[Feature Request] Add support for CPU/Memory ratio constraint in VPA

[Jrmy2402]

Which component are you using?:
/area vertical-pod-autoscaler

Describe the solution you'd like:
I’m wondering if there is any interest in adding a feature to Vertical Pod Autoscaler (VPA) to allow users to define a constraint or ratio between CPU and Memory resources (for example, always keep a 1:2 CPU:RAM ratio).

Additional context:
Currently, VPA provides recommendations for CPU and Memory independently, based on observed usage. However, in some use cases (for example, when workloads require a fixed ratio between CPU and RAM for optimal performance or licensing reasons), it would be useful to ensure that VPA maintains a specific ratio between these resources.

Questions:
Has this feature been discussed before?
Would other users find it valuable to have a way to configure a CPU:Memory ratio in VPA recommendations?
Are there known workarounds or best practices in the community for enforcing such a constraint today?

If there is community interest, I’d be happy to discuss use cases and possible implementation approaches.

Thanks!

[vitanovs]

Hi @Jrmy2402

Since you've mentioned that such a ratio will be useful with InPlaceOrRecreate, could you please provide more details around it ? I am interested in knowing if you've encountered any side effects and how does it affect the in-place update patch(es).

Thanks!

[Jrmy2402]

Hi @vitanovs,

Today I used a home-made "VPA" to do vertical scaling of my pods, but since VPA now supports InPlaceOrRecreate, I'm interested in using it instead of my custom "VPA".
For my needs, I need to maintain a ratio between RAM and CPU.
For now, I can't use VPA without that, but during my POC I didn’t encounter any side effects, it works fine :)

[iamzili]

@Jrmy2402 - may I ask why you specifically want to maintain the ratio between CPU and memory requests? one of the reasons you mentioned is licensing - is this something that actually exists out there? have you encountered this type of requirement before?

I can relate to your second reason which is about ensuring optimal performance. E.g. if an application's documentation states that it should use 4 GB of memory per CPU core, then it makes sense to follow that guidance. In such cases we could configure the VPA to provide recommendations only for the CPU request, and then adjust the memory request based on a user defined ratio.

[Jrmy2402]

@iamzili The examples I gave earlier were more general ideas of potential use cases for such a feature.
In my specific case, we want to maintain a fixed CPU/memory ratio for our services because our internal billing system is based on that ratio.

I’ve implemented this using a mutating admission webhook that automatically adjusts memory when CPU changes.
However, one issue I’m facing with this webhook-based approach is that the VPA updater continuously attempts to resize pods that have scaled, because it considers the memory value to be incorrect.

To avoid the updater performing these repeated resizes, I must configure the VPA to provide recommendations only for the CPU request. This works fine in that case, but I lose automatic scaling for memory.

This is why I think it might make more sense to have the ability to define a CPU/memory ratio directly in the VPA recommender, so that both CPU and memory can be scaled together consistently.

[Jrmy2402]

I’ve just opened a PR to propose the implementation of the feature discussed.
This PR adds the --enforce-cpu-memory-ratio flag to the VPA recommender, allowing users to enforce a fixed memory-to-CPU ratio (expressed in bytes per millicore) across all recommendations.

Feedback and suggestions are welcome 🙂