🖼️ See the drawing + lab 🏠 Course home 🌿 Branch on GitHub ✏️ View source

🖼️ आकृती आणि labThe drawing + lab पूर्ण पानावर उघडा ↗Open full page ↗

🏫🏫 धडा 15 — Multi-AZ आणि scaling ची शिडी

📍 तुम्ही इथे आहात: धडा 15 / 26 — भाग 2 इथे संपतो · मागील: lesson-14-deploy-gitops · पुढील: lesson-16-debugging — भाग 3 सुरू!

📦 या ब्रँचमध्ये काय आहे

आधीचे सर्व धडे, आणि प्रत्येक खऱ्या cluster ला उत्तर द्यावे लागणारे दोन प्रश्न: संपूर्ण इमारतच जळून गेली तर? (multi-AZ) आणि निकालाच्या दिवशी काय होते? (scaling ची शिडी). खऱ्या files:

terraform/vpc.tf — अनेक AZs मध्ये बनवलेले subnets (हाच आराखडा सुरुवातीपासून होता!)
terraform/eks.tf — node group त्या subnets मध्ये पसरलेला आहे
k8s/hpa.yaml — शिडीची पायरी 1 (धडा 09)

🧒 5 वर्षांच्या मुलाला समजावल्यासारखे

आगीचा प्रश्न. 🏫🔥 प्रत्येक मुलाला एकाच इमारतीत बसवणारी शाळा, एका आगीनंतर शाळाच उरणार नाही अशा स्थितीत असते. खऱ्या शाळा campus वापरतात: इमारती A आणि B (availability zones — वेगळी वीज, सगळेच वेगळे), आणि एक नियम: "संपूर्ण वर्ग कधीच एका इमारतीत बसवू नका." इमारत A जळली? दुःखाचा दिवस, पण वर्ग B मध्ये सुरू राहतात. हेच multi-AZ.

परीक्षेच्या दिवसाचा प्रश्न. 📝 निकालाच्या दिवशी हजार पालक एकदम येतात. शाळा एक शिडी चढते, आधी सर्वात स्वस्त पायरी:

प्रत्येक खोलीत जास्त मदतनीस 🙋 (HPA, धडा 09): transport manager pods वाढवतो — काही सेकंद, मोफत, पण बाकांवर जागा असेपर्यंतच चालते.
जास्त बाक 🖥️ (cluster autoscaler): pods ना बसायला जागा नाही, ते उभे आहेत (Pending!) → तो ASG कडे आणखी EC2 बाक मागतो — एक-दोन मिनिटे लागतात, पैसे लागतात, पण खरी मर्यादा सोडवतो.
जुळ्यांना वेगळे बसवा 👯 (topology spread): school-api च्या सगळ्या copies एकाच इमारतीत ठेवू नका — 2 replicas = 2 इमारती. आगीचा नियम आणि scaling ची गोष्ट इथे भेटतात.
सगळे एकदम जेवायला नाही 🍽️ (PodDisruptionBudget): बाकांची हलवाहलव (upgrades, scale-down) चालू असताना किमान N मदतनीस कामावर राहिलेच पाहिजेत. स्वेच्छेने केलेली हलवाहलव थांबते; वर्ग कधीच एकटा पडत नाही.

🗺️ आकृती

flowchart TB
    subgraph campus["🏞️ the campus - one region, one VPC"]
        subgraph aza["🏫 building A - AZ a"]
            n1["🖥️ desk<br/>🪑 api-1 · 🪑 analytics-1"]
            n3["🖥️ new desk<br/>added by autoscaler"]
        end
        subgraph azb["🏫 building B - AZ b"]
            n2["🖥️ desk<br/>🪑 api-2 · 🪑 analytics-2"]
        end
    end
    alb["🌍 ALB - stands OUTSIDE both buildings,<br/>sends visitors to whichever is healthy"]
    ladder["🪜 the ladder:<br/>1 HPA: more pods - seconds<br/>2 autoscaler: more desks - minutes<br/>3 spread: copies across buildings<br/>4 PDB: never all away at once"]
    alb --> aza
    alb --> azb
    ladder -.-> campus

