Notes from studying for the AWS Certified Solutions Architect Exam. I felt well-prepared for the exam and passed with a 94%. Please reach out with any corrections or questions.

AWS-CSA-A-Notes.md

External Resources

• linuxacademy.com (Training)
• Acloud.guru (training)
• Read the top-level Documentation and FAQs for all the major AWS resources (EC2, S3, RDS, Auto Scaling, etc). The answers to the "nit-picky" questions can be found here. It's also helpful to go deeper on VPCs and networking-related concepts.
  • VPC http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/using-vpc.html
  • VPC Security http://docs.aws.amazon.com/AmazonVPC/latest/UserGuide/VPC_Security.html
  • VPC Route Tables http://docs.aws.amazon.com/AmazonVPC/latest/UserGuide/VPC_Route_Tables.html
• Everyone says read the white papers. The ones I read were:
  • Security Best Practices https://d0.awsstatic.com/whitepapers/Security/AWS_Security_Best_Practices.pdf
  • Cloud Best Practices https://d0.awsstatic.com/whitepapers/AWS_Cloud_Best_Practices.pdf

Core

• Elasticity
  • Ability to scale up/down on demand
  • Reduce cost
  • Scaling
    • Proactive - fixed interval
    • Proactive Event-based
    • Auto-Scaling based on demand
    • Out: more instances, Up: Bigger instances
    • Fault tolerant
    • Operationally efficient
    • More resources, less resources cost
• Scalability
  • grow over time
  • economies of scale
  • vertically: more powerful resources
  • horizontally: increased number of resources
• Reserved
  • reduced price
  • guaranteed capacity (useful for disaster recovery)
  • within AZ
  • Can submit modification request to
    • change AZ
    • scope up/down AZ to Region
    • instance size w/in same type
• "Tightly Coupled" vs "Loosely Coupled"
  • Tightly
    • one thing fails, all fail
  • Loosely
    • individual component failure won't break everything
    • Can scale out individual components
• Shared Security Model
  • AWS: host OS/Virtualization, physical security
    • DDoS
      • Could use CloudFront to absorb requests
    • Do INGRESS filtering on incoming traffic
  • You: guest OS/VPC on up, security group, software updates, data in transit/rest
    • Also use software-level security: IPTABLES, Firewalls, etc
    • DDoS
      • Block CIDR at Network ACL (Subnet level)
      • Must have permission to do Port Scanning w/in cloud
  • Encryption (AES-256):
    • EBS
      • done on instance itself, not good for small instances
      • snapshots automatically encrypted
    • S3 - at rest
    • Glacier
    • Redshift
    • SQL RDS
      • MySQL//Aurora, Oracle, Postgres, MSSQL
      • Snapshots, backups, read replicas all encrypted
      • SSL connection encryption
• Disaster Recovery
  • Recovery Time Objective (RTO)
    • restore to level of service, measured in time
  • Recovery Point Objective (RPO)
    • acceptable amount of dataloss, measured in time
  • Methodology
    • Pilot light
      • Minimal version of production environment in AWS
      • Scale out and DNS switch if disaster
      • Make sure it's up to date
      • Requires extra time to spin up
    • Warm standby
      • Larger footprint than pilot light
      • Running business critical applications
    • Multi-site
      • clone production environment
      • active-active
      • also use as load balance
      • less downtime, more costly
  • DR Services
    • ELB and Auto Scaling
    • Route53 failover DNS, or latency based
    • Storage Gateway
    • lots of AWS tools to get data, AMIs out of on-premises

