My team works on on-prem clusters with fixed, not elastic, capacity. We got tired of the Sisyphean task of getting developers to set their resource requests and limits correctly, so we came up with ballast.
Its job is to watch the IRL behaviors of workloads, develop a historical profile, and then either adjust pods on admission or in-place (CPU/memory only).
It's not for eviction - descheduler is a better tool for that. But it solves the immediate need of better workload packing on constrained nodes.
Looking for feedback and (ideally) beta testers who can provide constructive input on how to improve the project.
My team works on on-prem clusters with fixed, not elastic, capacity. We got tired of the Sisyphean task of getting developers to set their resource requests and limits correctly, so we came up with ballast.
Its job is to watch the IRL behaviors of workloads, develop a historical profile, and then either adjust pods on admission or in-place (CPU/memory only).
It's not for eviction - descheduler is a better tool for that. But it solves the immediate need of better workload packing on constrained nodes.
Looking for feedback and (ideally) beta testers who can provide constructive input on how to improve the project.