Control multicluster costs with right-sizing in Red Hat Advanced Cluster Management 5

Red Hat Advanced Cluster Management for Kubernetes 5 makes the multicluster observability add-on (MCOA) delivery model and Perses dashboards generally available (GA) for namespace and OpenShift Virtualization right-sizing recommendations. The release also adds configurable percentile profiles for more precise CPU and memory recommendations.

For platform engineering and FinOps teams, right-sizing a fleet is a balancing act. Requests that remain well above actual use lock capacity away from other workloads and increase infrastructure cost. Requests that remain below demand can cause throttling, degraded performance, and unstable applications. The challenge grows when the same analysis must cover many clusters, namespaces, and virtual machines (VMs).

Red Hat Advanced Cluster Management for Kubernetes 2.16 made Namespace and OpenShift Virtualization right-sizing recommendations generally available. The capability used long-term metrics from the multicluster observability stack to compare actual CPU and memory use with configured requests and present the results in Grafana dashboards.

In Red Hat Advanced Cluster Management 2.17, we introduced a developer preview of right-sizing recommendations delivered by the multicluster observability add-on (MCOA) and visualized with Perses. Customers could opt in to evaluate the new architecture while retaining the option to return to the earlier policy-based delivery model.

Red Hat Advanced Cluster Management 5 completes that transition. The MCOA delivery model and Perses right-sizing dashboards are now generally available and have become the standard experience. Existing right-sizing configuration continues to work, upgrades migrate automatically, and new percentile profiles give administrators more control over how recommendations account for workload peaks.

What's new with right-sizing GA in Red Hat Advanced Cluster Management 5

While the primary objective of right-sizing remains unchanged—analyzing historical workload behavior to balance application performance and infrastructure costs across your fleet—Red Hat Advanced Cluster Management 5 fundamentally modernizes how recommendations are delivered, visualized, and managed:

• MCOA distributes the right-sizing PrometheusRule resources to selected managed clusters by using ManifestWork.
• Perses dashboards provide the primary integrated right-sizing experience in the Red Hat Advanced Cluster Management console.
• Namespace and OpenShift Virtualization right-sizing continue to use the same MultiClusterObservability custom resource configuration and the same rs-namespace-config and rs-virt-config customization model.
• Upgrades from Red Hat Advanced Cluster Management 2.17 migrate automatically, whether or not the 2.17 MCOA developer preview was enabled.
• Red Hat Advanced Cluster Management 5 removes the earlier policy-based delivery path. Grafana dashboards remain available in Red Hat Advanced Cluster Management 5, with their retirement handled separately.
• The release supports dual metrics collection paths, allowing you to use end-to-end MCOA platform metrics federation (ScrapeConfig + PrometheusAgent) or retain your existing observability-metrics-allowlist pipeline.

There's no 2.17-style opt-in annotation to manage in Red Hat Advanced Cluster Management 5. A fresh installation uses MCOA and Perses by default. During an upgrade, the platform removes the old right-sizing policy resources and creates the required ManifestWork resources and Perses dashboards without requiring a manual migration.

What to review when upgrading

For the standard configuration, upgrading from Red Hat Advanced Cluster Management 2.16 or 2.17 requires no migration action. Existing rs-namespace-config and rs-virt-config customizations are preserved while the platform removes legacy Policy, Placement, and PlacementBinding resources.

Upgrades require no pre-upgrade cleanup. Leave existing settings in place until the upgrade completes so the migration controller can use any custom namespaceBinding value to locate legacy resources. After migration, 2 settings from the earlier delivery model no longer control right-sizing in Red Hat Advanced Cluster Management 5:

• The system ignores the observability.open-cluster-management.io/right-sizing-capable annotation used to opt in to the 2.17 developer preview, and you can remove it after the upgrade.
• The namespaceBinding field previously selected the namespace where the system created legacy governance resources and bound them to a managed cluster set. It's retained for migration compatibility, but it no longer controls right-sizing delivery. In Red Hat Advanced Cluster Management 5, placementConfiguration in the right-sizing ConfigMaps determines which managed clusters receive the rules.

For example, the following placementConfiguration targets clusters labeled environment=prod:

data:
  placementConfiguration: |
    spec:
      predicates:
        - requiredClusterSelector:
            labelSelector:
              matchLabels:
                environment: prod
    ...