EC2

• AMI
  • Unique to a Region (need to manually copy to another Region)
  • Permissions: can make public, or available to another AWS account
  • PV Paravirtual
    • cannot take advantage of hardware extensions
    • historically was faster, but basically no diff now
  • HVM Hardware Virtual - preferred method, on current instance types
  • Copying an image, can encrypt from unencrypted, but default is keep same
• Security Group
  • EC2-Classic: Can't add/remove SGs to running instance
  • VPC: Live in that VPC
  • Stateful: response traffic is allowed
  • 500 Security Groups / VPC, 50 rules/SG, 5 SGs/network interface
  • Can't delete the default security group
  • Instance needs at least one security group
• Dedicated instance - hardware w/ just our stuff
• Elastic Load Balancer
  • Can distribute traffic across AZs
  • Can live in a public subnet and serve traffic to private subnets (eg Auto Scaling Groups)
  • Can apply an SSL directly to it
  • Needs to split between two subnets?
  • Configure
    • What to listen on (80)
    • Apply a Security Group to the ELB
    • Health Check on target instances (eg port 80)
    • target instances
  • Instances Security Groups still need to allow traffic from ELB
    • Instances don't need a Public IP, just a public subnet
    • Requests look like they're from ELB, might want to log traffic at ELB level
  • ELB for Apex domain required Route53 Alias record
  • Internal ONLY accepts traffic from w/in VPC (use w/ private subnet for multi-tier apps)
  • Connection draining: wait for connections to complete
  • Cross Zone load balancing required across AZs
  • AWS will increase ELB as needed
    • Contact AWS for "Prewarming" to handle rapid, drastic spikes in traffic
  • Lots of built in Metrics on the ELB (connections, responses, etc)
    • Can create Alerts off of it
  • Classic: register instances, Application: target group
  • Troubleshooting
    • Make sure target is available for health check (correct "index.html" target)
    • Make sure ELB and Instances have port 80 open
    • Enable Access Logs to Amazon S3 (otherwise ELB logs on Instance)
    • Add specific subnets to ELB
• Auto Scaling Group (ASG)
  • Launch additional instances as needed
    • Load
    • Proactive
    • Event
  • Consists of Launch Configuration + Scaling Plan
  • Spans AZs, not Regions
  • Tell it to receive traffic from an ELB
  • static size = "self healing", or
  • scaling policy to adjust capacity w/ CloudWatch Alarm for Increase and Decrease group size
  • Can use with internal multi-tier apps
  • Deleting ASG terminates running instances
  • Can't pass EC2 instance cap
  • Launch Configuration (LC)
    • Info on the instance that an Auto Scaling Group uses (AMI, instance type, etc)
    • LC can be used w/ multiple ASGs
    • ASG requires an LC
    • Can't modify LC after launch
      • Create new LC, apply to ASG
        • Only affects new instances! Existing are kept.
  • Scaling Plan
    • Tells Auto Scaling when and how to scale
      • Manual
      • Scheduled
      • Demand (In and Out)
        • Alarm - What to look for (via CloudWatch)
          • Could be SQS Queue Size
        • Policy - how to respond to Alarm
          • Can't go outside min/max group size
          • Adjustment Types
            • ChangeInCapacity (+/-)
            • ExactCapacity
            • PercentChangeInCapacity
          • Scaling Policy Types
            • Simple
              • has cooldown
            • Step
              • based on size of alarm breach
              • continuously evaluated
  • Troubleshooting
    • "thrashing" up and down
      • Change thresholds
      • Decrease checking frequency
      • Increase cooldown (not for step or scheduled)
    • doesn't happen
      • max too low
• EBS Volumes
  • cannot cross AZs (auto replicated WITHIN AZ)
    • create snapshot in other AZ to access there
  • IOP is max 256KB
  • SSD General Purpose
    • 1GiB to 16TiB
    • Burstable IOP credits
    • baseline 3 IOPS/GiB
  • Provisioned IOPS
    • 4GiB to 16TiB
    • up to 20,000 IOPS
    • Critical apps requiring sustained IOPS, large DB workloads
  • Magnetic
    • 1Gib to 1024GiB
  • AWS EBS encryption uses AWS Key Management Service (don't use on smaller than m3)
    • Customer Master Key (CMK)
    • Snapshots/Volumes by default inherit status of their source
      • Can't change CMK of volume/snapshot but
      • encrypt > encrypt w/ new CMK w/ copy snapshot
      • unencrypted > encrypt: w/ copy snapshot
      • encrypt > unencrypt: mount both, copy over
    • Can share an encrypted snapshot but using non-default CMK and sharing both w/ other account
  • Snapshots
    • Stored on S3 under the hood (can't see them)
    • incremental in nature (behind the scenes first snapshot might still exist)
    • When creating snapshots of EBS volumes that are configured in a RAID array, it is critical that there is no data I/O to or from the volumes when the snapshots are created
      • freeze filesystem, unmount, or stop instance
• Elastic File System (EFS)
  • NFS for EC2
• EC2 Classic
  • internal IPs are unstable after reboot
• Placement Groups
  • instances w/in same AZ, low latency (close together), min 10 Gbps network
  • stop / start as a group
  • failure is "insufficient capacity error"
  • CAN span peered VPCs, apparently
• Spot Instances
  • If AWS terminates, no charge for last hour (you kill, you pay)
  • The price per instance-hour for a Spot instance is set at the beginning of each instance-hour for the entire hour. Any changes to the Spot price will not be reflected until the next instance-hour begins.

