“We were essentially able to reduce the cost of that cluster by about 75%. On AWS, DevZero demonstrated they could achieve significantly higher savings than we initially thought possible.”
Mihir Nair
Head of Architecture, Databahn
VPA right-sizes your pods
DevZero does it without touching them
VPA's Auto mode evicts pods. Its Recommendation mode requires manual application. DevZero rightsizes continuously, applies via live migration, and shows you the dollar value of every change; within policy guardrails you control.
VPA
Pod eviction on resize
Auto mode kills and restarts pods to apply new resource values
Conflicts with HPA
VPA and HPA targeting the same workload is officially unsupported
No policy controls
No way to scope which workloads get resized or how aggressively
No cost projections
Recommendations have no dollar value — you can't prioritize them
No node optimization
Workloads are rightsized but underlying node waste is untouched
Live migration, no restarts
CRIU checkpoint/restore resizes running workloads without eviction
HPA-aware scaling
Replica count adjustments respect existing HPA targets and multipliers
Policy-gated changes
Conservative, balanced, or aggressive modes — nothing applied without a rule
Dollar savings per recommendation
Every WorkloadRecommendation ships with a projected $/month saving
Node consolidation included
Node Operator drains underutilized nodes and right-sizes instance types
Companies who slashed their Kubernetes
spend using DevZero
What VPA does well and where it stops
VPA monitors CPU and memory usage and recommends updated resource requests. The gap is in how it applies them. Every update evicts the pod, waits for the scheduler to recreate it with new values, and fires the admission webhook restart cycle. For StatefulSets or Java heaps, that eviction is a multi-minute outage. And throughout all of it, VPA has no concept of node cost, bin packing efficiency, or what rightsizing actually saves your team in dollars.
Availability during VPA update
Serving
Evicted + recreating (downtime)
New pod online
pod-api-7x
VPA watching
CPU requested80m / 340m
4x over-provisioned
Recommender fires→
terminated
Pod evicted
Updater kills pod in Recreate mode
Scheduler wait begins
Node check · image pull · init containers
Webhook injects→
pod-api-9f
CPU now320m
New pod, new name, new state
Connections lost. State reset.
Every VPA recommendation that touches a running pod in Recreate mode triggers this full cycle. For StatefulSets with large volumes or Java heaps, the scheduler wait alone can run several minutes.
Before DevZero
Running
pod-api-7x
CPU340m actual demand
80m requested+260m gap
MEMORY480Mi actual demand
128Mi requested+352Mi gap
DevZero
After DevZero
Still running
pod-api-7x
CPU320m P95
320m requestedmatched
MEMORY480Mi actual demand
512Mi requestedzero waste
Throttling risk: resource gap
Correctly sized: zero throttling
In-place rightsizing no restart required
DevZero applies CPU and memory changes directly to running pods via cgroup v2. No eviction, no scheduler wait, no dropped connections. It does not write to the CPU metrics HPA reads as a scaling signal, so both autoscalers run cleanly on the same deployment. Rightsized pods then consolidate onto fewer nodes, with savings surfaced per namespace, team, and workload in real time.
Three gaps VPA was not built to close
These are the constraints teams hit in production, not edge cases.
pod-api · StatefulSet
VPA
req
DevZero
req
pod-api-7x → pod-9f
evicted
pod-api-7x same pod
running
Stateful workload safety
VPA evicts pods to apply new resource values. For StatefulSets with large volumes or Java heaps, each eviction is a multi-minute outage. DevZero resizes in place with no disruption.
no attribution$340 saved
$300$200$100$0
VPA
Monthly spend by namespace
DevZero
per namespace and team
Cloud cost visibility
VPA has no concept of node cost or cloud spend. It optimizes resource requests, not dollars. DevZero maps every rightsizing decision to cost reduction by namespace and team.
HPA autoscaler pairing
with VPAwith DevZero
HPA
VPA
HPA
DevZero
CPU metric conflict
no metric interference
safe alongside HPA
safe alongside HPA
HPA conflict avoidance
Running VPA and HPA on the same workload creates a CPU feedback loop. DevZero applies vertical changes without touching the metrics HPA reads, so both run cleanly.
What DevZero adds that VPA cannot
The controls your team actually needs, not a dashboard that requires a PhD to interpret.
Before
80m
No restart
fired
fired
After
320m
CPU and memory update on the running pod via cgroup. Changes land in under 30 seconds. No eviction, no scheduler cycle.
P95 forecast
$300$200$100$0
MonTueWedThuFriSat
XGBoost models forecast P90/P95 usage 24 hours ahead. Pods consolidate onto fewer nodes before over-provisioning accumulates.
H100
L40
V100
A100
DevZero targets GPU-specific resources
35%
GPU utilization: 35% actual vs 100% requested
65% idle — reclaimed automatically
DevZero targets GPU-specific resources across H100, L40, V100, and A100. GPU jobs consuming a fraction of their request are reclaimed automatically.
VPA vs DevZero at a glance
DevZero complements Karpenter. Every row marked unsupported for VPA is a layer DevZero adds on top.
Capability
DevZero
VPA
CPU and memory rightsizing
In-place resize (no pod restart)
HPA compatibility
Live migration for stateful workloads
GPU rightsizing
Node consolidation and bin-packing
Extended workload support
~
Policy governance
Cost visibility and attribution
Security scanning
Installation model
~
Multi-cloud support
~
What our customers say
Frequently asked questions
Most clusters are overprovisioned.Let's prove yours is.
Run a free assessment to identify overprovisioned workloads, idle capacity, and your potential savings, in minutes.
Optimize now