Data Scientist → AI Engineer • Training 04

Article-Training • AI Engineering Foundations

Cloud & Deployment

From Local Container to Reliable Production Service

Learn the deployment mindset: choose the right compute, externalize configuration, release safely, observe production behavior and scale AI workloads without losing control.

Image → Registry → Compute → Network → Observe → Scale → Improve

📦 IMAGE

→

☁️ CLOUD

→

📡 SERVICE

→

📈 SCALE

8learning modules

24interactive practices

5rapid review questions

50%certificate threshold

Participant name

Learning target

Understand core cloud deployment choices, environments, managed services, networking, CI/CD, observability, scaling, reliability and cost tradeoffs for AI APIs and applications.

Practice progress0 / 24

Complete 12 of 24 practices (50%) and enter your name to unlock the Certificate of Participation.

MODULE 01

🧭

Cloud Mental Model

Cloud deployment means renting managed compute, storage, networking and platform capabilities instead of owning every physical layer. The engineering job is still to define boundaries and reliability expectations.

👁️

See it this way

Start with the simplest managed option that meets requirements. Complexity has an operational cost.

Core ideas

Compute runs code
Storage keeps durable data
Networking connects services
Managed services reduce operational burden

Try this

Local: container on laptop
Cloud: container on managed compute

Same app, different operational environment

✅

Start with the simplest managed option that meets requirements. Complexity has an operational cost.

Practice — 3 cases

Practice 1 / Práctica 1

Which statement best reflects this module?

ACompute runs codeBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Compute runs code

Practice 2 / Práctica 2

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BStart with the simplest managed option that meets requirements. Complexity has an operational cost.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Start with the simplest managed option that meets requirements. Complexity has an operational cost.

Practice 3 / Práctica 3

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CManaged services reduce operational burden

Best answer: C. Managed services reduce operational burden

MODULE 02

🧱

Choose a Deployment Target

AI services can run on virtual machines, managed container platforms, serverless functions or orchestrated clusters. The right target depends on traffic, runtime, latency, GPU needs and team maturity.

👁️

See it this way

Do not choose Kubernetes because it sounds advanced. Choose it when the operational problem justifies it.

Core ideas

VM: maximum control
Managed containers: balanced default
Serverless: event/short workloads
Cluster/Kubernetes: complex scale/control

Try this

# Decision questions
# - long-running API?
# - bursty traffic?
# - GPU?
# - custom networking?
# - team ops capacity?

✅

Do not choose Kubernetes because it sounds advanced. Choose it when the operational problem justifies it.

Practice — 3 cases

Practice 4 / Práctica 4

Which statement best reflects this module?

AVM: maximum controlBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. VM: maximum control

Practice 5 / Práctica 5

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BDo not choose Kubernetes because it sounds advanced. Choose it when the operational problem justifies it.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Do not choose Kubernetes because it sounds advanced. Choose it when the operational problem justifies it.

Practice 6 / Práctica 6

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CCluster/Kubernetes: complex scale/control

Best answer: C. Cluster/Kubernetes: complex scale/control

MODULE 03

🛠️

Environments & Configuration

Development, staging and production should use the same artifact with environment-specific configuration. Separate credentials, endpoints and data while preserving behavior as much as possible.

👁️

See it this way

Environment parity reduces “surprises” between staging and production.

Core ideas

Build once
Configure per environment
Separate secrets
Promote tested artifact

Try this

APP_ENV=production
MODEL_VERSION=2026-09
LOG_LEVEL=INFO
DATABASE_URL=...

# same image, different config

✅

Environment parity reduces “surprises” between staging and production.

Practice — 3 cases

Practice 7 / Práctica 7

Which statement best reflects this module?

ABuild onceBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Build once

Practice 8 / Práctica 8

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BEnvironment parity reduces “surprises” between staging and production.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Environment parity reduces “surprises” between staging and production.

Practice 9 / Práctica 9

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CPromote tested artifact

Best answer: C. Promote tested artifact

MODULE 04

🧪

Networking, Domains & TLS

A production AI API usually sits behind DNS, TLS and a load balancer or managed ingress. Network policy should expose only what clients actually need.

👁️

See it this way

Public by default is not a deployment strategy. Minimize exposed surface.

Core ideas

Use HTTPS
Keep private services private
Route through managed ingress/load balancer
Control inbound/outbound traffic

Try this

client -> https://api.example.com
       -> load balancer
       -> AI service
       -> private database/model store

✅

Public by default is not a deployment strategy. Minimize exposed surface.

Practice — 3 cases

Practice 10 / Práctica 10

Which statement best reflects this module?

AUse HTTPSBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Use HTTPS

Practice 11 / Práctica 11

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BPublic by default is not a deployment strategy. Minimize exposed surface.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Public by default is not a deployment strategy. Minimize exposed surface.

Practice 12 / Práctica 12

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CControl inbound/outbound traffic

Best answer: C. Control inbound/outbound traffic

MODULE 05

⚙️

CI/CD: Release Without Fear

Continuous integration tests every change; continuous delivery/deployment packages and releases validated changes through an automated pipeline.

👁️

See it this way

Automation reduces manual variance, but only if rollback and verification are designed into the pipeline.