S3

• URL
  • to object bucketname.s3.amazonaws.com/path
  • to static site linuxacademy-big-bucket.s3-website-ap-southeast-2.amazonaws.com
• unlimited storage
  • no limit to number of objects in bucket
• 11 nines durability, 99.99% availability
• Created in a Region, stay there
  • synchronized across all AZ w/in Region automatically
  • new PUT read-after-write consistent
  • overwrite PUT, DELETE eventual consistent
• bucketnames unique across ALL regions
• 100 buckets / account, can't change owner
• Account (not user) owns bucket
• min object size: 0 bytes
• max: 5GB? multipart: 5TB (recc'd > 100MB)
• if rapid increase > 100 PUT/LIST/DELETE or > 300 GET reqs/sec, contact AWS
• CORS configuration
  • JavaScript thing
• Resource Based Policies
  • ACL to share bucket across account
  • Bucket Policy
    • Restrict off IP address, HTTP referrer,
    • Will overwrite "Public Permissions"
    • Can be edited by CloudFront Distributions
    • Use IAM to give a User access to a specific bucket
      • Resource format arn:aws:s3:::bucketname/folder
• User-based Policies
  • IAM
• RRS - Reduced Redundancy Storage
  • 99.99 durability, availability
  • cheaper, for reproducible objects
• Versioning
  • Off by default
  • Once on, can only be suspended, not disabled (old versions continue to exist)
• Lifecycle policies
  • Archive (Glacier) and Delete options
  • Can be applied to Versions

Glacier

VPC

• Spans all AZs in the Region
• Internet Gateway (IGs) attach to VPC
• /28 is smallest possible subnet, /16 largest
  • can't resize (requires redo)
• Subnet
  • Only belongs to one AZ
  • /28 is smallest possible subnet, up to VPC range
    • First 4, last 1 reserved by AWS
  • Route Tables
    • VPC comes w/ main route table
    • Subnet must have route table (only one at a time, implicitly main).
    • Route table can have multiple Subnets
    • Subnet automatically associated with the main route table for the VPC (modifiable)
  • private subnet - no internet gateway
    • public requires IG + Route Table
      • Use when serving traffic to Internet
      • default subnets in default vpc are public (3 of them)
      • Route Table
        • Says all traffic to IG
        • Attach RT to subnet
      • Can set all instances to receive public IP
    • Even in Public, Instances needs Elastic/Public IP to comm w/ internet
      • Use a NAT for a Private Subnet to get updates
    • By default, instances w/in VPC can all communicate w/ e/o regardless of pub/private via "local route"
  • Can assign public IPs by default (routed to private IP (NAT))
  • NAT
    • NAT Gateway (AWS Provided)
      • Add to Public subnet,
      • Edit Route to Private Subnet (default if not explicit) to all traffic to nat
    • Instance
      • allow an instance in your VPC to initiate outbound connections to the Internet but prevent unsolicited inbound connections from the Internet
      • AWS provides instance types
      • Needs Public Subnet, Public IP
      • SG setting: Allow 80/443 from subnet CIDR
      • Disable source/destination check on EC2 instance
      • Add Route Table to Private Subnet pointing to NAT instance id
• In VPC, reboot keeps internal + elastic IPs (not so for non-ElasticIP public ip classic)
• Network ACLs (firewall at subnet level)
  • Live w/in a VPC. Can associate w/ multiple Subnets (Subnet has at most 1 Network ACL)
  • Block all traffic from an IP/range at Network ACL level
  • "stateless" - return traffic must be allowed through outbound rule
    • Security Groups - stateful, always allow return traffic
  • deny at low number trumps allow at high number (reverse NOT true)
  • increment rules by 10
  • Kind of like a Security Group, but for Subnets
• VPC Peering Connections
  • Can't do across regions (only between AZs in Region)
    • Can do across accounts w/in same Region
  • Can't have overlapping CIDR
  • Can do 1:Many where children can't see each other (not transitive)
  • Can configure Routes to act at VPC or Subnet level, or instance level
    • Edit Route Tables of BOTH VPC/Subnets to point to the PC
  • Only works w/in AWS (can't peer Internal network, use a VPG or DC instead)
  • Can't "cross over" into S3 endpoints
• Virtual Private Gateway
  • VPN connection from Customer Gateway to VPG in VPC
  • Add on-site route tables to AWS Subnet, apply to VPG
  • Could also run OpenVPN (not site-to-site)
    • Works w/ mobile devices that have OpenVPN client
    • Use w/ ElasticIP + backup OpenVPN to support High Availability
• Bastion Host
  • Log in from the web, it has access to private subnet resources
• Limits
  • 5 / VPCs region
    • 1 IG / VPC limit
    • ergo 5 IGs/region
  • 50 customer gateways / region
  • 50 VPN connections /region
  • 200 subnets/VPC
  • 200 Network ACLs/VPC, 20 rules/ACL
  • 200 route tables/VPC, 50 entries/table
  • 5 elastic IP/region
  • 500 Security Groups / VPC, 50 rules/SG, 5 SGs/network interface

