Design High-Performing Architectures

Key Concepts

• High-performing architectures = right compute + right storage + right database + right networking for the workload
• Elasticity: Scale out (horizontal) preferred over scale up (vertical) for resilience and cost
• Decoupling: Use SQS, SNS, EventBridge to separate components so one slow part doesn't bottleneck another
• Caching layers: ElastiCache (Redis/Memcached), CloudFront, DAX (DynamoDB Accelerator) — reduce latency and backend load
• Auto Scaling: EC2 Auto Scaling Groups, Application Auto Scaling (ECS, DynamoDB, Lambda concurrency)
• Read Replicas: Offload read traffic from primary DB (RDS Read Replicas, Aurora Replicas)
• Aurora outperforms standard RDS — up to 5x MySQL, 3x PostgreSQL throughput; Aurora Serverless for unpredictable workloads
• S3 performance: 3,500 PUT/COPY/POST/DELETE and 5,500 GET/HEAD requests per second per prefix — add prefixes to scale
• EBS volume types:
• gp3: General purpose, baseline 3,000 IOPS (best default choice)
• io2/io2 Block Express: High-performance, up to 256,000 IOPS — use for critical databases
• st1: Throughput-optimized HDD — big sequential reads (data warehouses, logs)
• sc1: Cold HDD — infrequent access, cheapest
• Instance Store: Ephemeral, physically attached — highest possible disk throughput, but lost on stop/terminate
• Global Accelerator: Improves performance for global users by routing through AWS backbone (not just CDN caching)
• CloudFront: CDN for static/dynamic content — reduces latency via edge locations
• Placement Groups:
• Cluster: Low latency, high throughput (HPC, big data) — same AZ, same rack
• Spread: Max availability — different hardware per instance
• Partition: Large distributed workloads (Kafka, Hadoop) — isolated racks per partition
• Enhanced Networking / ENA: Up to 100 Gbps, lower latency — enable for HPC workloads
• Kinesis Data Streams: Real-time streaming; each shard = 1 MB/s in, 2 MB/s out
• DynamoDB: Single-digit millisecond performance at any scale; use DAX for microsecond reads

How It Works

Caching Flow (ElastiCache) Request → Check cache → Cache hit: return data instantly | Cache miss: query DB → store in cache → return data

Auto Scaling Decision CloudWatch metric breach → Scaling policy triggers → Launch/terminate instances → ELB registers/deregisters targets

S3 Prefix Scaling bucket/folder1/file and bucket/folder2/file = 2 separate prefixes = 2× the request rate limits — distribute keys across prefixes for high-throughput workloads

Commands / Syntax / Key Values