❓ काय

AZ = region मधले प्रत्यक्ष वेगळे data center (AWS कोर्स, धडा 07). EKS control plane आपोआप multi-AZ असतो — AWS office इमारतींमध्ये पसरवून चालवते; तुम्ही फक्त बाकांची बाजू design करता.
Multi-AZ nodes: node group ला ≥2 AZs मधले subnets द्या (vpc.tf + eks.tf हेच करतात) — मग ASG बाक इमारतींमध्ये पसरवतो आणि बदलतो.
Cluster autoscaler (किंवा त्याचा नवा चुलत भाऊ Karpenter): Pending pods वर लक्ष ठेवतो → ASG वाढवतो; बाक रिकामे बसले की कमी करतो. HPA pods हलवतो, autoscaler बाक हलवतो — जोडीने गाणे, प्रतिस्पर्धी नाहीत.
topologySpreadConstraints — Deployment च्या pod spec मध्ये जोडा:

topologySpreadConstraints:
  - maxSkew: 1                                  # buildings may differ by ≤1 copy
    topologyKey: topology.kubernetes.io/zone    # spread BY BUILDING
    whenUnsatisfiable: ScheduleAnyway           # prefer, don't deadlock
    labelSelector: { matchLabels: { app: school-api } }

PodDisruptionBudget — 2-replica app साठी minAvailable: 1: node drains आणि upgrades एका वेळी एकच बाक करत पुढे जातात, वर्ग कधीच रिकामा होत नाही.
data विसरू नका: pods सहज पसरतात; EBS drawers प्रत्येक इमारतीचे वेगळे असतात (एक volume एकाच AZ मध्ये राहतो!) — हा repo state RDS मध्ये ठेवतो याचे आणखी एक कारण, ज्याचा स्वतःचा multi_az switch आहे.

🤔 का

AZ outage कधी होणार हा प्रश्न आहे, होईल का हा नाही — आणि ते टिकणे फक्त तेव्हाच शक्य आहे जेव्हा replicas, nodes आणि traffic आगीच्या आधीच इमारतींमध्ये पसरलेले असतात. Multi-AZ शिवायचे scaling म्हणजे वेगवान single point of failure; autoscaling शिवायचे multi-AZ म्हणजे पहिल्याच परीक्षेच्या दिवशी कोसळणारी मजबुती. शिडी हेच तुम्ही शिकलेल्या सगळ्याचे सार आहे: धडा 03 चा monitor, धडा 09 च्या बसेस, AWS कोर्सचे ASG बाक — सगळे एकत्र चढताना.

🧪 करून पाहा

# which building is each desk in? (works on EKS; minikube has 1 "building")
kubectl get nodes -L topology.kubernetes.io/zone

# where do the copies sit right now?
kubectl -n school get pods -o wide

# rung 2, felt: ask for more pods than the desks can seat →
kubectl -n school scale deployment school-api --replicas=20
kubectl -n school get pods | grep Pending | head -5
# Pending = kids standing without desks — EXACTLY what the cluster
# autoscaler watches for. (On EKS with it installed: new nodes in ~2 min.)
kubectl -n school scale deployment school-api --replicas=2   # calm down

# rung 4: protect the class during desk-shuffles:
kubectl -n school create pdb school-api --selector=app=school-api --min-available=1
kubectl -n school get pdb    # ALLOWED DISRUPTIONS: 1 — never both at once 🎉

🔧 या धड्यासाठी kubectl

kubectl get nodes -L topology.kubernetes.io/zone · kubectl get pdb

🔗 यावर आधारित: 🪪 AWS L07 (AZs म्हणजे इमारती) · L12 (ASG)

✅ तपासा — तुम्हाला काय दिसले पाहिजे

kubectl get pods -o wide मध्ये copies वेगवेगळ्या nodes/zones वर दिसतात; बसू शकतील त्यापेक्षा जास्त pods मागा → autoscaler बाक जोडेपर्यंत Pending (EKS) — minikube वर ते Pending च राहतात.

