VPA recommender — the recommendation engine

The VPA Recommender is the recommendation engine. It observes the pod’s resource usage and computes the recommended requests. The recommendation is at the 90th percentile of the usage distribution. This lesson walks the recommender, the percentile approach, and the inspection of the recommendations.

The recommendation algorithm

The VPA Recommender uses a percentile-based approach:

flowchart LR
    A[Pod usage history] --> B[Compute percentiles]
    B --> C[90th percentile]
    C --> D[Recommended requests]
    D --> E[Bound by minAllowed/maxAllowed]
    E --> F[Final recommendation]

The recommender collects the pod’s usage history, computes the percentiles, and selects the 90th percentile as the recommendation.

The historical data

The recommender collects the pod’s usage from the metrics server:

Pod: nginx-1-abc
Sample 1: CPU 50m, memory 128Mi
Sample 2: CPU 60m, memory 130Mi
Sample 3: CPU 45m, memory 125Mi
...
Sample N: CPU 55m, memory 129Mi

The samples are collected every 15 seconds (default for the metrics server). The recommender aggregates the samples into a distribution.

The percentiles

The recommender computes the percentiles:

CPU usage distribution:
  10th percentile: 40m
  50th percentile (median): 55m
  90th percentile: 80m
  99th percentile: 120m

Memory usage distribution:
  10th percentile: 100Mi
  50th percentile: 128Mi
  90th percentile: 160Mi
  99th percentile: 200Mi

The 90th percentile is the recommendation.

The recommendation bounds

The recommendation is bounded by the VPA’s minAllowed and maxAllowed:

resourcePolicy:
  containerPolicies:
  - containerName: "*"
    minAllowed:
      cpu: 100m
      memory: 128Mi
    maxAllowed:
      cpu: 2
      memory: 4Gi

The recommendation is the 90th percentile, bounded by the min/max:

90th percentile: 80m
minAllowed: 100m
maxAllowed: 2

Recommendation: max(100m, min(80m, 2)) = 100m

The recommendation output

kubectl describe vpa nginx-vpa

Name:         nginx-vpa
Namespace:    default
Labels:       <none>
Annotations:  <none>
API Version:  autoscaling.k8s.io/v1
Kind:         VerticalPodAutoscaler
Metadata:
  ...
Spec:
  Target Ref:
    API Version:  apps/v1
    Kind:         Deployment
    Name:         nginx
Status:
  Conditions:
    Last Transition Time:  2026-08-16T10:00:00Z
    Status:                True
    Type:                  RecommendationProvided
  Recommendation:
    Container Recommendations:
      Container Name:  nginx
      Lower Bound:
        Cpu:     50m
        Memory:  100Mi
      Target:
        Cpu:     100m
        Memory:  200Mi
      Uncapped Target:
        Cpu:     80m
        Memory:  180Mi
      Upper Bound:
        Cpu:     200m
        Memory:  300Mi

The output shows:

• Lower Bound: the minimum recommended request.
• Target: the recommended request (the 90th percentile).
• Uncapped Target: the raw 90th percentile, before bounds.
• Upper Bound: the maximum recommended request.

The recommendation modes

The recommender supports different modes:

• Auto: the recommender uses the 90th percentile.
• Initial: the recommender uses the initial requests.
• Off: the recommender does not compute recommendations.

The default mode is Auto.

The tuning

The recommender can be tuned with the recommendation-margin-fraction:

spec:
  resourcePolicy:
    containerPolicies:
    - containerName: "*"
      recommendationMarginFraction:
        cpu: 0.15
        memory: 0.15

The margin adds 15% to the recommendation. The tuned recommendation is 90th_percentile * 1.15.

The margin is useful when the workload has spikes that are not captured by the 15-second metrics interval.

The historical window

The recommender uses the past 8 days of data by default:

spec:
  resourcePolicy:
    containerPolicies:
    - containerName: "*"
      historicalDataWindow: 8d

The historical window is configurable. A longer window captures seasonal patterns; a shorter window focuses on recent behavior.

The inspect metrics

The VPA Recommender exposes metrics:

vpa_recommender_recommendation_age_seconds
vpa_recommender_models_total
vpa_recommender_pods_count
vpa_recommender_cpu_recommendation
vpa_recommender_memory_recommendation

The metrics are scraped by Prometheus. The metrics indicate the recommender’s health.

Cross-course references

• The HPA course (Part LXXXII) covers the horizontal alternative.
• The Prometheus course covers metric collection.
• The Observability course covers VPA monitoring.

Quiz

Production discipline

• Inspect the VPA recommendations regularly. Use kubectl describe vpa.
• Compare the Target to the current request. Identify the right-size opportunity.
• Set minAllowed and maxAllowed. Bound the recommendations.
• Tune the recommendation-margin-fraction. Catch the spikes.
• Tune the historical window. Capture seasonal patterns.
• Document the recommendations. The Target, the bounds, the rationale.

The VPA Recommender is the recommendation engine. Operating it well is inspecting the recommendations, setting the bounds, and tuning the algorithm.