L06 — Scheduling & Scaling

Once pods exist, two questions: where should this pod run, and how many should I have? L06 covers both — the scheduling primitives (where Pods land) and the scaling family (how many Pods run, how much they get).

What you’ll understand after this level

• The kube-scheduler flow: PreFilter → Filter → PreScore → Score → Reserve → Permit → PreBind → Bind
• Taints and tolerations — keeping pods off (or onto) specific nodes
• Node affinity / pod anti-affinity — schedule based on labels
• Topology spread constraints — spread replicas across zones/nodes
• PriorityClass and preemption — the only signal the scheduler uses to evict a lower-priority Pod
• Scheduling gates — hold a Pod back from scheduling until an external signal
• Resource requests vs limits — what each does, QoS classes, cgroups, the limits debate
• HPA (horizontal scale replicas), VPA (vertical resize requests), Cluster Autoscaler + Karpenter (add nodes), KEDA (event-driven) — what each does and how they fit
• PodDisruptionBudgets — keep services available during voluntary disruption
• Restart policies — Always, OnFailure, Never and when each applies
• Extended resources — GPUs, FPGAs, and the device plugin model

Notes in this level

Scheduling primitives

Note	Status	What’s in it
Scheduling	✅	Taints, tolerations, node/pod affinity, anti-affinity, topology spread, all the operator semantics
Priority & Preemption	✅	PriorityClass, preemption algorithm, system classes, the PD deadlocks, QoS vs priority
Scheduler Internals	✅	The plugin pipeline, every default plugin, profiles, framework extensions, perf tuning
Scheduling Gates	✅	Pod scheduling readiness, holding Pods back, the StatefulSet join pattern
Extended Resources	✅	GPUs, device plugins, time-slicing, MIG, DRA, ResourceClaim, the integer rule

Resources and constraints

Note	Status	What’s in it
Resource Requests & Limits	✅	CPU/memory/ephemeral-storage, CFS throttling, OOM-kill, QoS classes, cgroup v2, the limits debate
Restart Policy	✅	Always / OnFailure / Never, the backoff algorithm, CrashLoopBackOff, exit codes, Job/CronJob behavior

Scaling family

Note	Status	What’s in it
Scaling — overview	✅	The L06 hub: HPA / VPA / Karpenter / CA / KEDA at a glance, how they combine
HPA	✅	The autoscaling control loop, custom / external metrics, behavior settings, scaling math, the HPA controller
VPA	✅	VPA modes (Off / Initial / Auto), the recommender, VPA + HPA coexistence, the OOM pattern
Karpenter	✅	NodePools, EC2NodeClass, consolidation, spot, the modern alternative to Cluster Autoscaler
Cluster Autoscaler	✅	ASG / MIG / VMSS, scale-up and scale-down logic, the CA vs Karpenter decision
KEDA	✅	Event-driven autoscaling, 60+ scalers, scale to zero, the external metrics API
PodDisruptionBudget	✅	minAvailable / maxUnavailable, the eviction API, the HPA + PDB deadlock, unhealthyPodEvictionPolicy

Suggested reading order

Scheduling path

1. Resource Requests & Limits — the foundation; everything else assumes you have this
2. Scheduling — the YAML-level primitives (taints, affinity, topology)
3. Restart Policy — short, foundational
4. Scheduler Internals — the framework that enforces the above
5. Priority & Preemption — when scheduling fails, what happens
6. Scheduling Gates — advanced: hold a Pod back from scheduling
7. Extended Resources — GPUs and the device plugin model

Scaling path

1. Scaling — overview — at-a-glance comparison of all the scaling primitives
2. HPA — the workhorse
3. VPA — the right-sizing complement
4. PDB — read before you start draining nodes
5. Cluster Autoscaler — the older node provisioner
6. Karpenter — the modern alternative
7. KEDA — event-driven, the right answer for queue-based workloads

Where to go next

→ L07 — Security: with workloads scheduled and scaled, decide who can do what to them.