🧹 साफसफाई

scale परत कमी करा; EKS वर जास्तीचे nodes/cluster त्याच दिवशी destroy करा

⚠️ नेहमीच्या चुका

'सध्यापुरता' एकच AZ
minAvailable = replicas असलेला PDB — drain कधीच पूर्ण होऊ शकत नाही

🎓 campus पूर्ण झाला

स्वतः बरे होणारे pods (03), पावसाळी दिवसाच्या बसेस (09), downtime शिवाय बदली (11), deploy करणारे robots (14) — आणि आता आग आणि निकालाचा दिवस दोन्ही टिकणारी शाळा. हा आता demo cluster राहिला नाही; हे architecture आहे. 🏫🏫

आणि आता ते उभे आहे म्हणून: भाग 3 — ते खरोखर चालवणे (धडे 16–26): debugging, hall passes, गृहपाठाच्या घंटा, बसण्याचे नियम, आणि सगळ्यावर नजर.

git checkout lesson-16-debugging   # 🩺 the nurse's triage chart

🏫🏫 Lesson 15 — Multi-AZ & the scaling ladder

📍 You are here: Lesson 15 of 26 — Part 2 ends here · Previous: lesson-14-deploy-gitops · Next: lesson-16-debugging — Part 3 begins!

📦 What's in this branch

All previous lessons, plus the two questions every real cluster must answer: what if a whole building burns down? (multi-AZ) and what happens on exam-results day? (the scaling ladder). Real files:

terraform/vpc.tf — subnets created in multiple AZs (this was the plan all along!)
terraform/eks.tf — the node group spans those subnets
k8s/hpa.yaml — rung 1 of the ladder (lesson 09)

🧒 Explain like I'm 5

The fire problem. 🏫🔥 A school that puts every kid in ONE building has one fire away from no school at all. Real schools use a campus: buildings A and B (availability zones — separate power, separate everything), and one rule: "never seat the whole class in one building." Building A burns? Sad day, but classes continue in B. That's multi-AZ.

The exam-day problem. 📝 On results day, a thousand parents arrive at once. The school climbs a ladder, cheapest rung first:

More helpers per room 🙋 (HPA, lesson 09): the transport manager adds pods — seconds, free, but only works while desks have space.
More desks 🖥️ (the cluster autoscaler): pods are standing with nowhere to sit (Pending!) → it asks the ASG for more EC2 desks — takes a minute or two, costs money, solves the real limit.
Spread the twins 👯 (topology spread): don't put all school-api copies in one building — 2 replicas = 2 buildings. The fire rule and the scaling story meet here.
Never everyone at lunch at once 🍽️ (PodDisruptionBudget): during desk-shuffling (upgrades, scale-down), at least N helpers must stay on duty. Voluntary moves wait; the class is never unattended.

🗺️ Diagram

flowchart TB
    subgraph campus["🏞️ the campus - one region, one VPC"]
        subgraph aza["🏫 building A - AZ a"]
            n1["🖥️ desk<br/>🪑 api-1 · 🪑 analytics-1"]
            n3["🖥️ new desk<br/>added by autoscaler"]
        end
        subgraph azb["🏫 building B - AZ b"]
            n2["🖥️ desk<br/>🪑 api-2 · 🪑 analytics-2"]
        end
    end
    alb["🌍 ALB - stands OUTSIDE both buildings,<br/>sends visitors to whichever is healthy"]
    ladder["🪜 the ladder:<br/>1 HPA: more pods - seconds<br/>2 autoscaler: more desks - minutes<br/>3 spread: copies across buildings<br/>4 PDB: never all away at once"]
    alb --> aza
    alb --> azb
    ladder -.-> campus

