Allow VPA to skip eviction when in-place update fails

[Jrmy2402]

Which component are you using?:

/area vertical-pod-autoscaler

Is your feature request designed to solve a problem? If so describe the problem this feature should solve.:

I’m using the Vertical Pod Autoscaler on StatefulSets with a single replica.
When an in-place update fails, VPA currently falls back to evicting and recreating the Pod.
This causes downtime, which is unacceptable for stateful workloads that can’t tolerate disruption.
Ideally, I’d like VPA to simply skip the update and retry later instead of evicting.

Describe the solution you'd like.:

It would be great to have an option (either a new updateMode or a flag in updatePolicy) to tell VPA:
“If in-place update fails, do not evict or recreate the Pod, just retry later.”

Describe any alternative solutions you've considered.:

I tried setting spec.updatePolicy.minReplicas to 2, but since my StatefulSet only has 1 replica, this prevents VPA from performing any in-place updates at all.
So while it avoids eviction, it also stops updates from happening, which doesn’t fit my use case.

[maxcao13]

Sounds like a reasonable use case to me, thanks for opening this.

I'm more privy to a new update mode called InPlace. Mostly because as InPlacePodVerticalScaling node feature continues to mature, it probably makes since that we have an updateMode naming that aligns the most with in-place resizing logic, at least for intuitiveness for the user. Looking back, InPlaceOrRecreate may not be intuitive at first glance, and is a combination of 2 different features/updateModes basically.

In fact, I would probably say if we started the other way around where we implemented InPlace update mode, and allowed a configurable vpa-level toggle to fallback to recreate optionally, then that would've been better? But it's a bit late for that.

[omerap12]

Thanks for opening this!
We’ve gathered enough input to move forward.
Next steps: draft an AEP, then open a PR to implement it.
/triage accepted