If you customized namespaceBinding, verify your cluster targeting after the upgrade. Existing placementConfiguration predicates are preserved. If a custom namespace and its managed cluster set binding previously limited the eligible clusters, express that intent with placementConfiguration predicates and managed cluster labels. You can then remove the obsolete namespaceBinding value.

A consistent add-on architecture for observability

Before this transition, right-sizing recommendations were among the remaining observability capabilities delivered through the governance policy framework. Metrics, logging, and tracing had already moved to the MCOA framework. Bringing right-sizing into the same model gives observability capabilities a consistent delivery and lifecycle path, as summarized in the following architectural comparison.

The flow in Red Hat Advanced Cluster Management 5 is straightforward:

1. The MultiClusterObservability resource defines whether namespace and virtualization right-sizing are enabled.
2. The analytics controller synchronizes right-sizing state and configuration.
3. MCOA builds the resources required for each matching managed cluster.
4. ManifestWork delivers the right-sizing PrometheusRule resources to that cluster.
5. Aggregated metrics return to hub Thanos and populate the Perses dashboards.

This architecture keeps cluster targeting, namespace filters, label filters, and recommendation thresholds under administrator control while aligning right-sizing with the rest of the observability architecture.

Explore recommendations in Perses

Perses is an open source Cloud Native Computing Foundation project built for cloud-native observability. Its Kubernetes-native, dashboards-as-code model lets the right-sizing dashboards be versioned, reviewed, and tested with the application code. The dashboards are integrated into the Red Hat Advanced Cluster Management console and provide links that support top-down investigation.

Open the dashboards from Observe → Dashboards (Perses) in the OpenShift console, choose Project as observability-analytics, and open the Namespace or OpenShift Virtualization right-sizing dashboard.

The namespace right-sizing dashboard begins with selectors for the managed cluster, recommendation profiles, and aggregation period. It includes:

• Cluster-level CPU and memory summaries for recommendation, usage, request, and utilization
• Time-series views of the namespaces with the highest CPU and memory utilization ratios
• CPU and memory tables that compare usage, requests, recommendations, and namespace quota limits

The OpenShift Virtualization dashboard summarizes CPU and memory overestimation and underestimation across running virtual machines. Per-VM tables show usage, request, recommendation, utilization, and the difference between the current request and recommendation. Links from each virtual machine open detailed dashboards with current values and usage trends over time.

See Perses right-sizing in action

See the new Perses right-sizing dashboards in action. Walk through percentile profile selections, cluster filters, and workload recommendations for namespaces and OpenShift Virtualization in the following interactive tour:

Tune CPU and memory with percentile profiles

Red Hat Advanced Cluster Management 5 adds configurable CPU and memory aggregation profiles for both namespace and virtualization right-sizing. The default profiles are:

• Max OverAll, which bases the recommendation on the maximum observed value in the aggregation window
• P99, which uses the 99th percentile
• P95, which uses the 95th percentile

Maximum-based recommendations preserve the most headroom but can be influenced by an isolated spike. Percentile profiles let teams reduce the effect of rare outliers while retaining headroom for normal variation. P99 discounts fewer high-end samples than P95. Teams should validate any profile against workload service-level objectives before changing resource requests.

MCOA records each profile from the 5-minute usage series over a rolling 1-day window. Max OverAll uses max_over_time, while percentile profiles use quantile_over_time. For summary and comparison values, the dashboards then use the maximum recorded value across the selected number of days. The recommendation applies the configured recommendationPercentage to the 1-day usage profile; the default is 110, which adds 10% headroom.

CPU and memory profiles are independent. For example, an application with bursty CPU use and stable memory use might apply P99 for CPU and P95 for memory. The Perses dashboards keep both selections as users move from overview panels to detailed views.

Administrators can also define other integer percentile profiles from P1 through P99. Add or update cpuAggregator and memoryAggregator inside the existing prometheusRuleConfig value in either of these ConfigMaps:

• rs-namespace-config for namespace right-sizing
• rs-virt-config for OpenShift Virtualization right-sizing

Both ConfigMaps are in the open-cluster-management-observability namespace. The following focused example adds P90 for CPU while retaining the default memory profiles. Keep the other existing fields in prometheusRuleConfig when editing the ConfigMap.

Edit namespace profiles with the following command:

oc edit configmap rs-namespace-config -n open-cluster-management-observability

For virtualization profiles, edit rs-virt-config in the same namespace.

data:
  prometheusRuleConfig: |
    cpuAggregator:
      - Max OverAll
      - P99
      - P95
      - P90
    memoryAggregator:
      - Max OverAll
      - P99
      - P95
    ...

