In one sentence
Match compute, storage, database, networking, and caching choices to the workload’s access patterns and bottlenecks.
Key points
- Scale horizontally when possible and use managed or serverless services where their scaling model fits.
- Choose storage by object, block, or file access and databases by relational, key-value, document, graph, or cache patterns.
- Use CloudFront, read replicas, DAX, or ElastiCache only when their cache layer matches the data and consistency needs.
Exam takeaway
Measure latency, throughput, concurrency, and cost; do not optimize from service names alone.
Analytics & AI/ML performance
| Performance or architecture need | Start with |
|---|---|
| Ad hoc SQL directly over S3 | Athena + partitions / columnar formats / compression |
| Sustained, frequent, complex analytics and BI | Amazon Redshift |
| Real-time streams, multiple consumers, custom processing | Kinesis Data Streams |
| Low-operations near-real-time delivery to S3 and other targets | Amazon Data Firehose |
| Serverless ETL and data cataloging | AWS Glue |
| Large-scale custom Spark or Hadoop processing | Amazon EMR |
| Train and deploy models with proprietary data | Amazon SageMaker AI |
| Build generative-AI apps through managed foundation-model APIs | Amazon Bedrock |
Design reminders
- Isolate production DynamoDB traffic from large historical analytics and training workloads.
- Athena scan volume depends on partitioning, compression, and file format.
- Firehose buffers records, so verify whether its delivery latency meets the requirement.
- Training and inference are different workloads; plan compute, scaling, and monitoring separately.
- When model quality drops, check data quality, schema, and training-data freshness first.
Region, edge, and scope
- Choose a Region by compliance, latency, service availability, and price.
- Edge Locations / PoPs bring CloudFront, Route 53, and Global Accelerator closer to users but are not ordinary EC2 deployment locations.
- CloudFront caches HTTP/HTTPS content, Route 53 returns DNS answers, and Global Accelerator optimizes TCP/UDP network paths.
- Regional services and zonal resources have different placement scopes.
- Service availability in a Region is not the same as high availability.