Cheat Sheet: Compute

1. EC2 Instance Types and Families

1.1 Instance Family Categories

1.2 Burstable Performance Instances

• T3/T3a/T4g: Accumulate CPU credits when below baseline, consume when bursting
• Unlimited mode: Can burst beyond credits for additional cost
• Use for workloads with variable CPU usage patterns

1.3 Placement Groups

2. EC2 Purchasing Options

2.1 Pricing Models Comparison

2.2 Spot Instance Strategies

• Spot Fleet: Collection of Spot and On-Demand instances; can define multiple launch pools with different instance types and AZs
• Allocation strategies: lowestPrice, diversified, capacityOptimized, priceCapacityOptimized
• Spot Block: Request Spot instances for 1-6 hours without interruption (deprecated, use EC2 Fleet instead)
• EC2 Fleet: Provision mix of instance types and purchase options in single request; supports multi-AZ

3. Auto Scaling

3.1 Auto Scaling Components

3.2 Scaling Policy Types

3.3 Auto Scaling Configuration

• Default termination policy: Oldest launch template → oldest launch configuration → closest to next billing hour → random
• Cooldown period: Default 300 seconds; prevents ASG from launching/terminating before previous scaling takes effect
• Health checks: EC2 (instance status), ELB (target health); grace period before first health check (default 300 seconds)
• Lifecycle hooks: Perform actions before instance launch (Pending:Wait) or termination (Terminating:Wait); max timeout 7200 seconds
• Warm pools: Pre-initialized instances in Stopped or Hibernated state; faster scaling response

3.4 Multi-AZ and Load Balancer Integration

• ASG automatically distributes instances across multiple AZs for high availability
• Rebalancing: ASG launches new instances before terminating old ones in over-provisioned AZs
• Integration with ALB/NLB/CLB: ASG registers instances with load balancer target groups automatically
• Connection draining: ALB/NLB deregistration delay (default 300 seconds) before instance termination

4. Elastic Load Balancing

4.1 Load Balancer Types Comparison

4.2 ALB Advanced Features

• Content-based routing: Route to different target groups based on URL path, hostname, headers, query strings
• Host-based routing: Multiple domains on single ALB using host conditions
• Weighted target groups: A/B testing and blue/green deployments with traffic distribution control
• Fixed response: Return custom HTTP response directly from ALB without forwarding to target
• Redirect rules: HTTP to HTTPS, domain redirects at ALB level
• SNI: Multiple TLS certificates on single ALB; serves correct certificate based on hostname
• Request tracing: X-Amzn-Trace-Id header for end-to-end request tracking

4.3 Cross-Zone Load Balancing

4.4 Connection and Request Handling

• Connection draining (CLB) / Deregistration delay (ALB/NLB): 1-3600 seconds (default 300); completes in-flight requests before deregistering target
• Idle timeout: ALB default 60 seconds; client and target must send data within timeout to keep connection open
• Health checks: Configurable interval (5-300 seconds), timeout (2-120 seconds), healthy/unhealthy thresholds
• Target types: instance (route by instance ID), ip (route to IP including on-premises), lambda (invoke function)

5. Container Services

5.1 ECS Launch Types

5.2 ECS Core Components

5.3 ECS Task Placement

5.3.1 Placement Strategies (EC2 Launch Type)

• Binpack: Minimize number of instances by placing tasks based on least available CPU or memory
• Random: Place tasks randomly across instances
• Spread: Distribute tasks evenly based on specified field (instance ID, AZ, custom attribute)
• Can combine multiple strategies in order of execution

5.3.2 Placement Constraints (EC2 Launch Type)

• distinctInstance: Each task on different instance
• memberOf: Tasks placed on instances satisfying expression (cluster query language)

5.4 ECS Networking

5.5 ECS Service Auto Scaling

• Target Tracking: Scale based on CloudWatch metric (CPU, memory, ALB request count)
• Step Scaling: Scale based on CloudWatch alarms with step adjustments
• Scheduled Scaling: Time-based scaling for predictable patterns
• ECS Cluster Capacity Provider: Automatically scale EC2 instances for EC2 launch type using ASG
• Fargate: Scales tasks automatically; no infrastructure scaling needed

5.6 EKS Architecture

• Managed Kubernetes control plane across multiple AZs; automatic version upgrades and patching
• Node types: Managed node groups (EC2 automated), self-managed nodes (full control), Fargate (serverless)
• IAM integration: IAM roles for service accounts (IRSA) using OIDC provider; pod-level IAM permissions
• Networking: VPC CNI plugin assigns VPC IP addresses to pods; supports security groups for pods
• Storage: EBS CSI driver for persistent volumes; EFS CSI driver for shared storage

