What is Wrong with DaemonSet and Why Should You Care?
summary
DaemonSet often breaks expectations: one cluster shows log producers at 1 MB/s vs 50 MB/s with collectors needing 50m CPU vs 1 core, so a single DaemonSet can’t set sane requests. VPA doesn’t account per-node and historically has a 2023 bug around eviction (a PR was needed to avoid evicting all DaemonSet pods), and PodDisruptionBudget can’t count DaemonSet replicas (issue opened 2022). Taints/tolerations misfire, hostNetwork blocks rolling updates, and tools like OpenKruise can pre-download images while Datadog’s DaemonSet controller adds canary logic but each has tradeoffs.
AI-written summary · about
chapters (10)
- 0:00DaemonSets and platform engineering
- 1:08What is a DaemonSet
- 3:56DaemonSet controller and nodes
- 5:44Resource consumption problems
- 9:20Vertical Pod Autoscaler limits
- 11:25Taints and tolerations challenges
- 14:23Rolling update issues
- 17:15Solutions for DaemonSet problems
- 18:40Summary and future outlook
- 19:44Q&A
related talks
36mThe Next Generation of DaemonSet Autoscaling
20mDeployments vs StatefulSets vs Daemonsets
36mKubernetes Confessions: Real Mistakes We’ve All Made (and What They Teach Us)
31mGhosts in the Runtime: Who Ate My Capabilities & Other Mysteries
28mThe Daemonset in the Details: Pitfalls of OpenTelemetry Collector Deployments
20m10 Common Things People Do Wrong in K8s Environments
26mOptimizing Kafka Consumers with Kubernetes and KEDA...
16mUnconference Session: Reliable K8s Operator: How to Survive Eventual Consistency