Modern applications need to be anti-fragile and support hyperscale, which requires elasticity at the workload level (horizontal/vertical) and the infrastructure level (cluster scaling).
Autoscaling relies on events for cluster-level actions and a multi-layered metrics API for workload scaling based on resource, custom, or external data sources.
The Horizontal Pod Autoscaler (HPA) can scale pods based on simple resource metrics like CPU, custom pod metrics, or external metrics from Prometheus.
The Vertical Pod Autoscaler (VPA) can automatically adjust pod resources, but its recommendation mode is most useful for determining optimal CPU and memory settings.
The default cluster autoscaler automatically provisions new nodes when it detects unschedulable pods due to resource constraints, as demonstrated on Google Kubernetes Engine.
Carpenter provides a fast and flexible cluster autoscaling solution for AWS EKS, enabling cost optimization by using spot instances for scaled-out nodes.
KEDA (Kubernetes Event-driven Autoscaling) enables scaling workloads, including to zero, based on events from various sources like message queues or databases.
A recap of essential autoscaling components including the metric server, HPA, VPA, cluster autoscalers like Carpenter, KEDA, and the descheduler for cluster optimization.
Discussion on the production-readiness of autoscalers, the importance of observability, and how to achieve graceful pod termination during scale-down events.
21:43Mario Valderrama
42:45Marc Nimmerrichter
28:12Josip Stuhli
57:24Hannes Norbert Göring
27:52Maish Saidel-Keesing
50:51Alex Soto
29:21Um e Habiba
28:03Federico Fregosi
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
smartclip Europe GmbH
Hamburg, Germany
Mittwald CM Service GmbH & Co. KG
Espelkamp, Germany
iits-consulting GmbH
München, Germany
AUTO1 Group SE
Berlin, Germany
evoila Frankfurt GmbH
Mainz, Germany
