Cloud Service Cheatsheet
GCP vs AWS vs Azure (Interview & System Design Oriented)
This cheatsheet maps core cloud primitives across GCP, AWS, and Azure, with an emphasis on ML systems, data platforms, and production workloads.
The goal is not memorizing names, but understanding equivalent building blocks and trade-offs.
Compute
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Virtual Machines | Compute Engine | EC2 | Virtual Machines |
| Autoscaling | Managed Instance Groups | Auto Scaling Groups | VM Scale Sets |
| Containers (Managed) | GKE | EKS / ECS | AKS |
| Serverless (HTTP) | Cloud Run | Lambda | Azure Functions |
| Batch Jobs | Batch / Dataflow | Batch | Azure Batch |
Interview tip:
Serverless = fast iteration, bursty workloads
VMs / K8s = predictable latency, long-running services
Storage
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Object Storage | Cloud Storage | S3 | Blob Storage |
| Block Storage | Persistent Disk | EBS | Managed Disks |
| File Storage | Filestore | EFS | Azure Files |
ML usage notes - Object storage for datasets, checkpoints, artifacts - Block storage for low-latency model serving - File storage mostly for legacy or shared workloads
Databases & Data Platforms
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Managed SQL | Cloud SQL | RDS | Azure SQL |
| Globally Scalable SQL | Spanner | Aurora | Cosmos DB |
| NoSQL (Wide Column / KV) | Bigtable | DynamoDB | Cosmos DB |
| Cache | Memorystore | ElastiCache | Azure Cache for Redis |
| Data Warehouse | BigQuery | Redshift | Synapse |
Design signal:
- OLTP ≠ Analytics
- BigQuery / Redshift / Synapse are not serving databases
Messaging & Streaming
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Pub/Sub (Eventing) | Pub/Sub | SNS / SQS | Service Bus |
| Streaming | Pub/Sub | Kinesis | Event Hubs |
| Workflow Orchestration | Cloud Composer | Step Functions | Logic Apps |
ML relevance - Event-driven feature updates - Training triggers - Async inference workflows
Networking
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Virtual Network | VPC | VPC | VNet |
| Load Balancer | Cloud Load Balancing | ALB / NLB | Azure Load Balancer |
| DNS | Cloud DNS | Route 53 | Azure DNS |
| CDN | Cloud CDN | CloudFront | Azure CDN |
Interview note:
Talk in terms of L4 vs L7, not product names.
Observability & Reliability
| Concept | GCP | AWS | Azure |
|---|---|---|---|
| Logging | Cloud Logging | CloudWatch Logs | Azure Monitor |
| Metrics | Cloud Monitoring | CloudWatch Metrics | Azure Monitor |
| Tracing | Cloud Trace | X-Ray | Application Insights |
| Alerts | Alerting Policies | CloudWatch Alarms | Alerts |
Senior signal:
Observability is a first-class system requirement, not an afterthought.
ML & MLOps Platforms (Critical for MLE Interviews)
| Capability | GCP | AWS | Azure |
|---|---|---|---|
| ML Platform | Vertex AI | SageMaker | Azure ML |
| Training Jobs | Vertex Training | SageMaker Training | Azure ML Jobs |
| Feature Store | Vertex Feature Store | SageMaker Feature Store | Azure Feature Store |
| Pipelines | Vertex Pipelines | SageMaker Pipelines | Azure ML Pipelines |
| Model Registry | Vertex Model Registry | SageMaker Registry | Azure ML Registry |
| Online Serving | Vertex Endpoints | SageMaker Endpoints | Azure Online Endpoints |
| Monitoring | Vertex Model Monitoring | SageMaker Model Monitor | Azure ML Monitoring |
Key interview insight:
Most ML failures come from data + serving, not modeling.
How I Reason Across Clouds (Interview Framing)
Instead of saying:
“I’d use BigQuery”
Say:
“I’d use a managed columnar data warehouse (BigQuery / Redshift) optimized for analytical queries.”
Instead of:
“We use SageMaker”
Say:
“We use a managed ML platform for training, registry, deployment, and monitoring.”
Cloud Selection Heuristics
- GCP: Strongest for analytics & ML-native workflows
- AWS: Broadest ecosystem, best for infra-heavy systems
- Azure: Strong enterprise integration, identity-first
In interviews, emphasize trade-offs, not preferences.
TL;DR Mental Model
All clouds provide the same primitives:
Compute · Storage · Network · Data · ML · Observability
If you can reason at that level, service names are just details.
This cheatsheet is intended for ML system design interviews and real-world architecture discussions.