RDS

• Automatic point-in-time backups, updates
• Multi-AZ Deployment
  • primary DB instance is synchronously replicated across Availability Zones to the standby replica
  • InnoDB only for MySQL
  • Read replicas for heavy read only traffic
  • avoids need for user-initiated point-in-time restore
• DB Subnet Groups
  • Needs multiple Subnets across (at least two) AZs
  • generally private subnets
  • Use DNS to connect to DB instance; IP can change on failover
• Publicly Accessible needs to be True (plus SG, etc) to be accessible
• Can move VPCs for non-Aurora dbs
• Supported DBs
  • MySQL // Aurora
  • MariaDB
  • Postgres
  • Oracle
  • MSSQL
• 5 GB to 6 TB of storage
• Can't do MySQL Clusters, would need to run on EC2
• RDS does NOT support a cluster of instances w/ load balancing traffic
• Can use ElasticCache clusters for caching db session info
• Shares same Security Groups as EC2
• Can encrypt data at rest
• Can use an SSL certificate for connections
• CNAME can be used w/ Route53 to give it a different DNS name
• Automated Backup
  • automated point in time recovery (serious)
  • default 1 day (free), can be set up to 35
• Snapshots
  • user initiated
• Encryption

DynamoDB

• Fully managed NoSQL (MongoDB): HA, scaling
• Sync'd across AZ within Region
• Dev specifies table throughput
• Document and Key/Value
• Can use SSL
• Use case: User session data

AWS Database Migration Service

• Done while running
• homogenous, heterogenous migrations
• Can stream to Redshift

Redshift

• petabyte-scale data warehouse for BI
  • hRedshift columns = 1024kb
• Supports SQL tools w/ ODBC/JDBC connections (Postgres-based)
• Columnar data store
• Monitors and backups data, can enable encryption

ElasticCache - in-memory data store

• Redis
• Memcached

Amazon Storage Gateway

• Local Storage that backs up to S3
• Gateway-Cached Volumes:
  • iSCSI mounted on-premise. Writes to S3, caches locally. Store all data in S3, cache most-frequently accessed locally.
  • less-limited, only cache data locally, everything else in S3 (cheaper?)
• Gateway-Stored Volumes:
  • local storage, periodic incremental snapshots
  • limited to the amount of space you allocate to the VM (eg potentially more costly)
• Recommend use w/ Direct Connect

