Multi-dimensional autoscaling (MDA) optimizes both workload resources and replica counts to:
maintain stability and performance
eliminate workload over- and under-provisioning
identify workloads at risk
MDA combines two optimization capabilities:
Pod rightsizing: optimizes CPU and memory resources for individual Pods
Replica optimization: optimizes minimum replica counts for workloads
Pod rightsizing helps ensure that each Pod is allocated the appropriate CPU and memory resources based on actual workload demand.
However, resource optimization alone cannot eliminate all inefficiencies. Workloads can remain overprovisioned when they maintain more replicas than necessary, even if each replica is perfectly rightsized.
Minimum replica values are often configured conservatively to reduce operational risk and then remain unchanged as workload behavior evolves. As a result, workloads can continue consuming unnecessary compute resources even after CPU and memory requests have been optimized.
By combining Pod rightsizing and Replica optimization, MDA addresses both resource-level and replica-level inefficiencies, continuously aligning workload resources with actual demand across CPU, memory, and replicas.
MDA delivers these benefits by:
Reducing baseline overprovisioning caused by unnecessarily high minimum replica counts.
Capturing cost savings opportunities that Pod rightsizing alone cannot achieve.
Optimizing workload resources across CPU, memory, and replicas.
Reducing the need for ongoing manual tuning of workload resource settings.
Applying optimization through policy-driven automation.
MDA operating modes
Pod rightsizing can operate on its own, but Replica optimization requires Pod rightsizing to be active on the workload.
The following operating modes are supported:
Mode | Pod rightsizing solution | Replica optimization solution |
|---|---|---|
Pod rightsizing only | Active | Inactive |
Multi-dimensional autoscaling | Active | Active |
Multi-dimensional autoscaling generates recommendations only for resources that have requests configured for CPU, memory, or both.
For more information about Pod rightsizing, see Pod rightsizing.
For more information about Replica optimization, see Replica optimization.
Related optimization solutions
MDA optimizes workload resources across CPU, memory, and replicas. However, infrastructure inefficiencies can still remain when workloads are distributed in a way that prevents nodes from being consolidated and scaled down.
Adaptive Pod placement addresses a different aspect of resource optimization by improving workload placement and helping clusters reclaim additional capacity through node consolidation.
For more information, see Adaptive Pod placement.