5.7 ECS vs EKS Decision Factors

6. Serverless Compute

6.1 Lambda Configuration

6.2 Lambda Invocation Models

6.3 Lambda Integrations

• VPC: Lambda functions can access VPC resources; requires subnet and security group configuration; uses Hyperplane ENI (scalable)
• RDS Proxy: Connection pooling for RDS; reduces connection overhead; improves scalability with Lambda
• EFS: Mount file system for persistent storage across invocations; requires VPC configuration
• Layers: Share code, libraries, dependencies across functions; max 5 layers per function; 250 MB total unzipped size
• Extensions: Integrate monitoring, observability, security tools; run in Lambda execution environment

6.4 Lambda Performance Optimization

• Provisioned Concurrency: Pre-initialized execution environments; eliminates cold starts; predictable performance
• Cold start factors: Runtime choice (compiled languages faster), VPC configuration (Hyperplane ENI reduces impact), package size
• Memory allocation: More memory = more CPU; test optimal memory for cost-performance balance
• Lambda SnapStart (Java): Reduces cold start latency by caching initialized snapshot
• Function reuse: Leverage execution context reuse; initialize SDK clients outside handler; cache in /tmp

6.5 Lambda Scaling and Throttling

• Scaling: Automatic up to account concurrency limit; up to 1,000 concurrent executions per region by default
• Burst limits: 3,000 (US West, US East, Europe), 1,000 (other regions), 500 (specific regions) concurrent executions per minute increase
• Reserved concurrency: Guarantees concurrency for function; limits maximum concurrency; prevents function from using unreserved pool
• Throttling: Returns 429 TooManyRequestsException; synchronous invocations return error to caller; asynchronous retry automatically

6.6 Step Functions

• Orchestrate Lambda functions and AWS services into workflows; visual workflow designer
• State types: Task, Choice, Parallel, Wait, Succeed, Fail, Pass, Map
• Standard workflows: Up to 1 year execution; exactly-once execution; full execution history; higher cost
• Express workflows: Up to 5 minutes; at-least-once execution; CloudWatch Logs for history; high-volume event processing; lower cost
• Error handling: Retry with exponential backoff; Catch for error transitions; built-in fault tolerance
• Integration: Direct SDK integrations with 200+ AWS services; optimized integrations for 9 services including Lambda, ECS, SNS, SQS, Glue, SageMaker, EMR, Batch, DynamoDB

7. Batch Processing

7.1 AWS Batch Components

7.2 AWS Batch Features

• Managed compute: Automatically provisions optimal compute resources based on job requirements
• Spot integration: Use Spot instances for cost savings; automatically replaces interrupted instances
• Multi-node parallel jobs: Tightly coupled workloads across multiple EC2 instances
• Array jobs: Submit large number of similar jobs with single submission; child jobs inherit parent parameters
• Job dependencies: Sequential, N_TO_N dependencies between jobs
• Scheduling: Fair share scheduling with policies; priority-based queue selection

7.3 Batch vs Lambda Decision

8. Additional Compute Services

8.1 Lightsail

• Simplified compute service with fixed monthly pricing; includes VM, SSD storage, data transfer, DNS management, static IP
• Pre-configured application stacks: WordPress, LAMP, Node.js, MEAN, Magento
• Use cases: Simple web applications, dev/test environments, small business workloads
• Limited scalability compared to EC2; cannot use advanced AWS features

8.2 Elastic Beanstalk

• PaaS for deploying applications; handles provisioning, load balancing, auto scaling, monitoring
• Platforms: Java, .NET, PHP, Node.js, Python, Ruby, Go, Docker
• Deployment policies: All at once, Rolling, Rolling with additional batch, Immutable, Blue/Green, Traffic splitting
• Full control: Access underlying resources (EC2, RDS, ALB); customize via .ebextensions configuration files
• Free service; pay only for underlying resources

8.3 Outposts

• Fully managed on-premises AWS infrastructure; extends VPC to on-premises
• Compute options: EC2 instances, ECS, EKS; consistent APIs and tools with AWS Regions
• Form factors: 42U racks, 1U and 2U servers
• Use cases: Low latency, local data processing, data residency requirements, hybrid cloud
• Local services: EBS, S3, RDS, EMR; connects to Region via service link

8.4 Wavelength

• Embeds AWS compute and storage within 5G networks at telecommunication carrier data centers
• Ultra-low latency: Single-digit millisecond latency to mobile devices and end users
• Services: EC2 instances, EBS, VPC; extends VPC into Wavelength Zone
• Use cases: ML inference at edge, AR/VR, real-time gaming, live video streaming

