Autoscaling · CuraDevOps

tag: Autoscaling · 7 items

Platform/SRE — Learn: Useful framing on how GPU-bound and bursty AI workloads strain the classic HPA/VPA/cluster-autoscaler division; no version change or deadline, but informs how you’d architect autoscaling for AI teams onboarding to the cluster.
CI/CD — Skip
Leader — Learn: Useful context for platform strategy as teams push AI/ML workloads onto shared Kubernetes; no near-term decision or cost/vendor change implied, but shapes how you’d evaluate autoscaling capabilities when reviewing cluster investments.

Platform/SRE — Learn: LiveTaskCount covers tasks in ACTIVATING, RUNNING, and DEACTIVATING states, making it a more complete signal for autoscaling policies than RunningTaskCount; worth evaluating when next tuning ECS autoscaling or capacity planning, but no existing configuration requires updating.
CI/CD — Skip
Leader — Skip

Platform/SRE — Plan: If you run LLM inference workloads on SageMaker HyperPod, this GA feature cuts scale-out time ~60% by pre-loading model weights to NVMe and pre-pulling container images; evaluate adding a modelCacheConfig block to InferenceEndpointConfig resources this quarter.
CI/CD — Skip
Leader — Learn: HyperPod’s new model caching substantially narrows the cold-start penalty for large-model inference, which is relevant context if evaluating HyperPod as the org’s managed inference platform versus self-managed alternatives.

Platform/SRE — Learn: Beta (pre-GA) feature that eliminates the need for external add-ons to scale queue consumers and GPU/CPU batch workloads to zero; worth evaluating in non-production clusters now, with a plan to adopt when it reaches GA.
CI/CD — Skip
Leader — Learn: The scale-to-zero HPA capability has meaningful cost-reduction potential for GPU and dedicated-CPU workloads, but it is still Beta; track for inclusion in platform cost-optimization standards when it reaches GA.
Signals: Kubernetes 1.37 EOL 2027-10-28

Platform/SRE — Learn: Covers a real incident pattern — GPU pods pending during traffic spikes — and predictive scaling approaches; worth reading to inform GPU cluster design, but no GA tool, deadline, or breaking change anchors an action now.
CI/CD — Skip
Leader — Skip

Platform/SRE — Learn: EKS Provisioned Control Plane clusters now get significantly faster HPA-driven scaling with no configuration changes required — worth knowing if you run large clusters with many HPA objects, but no action needed.
CI/CD — Skip
Leader — Skip

Platform/SRE — Learn: A solid walkthrough for extending HPA with custom signals (queue depth, connection counts) via a Prometheus-compatible exporter — useful design reference but no operational change required to existing clusters.
CI/CD — Skip
Leader — Skip