❓ What

AZ = physically separate data center in a region (AWS course, lesson 07). The EKS control plane is multi-AZ automatically — AWS runs the office across buildings; you only design the desks' side.
Multi-AZ nodes: give the node group subnets in ≥2 AZs (vpc.tf + eks.tf do) — the ASG then spreads and replaces desks across buildings.
Cluster autoscaler (or its newer cousin Karpenter): watches for Pending pods → grows the ASG; shrinks it when desks idle. HPA moves pods, autoscaler moves desks — a duet, not rivals.
topologySpreadConstraints — add to a Deployment's pod spec:

topologySpreadConstraints:
  - maxSkew: 1                                  # buildings may differ by ≤1 copy
    topologyKey: topology.kubernetes.io/zone    # spread BY BUILDING
    whenUnsatisfiable: ScheduleAnyway           # prefer, don't deadlock
    labelSelector: { matchLabels: { app: school-api } }

PodDisruptionBudget — minAvailable: 1 for a 2-replica app: node drains and upgrades proceed one desk at a time, never emptying the class.
Don't forget the data: pods spread easily; EBS drawers are per-building (a volume lives in ONE AZ!) — one more reason this repo keeps state in RDS, which has its own multi_az switch.

🤔 Why

An AZ outage is a when, not an if — and it's survivable only if replicas, nodes AND traffic already span buildings before the fire. Scaling without multi-AZ is a fast single point of failure; multi-AZ without autoscaling is resilience that falls over at the first exam day. The ladder is the whole point of everything you've learned: lesson 03's monitor, lesson 09's buses, the AWS course's ASG desks — climbing together.

🧪 Try it

# which building is each desk in? (works on EKS; minikube has 1 "building")
kubectl get nodes -L topology.kubernetes.io/zone

# where do the copies sit right now?
kubectl -n school get pods -o wide

# rung 2, felt: ask for more pods than the desks can seat →
kubectl -n school scale deployment school-api --replicas=20
kubectl -n school get pods | grep Pending | head -5
# Pending = kids standing without desks — EXACTLY what the cluster
# autoscaler watches for. (On EKS with it installed: new nodes in ~2 min.)
kubectl -n school scale deployment school-api --replicas=2   # calm down

# rung 4: protect the class during desk-shuffles:
kubectl -n school create pdb school-api --selector=app=school-api --min-available=1
kubectl -n school get pdb    # ALLOWED DISRUPTIONS: 1 — never both at once 🎉

🔧 kubectl for this lesson

kubectl get nodes -L topology.kubernetes.io/zone · kubectl get pdb

🔗 Builds on: 🪪 AWS L07 (AZs are buildings) · L12 (ASG)

✅ Verify — what you should see

kubectl get pods -o wide shows copies on different nodes/zones; ask for more pods than fit → Pending until the autoscaler adds a desk (EKS) — on minikube they stay Pending.

🧹 Clean up

scale back down; on EKS destroy the extra nodes/cluster same day

⚠️ Common mistakes

one AZ 'for now'
PDB with minAvailable = replicas — a drain can never finish

🎓 The campus is complete

Self-healing pods (03), rainy-day buses (09), zero-downtime substitutions (11), robots that deploy (14) — and now a school that survives fires and results day. That's not a demo cluster anymore; that's an architecture. 🏫🏫

And now that it stands: Part 3 — running it for real (lessons 16–26): debugging, hall passes, homework bells, seating rules, and eyes on everything.

git checkout lesson-16-debugging   # 🩺 the nurse's triage chart

← Previousdeploy gitops Next →debugging

This page is the lesson's README from the lesson-15-multi-az-scaling branch, shown here so the whole School stays on one site. Code files open on GitHub at the same branch.