Cloud Computing Advanced and Emerging Cloud Topics Syllabus

Give examples of pre-trained vision, speech, and language APIs across the clouds.

Vision: AWS Rekognition, Azure Computer Vision, Google Vision AI. Speech: AWS Transcribe/Polly, Azure Speech, Google Speech-to-Text. Language: AWS Comprehend/Translate, Azure Language, Google Natural Language/Translation.

What is a foundation model?

A foundation model is a large model trained on broad, unlabeled data at scale that can be adapted (via fine-tuning or prompting) to a wide range of downstream tasks. LLMs like GPT and Claude are examples.

What is the difference between fine-tuning and prompt engineering (in-context learning)?

Fine-tuning updates a model's weights by training on task-specific data; prompt engineering / in-context learning guides a frozen model's behavior purely through the input prompt and examples, without changing weights.

What is RAG (Retrieval-Augmented Generation) and why is it used with foundation models?

RAG augments a generative model by retrieving relevant documents (often via vector similarity search) and adding them to the prompt as context. It grounds responses in up-to-date, domain-specific data and reduces hallucination without retraining.

Name the managed generative-AI / foundation-model services on AWS, Azure, and Google Cloud.

AWS Bedrock, Azure OpenAI Service (and Azure AI Foundry), and Google Vertex AI (Gemini/Model Garden).

What is MLOps?

MLOps is the set of practices that combines Machine Learning, DevOps, and data engineering to reliably and efficiently deploy, monitor, and maintain ML models in production, covering CI/CD, versioning, automation, and governance.

What is model drift, and distinguish data drift from concept drift.

Model drift is degradation of model performance over time. Data drift is a change in the input feature distribution; concept drift is a change in the relationship between inputs and the target (the underlying pattern changes).

What is a feature store in MLOps?

A feature store is a centralized repository for storing, sharing, and serving curated ML features consistently between training and inference (avoiding training-serving skew), with support for both offline (batch) and online (low-latency) access.

What does CI/CD/CT mean in the MLOps context?

Continuous Integration (test code and data), Continuous Delivery/Deployment (automatically deploy models/pipelines), and Continuous Training (CT)—automatically retraining models on new data, which is unique to ML systems.

What is edge computing?

Edge computing is a distributed paradigm that processes data near the source of generation (at or close to devices/sensors) rather than in a centralized cloud, reducing latency, bandwidth use, and dependence on constant connectivity.

State the primary benefits of edge computing over pure centralized cloud processing.

Lower latency, reduced backhaul bandwidth/cost, improved reliability during network outages, better data privacy/locality, and real-time responsiveness for local decisions.

What is IoT device management and what core functions does it include?

IoT device management is the process of provisioning, authenticating, configuring, monitoring, updating (OTA firmware), and decommissioning connected devices at scale. It ensures security, health, and lifecycle control of fleets.

Name the managed IoT platforms on AWS, Azure, and Google Cloud.

AWS IoT Core, Azure IoT Hub, and Google Cloud IoT (with partners). These provide device registry, secure messaging (often MQTT), and cloud integration.

Which lightweight publish/subscribe protocol is most commonly used for IoT device messaging, and why?

MQTT (Message Queuing Telemetry Transport). It is lightweight, has minimal overhead, supports pub/sub with QoS levels, and works well over constrained, low-bandwidth, high-latency networks.

What is a device shadow (digital twin) in IoT management?

A device shadow is a persistent virtual JSON representation of a device's state stored in the cloud, allowing applications to read/set desired state and sync with the physical device even when it is temporarily offline.

What is fog computing, and how does it relate to edge and cloud?

Fog computing is an intermediate layer between edge devices and the cloud, distributing compute, storage, and networking across nodes (e.g., gateways, routers) in the local network. It aggregates and processes edge data before selectively forwarding to the cloud.

Distinguish edge computing from fog computing.