Import/Export

• Mail hard drives
• Example: Baseline data into S3 (use incremental updates over internet)
• Snowball
  • Secure appliance (encrypted)

Direct Connect - dedicated private connection from ISP to AWS

• Not over internet
• Private Virtual Interface
  • Only internal IP addresses inside of EC2
• Public Virtual Interface
  • connect to public AWS endpoints
• Cross Connect
• Alternative to VPN
• 1 to 10 GB

IAM

• Global users, all AWS regions
  • New users start with no permissions
• Federated - can integrate w/ existing LDAP or Kerberos
• Simple Token Service
  • Temporary permissions for users/role
• SAML - integrate w/ active directory
• Groups -
  • Collection of IAM users
  • Deny overrides allow
• Roles - other AWS resources (users, EC2 instances can assume)
  • Identify Provider Access - gran on-premises networks Role access
  • Don't have API credentials (password or access keys)
  • EC2 Instance can only assume ONE Role when it is first created (cannot change or add)
  • Roles are always preferred to API keys
  • Default PowerUser role has access to everything except IAM
  • Delegation requires
    • Trust Policy
    • Permissions Policy
• Temporary Credentials
• Policies
  • Last 5 versions are tracked
• ARN - Amazon Resource Name (includes amazon account id)
• Security Token Service - assume a temporary role, do something (write to database), expires
• Cloudtrail - logs API calls
• Resource-level permissions
  • EC2
  • EBS: attach, delete, detach
  • Could require MFA for actions
• Cannot apply permissions to Root

AWS WAF

• Web application firewall, blocks common attacks (SQL injection, cross-site scripting)

Route53

• Apex of domain is "bare", w/o subdomain "example.com"
  • Route53 can Alias to
    • ELB
    • CloudFront
    • Elastic Beanstalk
    • S3 Bucket configured Static
    • other R53 record in zone
• Routing Policy
  • Simple
  • Latency based routing / multiregion failover
    • "Active-Active"
    • Routes to whichever has lowest latency
    • Requires duplicate architecture (use CloudFormation)
  • Weighted
    • "Active-Active"
    • Probabilistic off weighting
    • Useful for A/B tests
  • Geo-based Routing
    • Compliance w/ laws
  • Failover
    • Active-Passive
    • Example was ELB-Instance failing to CloudFormation-S3

SNS

• Notifications when events occur in AWS
• Topic: what a message is sent to
• Subscription/Subscriber: who/what gets the message
  • SMS, HTTPS,JSON,SQS
• Exists at Region level (e.g., don't pick an AZ)

SQS - Simple Queue Service

• Exists at Region level (e.g., don't pick an AZ)
• Distributed and Decoupled applications (Fault Tolerant)
• Messages up to 256KB
• Default message retention in queue is 4 days
  • Can set to 60s to 14 days
• Standard Queue
  • at-least-once message delivery
• FIFO (First In First Out)
• Maximum inflight messages
  • 120,000 standard
  • 20,000 FIFO
• Default Visibility Timeout is 30s
• Guaranteed delivery
• Order best effort
• Can create an Auto Scaling group for component based off queue size
• Long Polling
  • 1-20s waits for all messages in queue
  • Reduces cost when you get lots of empty responses
  • Queries all servers
• Short Polling (default)
  • increases API requests (increases cost)
  • returns some (not all) messages in queue (queries subset of servers)
• dead-letter queue
  • messages that cannot be processed successfully (poison-pill management)
  • Must be in the same Region as the feeder queue

SWF - Simple Work Flow Service

• Distributed and Decoupled applications (Fault Tolerant)
• coordinates asynchronously across multiple devices
• guaranteed order, no duplicates
• Execution can last up to 1 year
• Activity Task
• Decision Task
• API is task-oriented (SQS is message-oriented)

Amazon API Gateway

• Service to build RESFUL API to expose lambda, http, other

EMR - Elastic MapReduce

• Hadoop Master/Slave
  • Master, Core, Task Nodes
• Can launch apps on it (Hive, Pig compatible
• Launches a pre-built Hadoop cluster
  • You can login (unlike RDS)
• Input data from S3, DynamoDB, RedShift
  • S3 mounted by default
• Chunk out data into 128MB sizes (default, can be changed)
  • split files loaded into memory
• Preconfigured Hadoop AMI w/ Mappers/Reducers per instance size
• Can use CloudWatch to configure number of workers
• customers may encrypt the input data before they upload it to Amazon S3 (using any common data compression tool); they then need to add a decryption step to the beginning of their cluster when Amazon EMR fetches the data from Amazon S3.

