Yaya Horizontal Pod Autoscaler ke aiki, kuma yaya yake da alaƙa da Cluster Autoscaler?

Question

IT Interview Practice · Accepted Answer

**Horizontal Pod Autoscaler (HPA)** yana ƙara ko cire **Pod replicas** bisa metrics da aka lura (CPU, memory, ko na musamman). **Cluster Autoscaler (CA)** yana ƙara ko cire **nodes** lokacin da ba a iya tsara Pod ko ba a amfani da nodes sosai. Suna aiki a matakai daban kuma suna aiki tare.

## Autoscalers biyu, matakai biyu

```text
HPA  → yana scale PODS  (replicas na Deployment fiye/ƙasa)     — yana amsa LOAD
CA   → yana scale NODES (yana faɗaɗa/rage injunan cluster)     — yana amsa Pod na PENDING
VPA  → (ƙari) yana daidaita requests/limits na Pod             — yana amsa girman daidai

Yawan gudana: load ya tashi → HPA ya ƙara Pod → Pod ba su dace → CA ya ƙara node →
              Pod suna tsara. Load ya faɗi → HPA ya cire Pod → node babu aiki → CA ya cire shi.
```

## Yadda HPA control loop ke aiki

```text
Kowane ~15s HPA controller:
  1. yana karanta metric na yanzu (misali matsakaicin CPU a fadin Pod) daga metrics-server / custom API
  2. yana ƙididdiga:  desiredReplicas = ceil(currentReplicas * currentMetric / targetMetric)
  3. yana scale Deployment zuwa gare shi (an iyakance da minReplicas/maxReplicas)
Misali: Pod 4 a 80% CPU, manufa 50% → 4 * 80/50 = 6.4 → replica 7.
Stabilization windows suna dakushe kaɗawa (musamman a scale-down).
```

## Manifest na HPA

```yaml
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: web-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: web
  minReplicas: 3
  maxReplicas: 20
  metrics:
    - type: Resource
      resource:
        name: cpu
        target:
          type: Utilization
          averageUtilization: 60      # kiyaye matsakaicin CPU ~60% na REQUEST
  behavior:                           # na zaɓi: daidaita saurin scale up/down
    scaleDown:
      stabilizationWindowSeconds: 300
```

```bash
kubectl get hpa                       # metric na yanzu/manufa, adadin replica
kubectl top pods                      # yana buƙatar metrics-server
```

## Muhimman dependencies & gargaɗi

```text
• HPA a kan CPU/memory YANA BUƘATAR an girka metrics-server; utilization dangi da REQUESTS ne
  → idan requests ba a saita/kuskure ba, lissafin HPA ba shi da ma'ana. Saita requests daidai.
• Custom/external metrics (RPS, zurfin layi) suna buƙatar adapter (Prometheus Adapter, KEDA).
• HPA da `replicas` na hannu suna cin karo — bari HPA ya mallaki adadin replica.
• Cluster Autoscaler yana buƙatar requests daidai kuma: yana kallon REQUESTS na Pod na Pending don
  yanke scaling na node. Node scale-up yana ɗaukar mintuna (samarwa), don haka kiyaye sarari.
```

## Me ya sa yake da muhimmanci

Autoscaling shine abin da ke bari Kubernetes ya daidaita ƙarfi da buƙata kai tsaye, kuma fahimtar matakan biyu ilimi ne na matakin senior wanda ke raba clusters masu inganci na kuɗi da masu jurewa daga waɗanda aka wuce ko aka rage samar da su. Mahimmancin tsarin tunani shine cewa **HPA yana scale Pods** don amsa load (yana daidaita adadin replica daga CPU, memory, ko metrics na musamman da aka lura) yayin da **Cluster Autoscaler yana scale nodes** don amsa matsin scheduling (yana ƙara injuna lokacin da Pod suke Pending, yana cire su lokacin da nodes babu aiki) — suna aiki a matakai daban kuma suna haɗin kai: HPA zai iya ƙara Pod kawai idan akwai wuri, kuma CA yana ba da wannan wuri. Fahimtar HPA control loop — tsari na `desiredReplicas = ceil(currentReplicas * currentMetric / targetMetric)`, reconciliation na lokaci-lokaci, da stabilization windows waɗanda ke hana kaɗawa — yana bari ka saita ka kuma gyara shi maimakon ka ɗauke shi a matsayin sihiri. Gargaɗin su ne inda hukuncin senior ke bayyana: HPA a kan CPU/memory **yana buƙatar metrics-server** kuma yana ƙididdiga utilization **dangi da requests**, don haka requests marasa daidai ko marasa saitawa suna sa lissafinsa ba shi da ma'ana; metrics na musamman (requests-per-second, zurfin layi) suna buƙatar adapter kamar Prometheus Adapter ko KEDA; kuma node scale-up yana ɗaukar mintuna, don haka sarari yana da muhimmanci. Saboda girman ƙarfi daidai yana shafar aminci kai tsaye ƙarƙashin karuwar zirga-zirga da kashe kuɗin cloud a lokutan natsuwa, kuma saboda autoscalers biyu suna dogara ga resource requests da aka saita daidai, ƙware HPA da Cluster Autoscaler na tushe ne don gudanar da production Kubernetes cikin inganci da aminci.

Autoscalers biyu, matakai biyu

Yadda HPA control loop ke aiki

Manifest na HPA

Muhimman dependencies & gargaɗi

Me ya sa yake da muhimmanci

Tambayoyi masu alaƙa