Core ideas

Test code
Build image
Scan/package
Deploy and verify

Try this

git push
  -> tests
  -> docker build
  -> push registry
  -> deploy staging
  -> smoke test
  -> promote production

✅

Automation reduces manual variance, but only if rollback and verification are designed into the pipeline.

Practice — 3 cases

Practice 13 / Práctica 13

Which statement best reflects this module?

ATest codeBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Test code

Practice 14 / Práctica 14

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BAutomation reduces manual variance, but only if rollback and verification are designed into the pipeline.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Automation reduces manual variance, but only if rollback and verification are designed into the pipeline.

Practice 15 / Práctica 15

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CDeploy and verify

Best answer: C. Deploy and verify

MODULE 06

📡

Observability for AI Services

Logs, metrics and traces tell you whether the system is healthy. AI adds extra signals: model latency, token usage, retrieval quality, drift and business outcomes.

👁️

See it this way

If you cannot observe it, you cannot operate it responsibly.

Core ideas

Logs explain events
Metrics quantify health
Traces connect distributed calls
AI metrics measure model/application quality

Try this

metrics:
  request_rate
  p95_latency
  error_rate
  cpu_memory
  model_confidence
  token_cost

✅

If you cannot observe it, you cannot operate it responsibly.

Practice — 3 cases

Practice 16 / Práctica 16

Which statement best reflects this module?

ALogs explain eventsBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Logs explain events

Practice 17 / Práctica 17

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BIf you cannot observe it, you cannot operate it responsibly.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. If you cannot observe it, you cannot operate it responsibly.

Practice 18 / Práctica 18

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CAI metrics measure model/application quality

Best answer: C. AI metrics measure model/application quality

MODULE 07

🛡️

Scaling, Reliability & Resilience

Production services need enough capacity and graceful failure behavior. Scale horizontally when possible, use timeouts/retries carefully and design health checks and rollback paths.

👁️

See it this way

Retries can amplify outages. Reliability patterns must be applied with context.

Core ideas

Autoscale based on useful signals
Use timeouts
Retry only safe operations
Design rollback and redundancy

Try this

replicas: 3
autoscale: 3 -> 20
timeout: 5s
health: /health
rollback: previous image tag

✅

Retries can amplify outages. Reliability patterns must be applied with context.

Practice — 3 cases

Practice 19 / Práctica 19

Which statement best reflects this module?

AAutoscale based on useful signalsBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Autoscale based on useful signals

Practice 20 / Práctica 20

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BRetries can amplify outages. Reliability patterns must be applied with context.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. Retries can amplify outages. Reliability patterns must be applied with context.

Practice 21 / Práctica 21

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CDesign rollback and redundancy

Best answer: C. Design rollback and redundancy

MODULE 08

🚀

Cost, Governance & Production Readiness

Cloud makes capacity easy to consume, not free. AI workloads can create large compute, GPU and token bills, so cost visibility and governance belong in architecture decisions.

👁️

See it this way

A production system is not finished when it deploys. It is finished when the team can operate, secure, measure and afford it.

Core ideas

Tag resources
Set budgets/alerts
Right-size compute
Track cost per useful outcome

Try this

unit_economics = monthly_ai_cost / successful_business_outcomes
print(unit_economics)

✅

A production system is not finished when it deploys. It is finished when the team can operate, secure, measure and afford it.

Practice — 3 cases

Practice 22 / Práctica 22

Which statement best reflects this module?

ATag resourcesBIgnore interfaces and let every caller depend on internal implementation details.CRemove validation and monitoring to reduce engineering effort.

Best answer: A. Tag resources

Practice 23 / Práctica 23

Which action is the best engineering choice?

AHard-code production secrets and configuration directly in source code.BA production system is not finished when it deploys. It is finished when the team can operate, secure, measure and afford it.CTreat every failure as a user error and return no diagnostic information.

Best answer: B. A production system is not finished when it deploys. It is finished when the team can operate, secure, measure and afford it.

Practice 24 / Práctica 24

Which option would you avoid in a production system?

AAssume a system is reliable because it worked once on a developer laptop.BChange production behavior manually without versioning or traceability.CTrack cost per useful outcome

Best answer: C. Track cost per useful outcome

5-Question Knowledge Check

Can you explain the core ideas clearly?

Open each item only after answering it in your own words.

1. What is the simplest cloud principle to remember?

Use managed building blocks to run your workload, but keep explicit control of configuration, security, reliability and cost.

2. Why promote the same image across environments?

It preserves artifact consistency and reduces environment-specific rebuild differences.

3. What does CI/CD reduce?

Manual variance in testing, packaging and releasing software.

4. What should AI observability include beyond CPU and errors?

Application/model quality signals such as latency, token/cost usage, retrieval quality, drift and business outcomes.

5. When is a deployment truly production-ready?

When it can be operated, secured, observed, scaled, rolled back and financed sustainably.

Decision Guide

What should you reach for?

Need	Recommended approach
Run simple container API	Managed container platform
Automate release	CI/CD pipeline
See production health	Logs + metrics + traces
Control cloud spend	Budgets + unit economics

Certificate of Participation

Complete at least 12 of the 24 practice cases (50%) and enter your name.

0 / 24 • 0%