DEVOURED
Why Kubernetes autoscalers keep fighting each other (and why the solution must be autonomous)
Independent Kubernetes autoscalers like HPA, VPA, and Karpenter are fighting because they lack a unified 'central nervous system' to manage resources.
What: Nikhil Kurup of Sedai highlights four failure modes—such as the VPA/HPA death spiral—caused by decoupled autoscalers, and argues for a singular, predictive control plane.
Why it matters: Modern clusters treat autoscaling as separate reflexive actions rather than a single optimization problem, leading to thrashing, inflated cloud costs, and performance instability.
Takeaway: If you are running multiple autoscalers, conduct a thorough audit of your resource requests versus actual usage to see if they are conflicting, as current tools often mask inefficiency.
Deep dive
- Autoscalers act as independent control loops that mutate the same fields (replicas and resource requests) without awareness of each other.
- The VPA/HPA death spiral: VPA lowers requests to save money, triggering HPA to spawn more pods, leading to smaller, more inefficient containers.
- KEDA and HPA can conflict when event-based scaling competes with CPU-based utilization metrics.
- GitOps tools can create 'drift ping-pong' if they try to reset auto-scaled resources to a static Git-defined baseline.
- Node autoscalers like Karpenter are often misled by highly inflated (over-provisioned) container requests.
- A unified autoscaler must ingest all metrics and compute a single action vector across horizontal and vertical scaling simultaneously.
Decoder
- Thrashing: A condition where a system spends more resources on managing state changes (scaling up/down) than on doing useful work.
- Death Spiral: A feedback loop where independent controller actions force the system into progressively worse states (e.g., smaller pods + more pods = more overhead/instability).
- SLA (Service Level Agreement): A formal commitment between a service provider and customer regarding performance, uptime, and reliability.
Original article
Full article content is not available for inline reading.
Explore all 54 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi