admission Controller operates when a new Pod is created. It intercepts requests sent to the Kubernetes API Server and automatically modifies resource requests and limits based on the current VPA recommendations.

This component is particularly important in modes such as Initial or Auto because it ensures that newly created Pods are initialized with an optimized resource configuration, even before they start running and generating metrics.

Vertical Pod Autoscaler Operating Modes

- Off mode: In Off mode, VPA only observes and analyzes resource usage without directly affecting Pods. All recommendations are stored and displayed for operators to review. This mode is suitable when VPA is first deployed or when a team wants to understand an application's resource usage behavior before allowing VPA to intervene automatically.

- Initial mode: In Initial mode, VPA applies recommendations only when a new Pod is created. Running Pods are not restarted or evicted. This is an intermediate mode that provides some of the benefits of VPA while reducing the risk of service disruption. Initial mode is commonly used in production environments where high stability is required.

- Auto mode: Auto mode is the most aggressive mode, in which VPA automatically updates resources and may restart Pods to apply the new configuration. This mode provides more comprehensive resource optimization but also introduces availability risks if not configured carefully. Auto mode is suitable for batch workloads, scheduled jobs, or systems that can tolerate restarts without significantly affecting users.

How Is the Vertical Pod Autoscaler Different from the Horizontal Pod Autoscaler?

Advantages and Disadvantages of the Vertical Pod Autoscaler

Advantages of VPA

The biggest advantage of the Vertical Pod Autoscaler is its ability to optimize resource requests based on actual usage data rather than manual estimates. In many Kubernetes clusters, resource requests are intentionally configured higher than necessary for safety, resulting in situations where nodes still have available resources but the scheduler cannot place additional Pods. VPA helps address this issue by adjusting resource requests closer to actual consumption, thereby improving resource utilization across the entire cluster.

VPA also plays an important role in maintaining application performance stability. When a Pod frequently reaches CPU throttling thresholds or experiences memory pressure, VPA can identify these trends and recommend higher resource allocations. This helps reduce OOMKilled errors, minimize abnormal latency, and prevent difficult-to-debug issues related to insufficient resources.

Disadvantages and Limitations of VPA

The biggest limitation of the Vertical Pod Autoscaler is that changing resources often requires restarting a Pod. In Auto mode, when VPA decides to apply a new recommendation, a Pod may be evicted and recreated with a different resource configuration. If the application is not designed to be stateless or does not have a proper graceful shutdown mechanism, this process may cause service disruption.

VPA is also not well suited to workloads that require a rapid response to traffic fluctuations. Because VPA relies on historical data and long-term trend analysis, it cannot immediately increase resources when traffic suddenly spikes for a few minutes. In scenarios such as flash sales, livestreaming, or APIs experiencing high traffic for short periods, the Horizontal Pod Autoscaler remains a more suitable choice.

When Should You Use the Vertical Pod Autoscaler?

The Vertical Pod Autoscaler is particularly suitable for monolithic applications or workloads that are difficult to scale horizontally, where simply increasing the number of Pods provides limited benefits. In these systems, allocating the right amount of CPU and memory to each Pod is often much more important than replicating Pods.

VPA is also well suited to batch jobs, cron jobs, and periodic workloads where resource consumption can vary over time but real-time processing is not required. Allowing VPA to automatically adjust resources can help these jobs run more consistently without requiring manual configuration for every scenario.

When Should You Not Use the Vertical Pod Autoscaler?

The Vertical Pod Autoscaler is not an ideal choice for systems that require absolute high availability, where even a brief Pod restart is unacceptable. Financial systems, real-time transaction processing platforms, and continuously running streaming systems are generally not well suited to vertical autoscaling mechanisms.

VPA is also not a good fit for microservices that are already designed to scale horizontally efficiently using HPA. In such systems, adding VPA may increase operational complexity without providing significant benefits and could even make scaling behavior more difficult to predict.

Vertical Pod Autoscaler Deployment Process in Kubernetes

Installing VPA on the Cluster

Deploying the Vertical Pod Autoscaler typically begins with installing its core components: the Recommender, Updater, and Admission Controller. These components are deployed as Pods within the cluster and require access to metrics and the Kubernetes API to operate correctly.

Installation should initially be performed in a staging or test environment to evaluate its impact, particularly VPA's Pod eviction behavior under different operating modes.

Defining the VerticalPodAutoscaler Manifest

Once the VPA core components are running, the next step is to define a VerticalPodAutoscaler resource for each specific workload. This manifest specifies the target workload (Deployment, StatefulSet, etc.), the VPA operating mode, and the resource ranges that VPA is allowed to modify.

Configuring the appropriate scope from the beginning helps prevent VPA from interfering with unintended Pods, especially in clusters shared by multiple teams.

Monitoring Recommendations and Making Adjustments

Once VPA begins collecting data, monitoring its recommendations is an essential step. These recommendations provide a clear view of how much CPU and memory an application actually needs, allowing the team to evaluate whether the recommendations are reasonable before enabling automatic adjustments.

At this stage, VPA is typically operated in Off or Initial mode to ensure system safety.

Best Practices for Rolling Out VPA in Production

In production environments, the recommended approach is to roll out VPA gradually, starting with Off mode, then moving to Initial mode, and enabling Auto mode only after the associated risks have been properly assessed and controlled.

Combining VPA with a Pod Disruption Budget, readiness probes, and comprehensive monitoring is critical to ensuring that the system does not experience unintended service disruptions.

Conclusion

The Vertical Pod Autoscaler is a powerful tool that helps Kubernetes operate more efficiently by optimizing Pod resources based on actual usage data. When properly deployed, VPA can not only improve application performance but also significantly reduce infrastructure costs.

If your business wants to deploy Kubernetes quickly, reliably, and cost-effectively, consider Viettel IDC's Viettel Open Kubernetes Service (vOKS) here. The Kubernetes platform service enables software developers to easily build, deploy, scale, and manage containerized applications:

https://viettelidc.com.vn/en/viettel-kubernetes-service