8.5 Local Zones

• AWS infrastructure closer to large population centers; extension of Region
• Lower latency than Region: Single-digit millisecond latency for end users
• Services: EC2, EBS, VPC, ECS, EKS, ALB, FSx
• Enable via opt-in; subnet in Local Zone connected to VPC in parent Region
• Use cases: Latency-sensitive applications, media rendering, real-time gaming

8.6 ParallelCluster

• Open-source cluster management tool for HPC on AWS
• Automates creation of VPC, subnet, compute fleet, shared file system
• Integrations: Slurm, AWS Batch as schedulers; FSx for Lustre, EFS for storage
• Supports multiple instance types, Spot instances, elastic scaling
• Configuration via YAML file; deploy via CLI or CloudFormation

8.7 App Runner

• Fully managed container and web application service; deploys from source code or container image
• Automatic scaling, load balancing, encryption; pay per use based on compute and memory
• Source: GitHub, Bitbucket, ECR; automatic deployments on code changes
• Use cases: Web applications, APIs, microservices without infrastructure management
• Simpler than ECS/EKS; less control but faster deployment

9. Compute Architecture Patterns

9.1 High Availability Patterns

• Multi-AZ: Deploy instances across at least 2 AZs; use Auto Scaling Groups with cross-zone load balancing
• Multi-Region: Active-active or active-passive; Route 53 health checks and routing policies; cross-region replication for data
• Stateless applications: Store session data in ElastiCache, DynamoDB; enables horizontal scaling
• Health checks: ELB target health, Route 53 endpoint health, ASG instance health; define appropriate grace periods

9.2 Scaling Patterns

• Vertical scaling: Increase instance size; requires downtime; limited by instance family maximum
• Horizontal scaling: Add more instances; no downtime with load balancer; preferred for cloud-native applications
• Scheduled scaling: Pre-scale for known traffic patterns; combine with dynamic scaling
• Predictive scaling: ML-based forecasting for proactive scaling; analyzes historical patterns

9.3 Cost Optimization Strategies

• Right-sizing: Use CloudWatch metrics to identify underutilized instances; Compute Optimizer recommendations
• Purchase options: Reserved Instances or Savings Plans for steady-state; Spot for fault-tolerant workloads; On-Demand for unpredictable
• Auto Scaling: Scale in during low demand; use target tracking with appropriate metrics
• Serverless adoption: Lambda for intermittent workloads; Fargate for containers without EC2 management overhead
• Instance generations: Use latest generation instances for better price-performance; Graviton2/Graviton3 for 20-40% better price-performance

9.4 Hybrid and Edge Patterns

• Outposts: Consistent AWS experience on-premises; VMware Cloud on AWS for VMware workloads
• Storage Gateway: Bridge on-premises applications to cloud storage
• Direct Connect: Dedicated network connection for hybrid workloads; more predictable latency than VPN
• Wavelength: Deploy at 5G edge for mobile applications
• Local Zones: Extend Region for latency-sensitive applications in specific metros

9.5 Migration Strategies

• Rehost (lift and shift): VM Import/Export, Application Migration Service; minimal changes
• Replatform: Move to managed services; RDS instead of self-managed database; Elastic Beanstalk for applications
• Refactor: Redesign for cloud-native; containerization, serverless, microservices
• Repurchase: Move to SaaS; replace custom applications with commercial solutions

10. Monitoring and Troubleshooting

10.1 CloudWatch Metrics

10.2 Logging and Tracing

• CloudWatch Logs: Centralized logging; EC2 requires CloudWatch agent; Lambda automatic; retention policies
• VPC Flow Logs: Network traffic monitoring; troubleshoot connectivity; security analysis
• ELB Access Logs: Request details stored in S3; includes timestamps, client IPs, latencies, request paths
• X-Ray: Distributed tracing; analyze and debug microservices; service map visualization; trace sampling
• CloudTrail: API call logging; governance, compliance, auditing; integrates with CloudWatch Logs

10.3 Common Troubleshooting Scenarios

10.4 Performance Optimization

• EC2: Use enhanced networking (ENA) for higher bandwidth and lower latency; placement groups for HPC
• EBS: Use provisioned IOPS for consistent performance; io2 Block Express for highest performance
• Load Balancer: Enable connection draining; configure appropriate idle timeout; use cross-zone load balancing
• Lambda: Optimize memory allocation; reduce package size; reuse execution context; use Provisioned Concurrency for consistent latency
• Containers: Right-size task CPU and memory; use Fargate Spot for cost-performance; implement health checks