Edge computing performs processing directly on or adjacent to the end devices/sensors; fog computing occurs at an intermediate infrastructure layer (gateways, local nodes) between the edge and the cloud, coordinating multiple edge devices.

What defines a latency-sensitive workload, and give two examples.

A latency-sensitive workload requires very low, predictable response times where delay degrades correctness or user experience. Examples: autonomous vehicle control, industrial robotics, real-time gaming, high-frequency trading, AR/VR.

Round-trip time is bounded below by physics. Approximately how long does light take to travel 3000 km through fiber one way, and why does this favor edge computing?

Light in fiber travels at roughly $2\times10^{8}\,\text{m/s}$, so $3000\,\text{km}$ takes about $t=\frac{3\times10^{6}}{2\times10^{8}}=15\,\text{ms}$ one way (~30 ms round trip). This irreducible propagation delay is why latency-sensitive workloads are placed at the edge.

What is a cloud abstraction layer, and what problem does it solve?

A cloud abstraction layer sits above provider-specific APIs to present a unified, provider-agnostic interface. It solves fragmentation across clouds, enabling portable code, consistent tooling, and reduced coupling to any single vendor's services.

Give examples of technologies that act as cloud abstraction / portability layers.

Kubernetes (container orchestration), Terraform/OpenTofu (infrastructure as code), Crossplane, and multi-cloud abstraction frameworks. These abstract compute/networking/storage across providers.

What are Google Anthos and Azure Arc used for?

They are hybrid/multi-cloud management platforms that extend a provider's control plane to on-premises and other clouds. Anthos (Google) and Azure Arc let you manage Kubernetes clusters, servers, and services consistently across environments from one control plane.

What is hybrid connectivity in cloud computing?

Hybrid connectivity provides secure, reliable network links between on-premises data centers and cloud environments (or between clouds), via VPN, dedicated private connections (AWS Direct Connect, Azure ExpressRoute, Google Cloud Interconnect), and consistent identity/policy management.

What is workload portability, and what technologies enable it?

Workload portability is the ability to move applications and their workloads across environments (clouds, on-prem) with minimal changes. It is enabled by containers (Docker), orchestration (Kubernetes), and open standards/IaC that abstract away provider specifics.

What is vendor lock-in, and name two strategies to avoid it.

Vendor lock-in is dependence on a specific provider's proprietary services making migration costly or difficult. Strategies to avoid it: use open standards/open-source tools (Kubernetes, Terraform, open formats), containerize workloads, adopt multi-cloud/abstraction layers, and avoid deep proprietary service coupling.

What is the trade-off between using managed proprietary cloud services and avoiding vendor lock-in?

Proprietary managed services offer higher productivity, less operational overhead, and advanced features, but increase lock-in. Portable open-source/self-managed approaches reduce lock-in but require more engineering effort and ops responsibility.

What is carbon-aware computing?

Carbon-aware computing schedules and shifts workloads in time or location to run when/where the electricity grid's carbon intensity is lowest (e.g., when renewable supply is high), reducing the carbon footprint of computation.

Define grid carbon intensity and give its typical unit.

Carbon intensity is the amount of CO2 emitted per unit of electricity consumed, typically expressed in grams of CO2-equivalent per kilowatt-hour, i.e. $\text{gCO}_2\text{e}/\text{kWh}$. It varies by region and time based on the energy mix.

What is PUE (Power Usage Effectiveness), how is it computed, and what is the ideal value?

PUE measures data center energy efficiency: $$\text{PUE}=\frac{\text{Total Facility Energy}}{\text{IT Equipment Energy}}$$ The ideal value is $1.0$ (all energy goes to computing); typical modern hyperscale data centers approach $1.1\text{–}1.2$.

Name three techniques used to build energy-efficient cloud architectures.

Autoscaling and rightsizing to match capacity to demand, serverless/spot usage to eliminate idle resources, workload consolidation and higher server utilization, efficient hardware (ARM/custom silicon, accelerators), and carbon-/renewable-aware scheduling.