4.4 KiB
title, type, source-type, category, tags, date-added, video-source, audio-source, status
| title | type | source-type | category | tags | date-added | video-source | audio-source | status | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Public Cloud Learning Sessions - EKS Optimization part 1 of 3 - Compute Optimization with Karpenter - 20250204 170113-Meeting Recording | cloud-learning | video | DevOps & SRE/04_EKS |
|
2026-04-14 | nas:///volume2/work/Public Cloud Learning Sessions/Public Cloud Learning Sessions - EKS Optimization part 1 of 3 - Compute Optimization with Karpenter - 20250204_170113-Meeting Recording.mp4 | summarized (Gemini 摘要) |
Public Cloud Learning Sessions - EKS Optimization part 1 of 3 - Compute Optimization with Karpenter - 20250204 170113-Meeting Recording
Source: NAS /volume2/work/Public Cloud Learning Sessions/Public Cloud Learning Sessions - EKS Optimization part 1 of 3 - Compute Optimization with Karpenter - 20250204_170113-Meeting Recording.mp4
Type: VIDEO | Category: 04_EKS
Status: 🟡 Awaiting Whisper transcription → Summary
EKS Optimization with Carpenter
This session introduces Carpenter, an open-source compute infrastructure management tool for Kubernetes clusters, addressing challenges associated with the traditional Cluster Autoscaler. Carpenter offers native integration with Kubernetes, direct EC2 fleet API communication, and intelligent workload placement and consolidation based on cost and utilization.
Key differences between Carpenter and Cluster Autoscaler:
- Carpenter integrates with Kubernetes workload scheduling constructs.
- It directly communicates with the EC2 fleet API, reducing latency.
- It provides native experiences for workload placement and node consolidation.
Two core components of Carpenter: node pools and node classes. Node pools define scheduling constraints and capacity limits, while node classes define instance provisioning details like subnets, node roles, and AMIs.
Carpenter supports Kubernetes scheduling constraints like node selectors, affinity, taints, tolerations, and topology spread, along with AWS placement requirements such as purchasing options, processor architectures, and availability zones. It can identify zonal requirements based on volume claims and storage classes, simplifying workload definitions compared to Cluster Autoscaler.
Carpenter has native integration with Kubernetes and it complements the native Kubernetes spot pod scheduling constraints that is available for your workloads.
Carpenter natively supports spot interruptions without requiring additional components like the node termination handler. It uses EventBridge and SQS to handle spot interruption notifications, instance rebalance notifications, health events, and instance state change events.
Node pools can be designed for various scenarios, including single node pools, mixed compute/accelerated nodes, or isolated node pools based on cost, security, or multi-tenancy. Weighted node pools can prioritize instances based on existing commitments or reservations.
Carpenter simplifies data plane management by removing pain points associated with node groups, integrating node termination handlers, and providing native integration with Kubernetes scheduling constraints. It also helps consolidate compute instances for greater cost efficiency.
Carpenter not only does the auto-scaling bit, but it also removes the pain points of working with node groups.
Carpenter can automatically upgrade AMIs or use defined AMIs, referring to the parameter store for the latest EKS optimized AMIs for the corresponding control plane version. It identifies drifts between the desired state and running machines, rolling out changes in a rolling upgrade fashion.
AMI selection can be pinned to specific versions or use custom AMIs. The AMI family setting tells Carpenter what user data to inject when spinning up instances.
Consolidation policies can be configured with fine-grained budgets, such as preventing consolidation during peak business hours or limiting the percentage of instances disrupted at a time.
Carpenter publishes logs and emits Prometheus metrics for observability, with community-maintained dashboards available for visualization.
Onboarding is simple, requiring Carpenter to be deployed on nodes not managed by Carpenter, such as a small node group or Fargate instances. Migration guides are available for migrating from Cluster Autoscaler.
The session is the first in a series of three, with subsequent sessions covering the Bottlerocket operating system and EKS Auto Mode.