Kinesis

• Collect data from multiple Producers, maintains order
• "Stream"
  • Preserves for 24hrs default, 7 days max
  • Data blob can be 1 MB
• Run SQL queries on streaming data
• Emit from Kensis Streams to S3, Redshift, EMR, Lambda (Consumers)
• Scale across multiple shards
• Aggregation refers to the storage of multiple records in a Streams record

Elastic Beanstalk

• Uses CloudFormation templates
• Deploy less-complex applications, single tier, core services
  • EC2
  • Auto Scaling
  • ELB
  • RDS
  • SQS
  • CloudFront
• Supported Platforms:
  • Docker
  • Java
  • Windows .NET
  • Node.js
  • PHP
  • Python
  • Ruby
• Don't use if:
  • Need software updates on boot?
• Integrates with version control
• Web server Environment vs Worker Environment
• Deploys an ELB, integrates w/ Route53

CloudFormation - create and provision resources w/ templates

• JSON Templates "infrastructure as code"
• Stack - template that deploys infrastructure
  • Can create from existing application architecture
  • 25 at a time
• Resources
  • AWS things to be launched
• Can configure Parameters that prompt for user input
• "Automatic rollback on error" enabled by default

CloudWatch

• Can be used to shut down inactive instances
• integrates w/ CloudTrail for AWS environment change monitoring
• Only accessible via SSL endpoint
• EC Instances "built in" hypervisor metrics
  • CPU
  • Network I/O
  • Disk Read/Write
• EC Instance custom metrics (non-RDS)
  • disk usage / free
  • swap
  • memory use /free
• Status Check
  • "self healing applications"
• Basic: 5 min, Detailed: 1 min

CloudTrail

• Log any action taken against AWS API, user details of who did it
  • account, IP, time, parameters, response

CloudFront - CDN, uses edge locations to serve data

• Edge Location - AWS datacenter w/o services
• origin (S3 bucket, ELB CNAME) to edge location
• distribution can have multiple origins (by device type)
  • Web or RTMP (Flash media)
  • Restrict Bucket Access: Only accessible via CloudFront, ignore Bucket policy
    • CloudFront adds ACL to Bucket Policy (OAI?)
    • S3 Bucket permissions still apply otherwise
• Private content
  • can create one time use "signed URLs" (or cookies) for private content
    • urls for RTMP, individual files
    • cookies for multiple files, or to not change URLs
  • specify ending datetime (optional: starttime, IP address range)
  • Limit S3 access to just CloudFront
    • Create Origin Access identity (OAI)
    • Give it S3 read permissions, lock out everyone else
• Price Class: varies by geographic spread
• Route53
  • can create alternate cnames for long CloudFront names (eg "cdn.mywebsite.com")
  • Can create two Alias records, one as primary one as failover (eg, ELB to Instance, CloudFront to S3)
    • "Active-Passive"
• caches files until
  • cache expires
  • or overwrite w/ new name
  • or create an "invalidation" (costs, might be cheaper to recreate a new CloudFront distribution & update CNAME)

AWS Config

• Point in time snapshot of AWS
• JSON
• Stores config changes

Trusted Advisor

• Notify you of security holes, High Availability issues, cost, performance
• According to AWS best practices

CloudHSM

• Hardware Security Module
• You can implement in multiple AZs and enable replication
• dedicated hardware to store keys, encryption, etc
• isolated physically, tamper resistant
• Compliance reasons, encrypt data/files
• AWS Engineers don't have access

Key Management Service

AWS CloudSearch

• Upload data, they index and provide a search interface

AWS ElasticSearch Service

ECS EC2 Container Service

• Start/Stop Docker containers

AWS Lambda

• Run code in response to "events"
• AWS handles all scaling