After a valid change, MCOA regenerates the corresponding PrometheusRule resources and the configured profiles become available in the dashboards. Percentile profile names are case-normalized and duplicate entries are removed. CPU and memory lists are validated independently. If either list contains an invalid profile name, MCOA restores that entire field to its previous valid list while preserving a valid change to the other field. If no previous value is available, the affected field returns to the default profiles. An absent or empty list also resolves to the defaults. This validation prevents an invalid value from leaving the 1-day aggregation rules empty.

During upgrade, Red Hat Advanced Cluster Management 5 adds missing aggregator keys to ConfigMaps created by earlier releases while preserving existing filters, recommendation percentages, placement, and other customizations.

Choose a metrics collection path

Red Hat Advanced Cluster Management 5 supports 2 paths for collecting right-sizing metrics:

• With platform metrics collection enabled, MCOA deploys the right-sizing PrometheusRule resources and a ScrapeConfig to each matching managed cluster. The platform PrometheusAgent federates the resulting acm_rs:* and acm_rs_vm:* metrics from the managed cluster and remote-writes them to hub Thanos. This is the recommended path because the add-on manages the pipeline end to end.
• Environments that don't enable platform metrics collection continue to work with the existing observability pipeline. MCOA deploys the PrometheusRule resources, and the existing metrics collector forwards the right-sizing metrics through the observability-metrics-allowlist. This compatibility lets customers adopt the Red Hat Advanced Cluster Management 5 right-sizing experience without changing their metrics collection architecture immediately.

Get started

After installing or upgrading to Red Hat Advanced Cluster Management 5:

1. Enable the multicluster observability service on the hub cluster.
2. Confirm that namespace and virtualization right-sizing are enabled in the MultiClusterObservability resource. Both are enabled by default with observability; OpenShift Virtualization must be available on the managed clusters where VM recommendations are required.
3. In the OpenShift console, select Observe → Dashboards (Perses), then choose Project: observability-analytics and open the Namespace or OpenShift Virtualization right-sizing dashboard.
4. Select a cluster, CPU profile, memory profile, and aggregation period to review utilization recommendations. Analyzing these sustained trends helps administrators identify over- or underused resources before adjusting workload requests, namespace quotas, or VM allocations.

Right-sizing recommendations are an analytical input rather than an automatic resource change. Apply adjustments gradually, verify application behavior, and keep enough headroom for known demand patterns.

Operational considerations

Keep the following behavior in mind when interpreting a new or upgraded deployment:

• Prometheus rules identify managed clusters by cluster name rather than cluster ID.
• Historical right-sizing data isn't backfilled. After installation, longer aggregation periods become useful only after enough new data has accumulated.
• It can take approximately 30 minutes for recommendations to appear after right-sizing is enabled because metrics must be collected and aggregated through Thanos.
• The platform deploys right-sizing resources only to OpenShift managed clusters.
• Cluster targeting, namespace inclusion or exclusion, the label filter, recommendation percentages, and CPU and memory aggregator profiles remain configurable through the right-sizing ConfigMaps.
• namespaceFilterCriteria accepts inclusion or exclusion criteria, but not both at the same time.
• labelFilterCriteria acts only on an entry whose labelName is label_env. The Red Hat Advanced Cluster Management 5 rule builder ignores other label names, and the label_env entry cannot combine inclusion and exclusion criteria.

From insight to efficient capacity

The Red Hat Advanced Cluster Management 5 release makes the modern right-sizing architecture ready for production use. MCOA and ManifestWork provide consistent fleet delivery, Perses brings Kubernetes-native dashboards into the console, and configurable percentile profiles let teams choose how much weight to give workload peaks for CPU and memory independently.

Together, these changes help platform engineers and FinOps teams move from a static view of requested capacity to a repeatable, evidence-based process for improving resource use across namespaces and virtual machines.

What's next

We're continuing to evolve right-sizing in Red Hat Advanced Cluster Management. In upcoming releases, look for expanded recommendation support at both the workload and individual pod levels, giving you deeper granular control over application capacity.

We value your feedback, which helps us shape the future of right-sizing in Red Hat Advanced Cluster Management. Share your questions and recommendations with us using the Red Hat OpenShift feedback form.