Decision rule: Give people and workloads temporary credentials through roles, grant the least they need where the access happens, and fence every member account with SCPs, remembering that an explicit deny anywhere always wins.
Evaluation rule
Result
Any matching explicit Deny
Denied, whatever allows exist
Same account
Identity policy or resource policy must allow
Cross-account
Caller's identity policy and owner's resource or trust policy allow
SCP, RCP, permissions boundary
Only cap permissions; never grant
Limit or behavior
Value
Role maximum session duration
1 to 12 hours, default 1 hour
Role chaining session
1 hour maximum
SCPs attached per root, OU or account
10
SCP document size
10,240 characters
RCPs attached per root, OU or account
5, including the automatic RCPFullAWSAccess policy
IAM groups per user
Groups can't be nested; they can contain only users
Access keys per IAM user
2
Root user MFA devices
Up to 8
SCPs and RCPs in the management account
No effect
SCPs and service-linked roles
No effect
If the question says…
Reach for…
Workforce sign-in to many accounts
IAM Identity Center with the company IdP or AD
No one in member accounts may…
SCP on the OU or root
New multi-account setup with guardrails
AWS Control Tower
Workload calls AWS APIs
IAM role (instance profile or execution role)
Third party needs access
Cross-account role with an external ID
Share a resource with the whole organization
Resource policy with aws:PrincipalOrgID
Delegate role creation safely
Permissions boundary
1.2 Secure workloads and applications
Decision rule: Put each tier in the most private subnet that still works, let stateful security groups that reference each other carry the allow rules, keep network ACLs for subnet-wide denies, reach AWS services through endpoints instead of the internet, and stop exploits with AWS WAF and floods with AWS Shield.
Filter or control
Behavior
Security group
Stateful, allow only, instance level; source can be another security group
Network ACL
Stateless, allow and deny, subnet level; rules checked in ascending number, first match wins
Network ACL return traffic
Needs rules for the client's ephemeral ports; the range depends on the client OS (32768-61000 on many Linux kernels, 49152-65535 on Windows Server 2008 and later)
Default network ACL
Allows all traffic; a custom network ACL denies everything until rules are added
Security group rule quota
60 inbound and 60 outbound per group (default); 5 groups per network interface
Network ACL rule quota
20 inbound and 20 outbound per ACL (default); up to 40 each on request
NAT gateway
Outbound only for private subnets; 5 Gbps, scales to 100 Gbps; 55,000 connections per destination per IP
Gateway endpoints
S3 and DynamoDB only; route-table entry; no additional charge
Interface endpoints
Most AWS services through PrivateLink; one network interface per chosen subnet; security group applies
Parameter Store
Standard: 4 KB, 10,000 per account per Region; advanced: 8 KB, 100,000 per account per Region
Secrets Manager
Up to 65,536 bytes per secret; RDS managed master password rotates every 7 days by default
Site-to-Site VPN
Two IPsec tunnels per connection; 1.25 Gbps per standard tunnel, 5 Gbps large bandwidth tunnel on TGW or Cloud WAN
Direct Connect encryption
Not encrypted by default; MACsec on 10, 100 and 400 Gbps dedicated connections at selected locations
Automatic at no additional charge; layer 3 and 4 DDoS protection
If the question says…
Reach for…
Only the app tier may reach the database
Database security group that names the app tier's group
Block one IP range from a subnet
Network ACL deny with a lower number than the allows
Private instances download updates
NAT gateway in a public subnet, private route to it
Private access to S3 or DynamoDB
Gateway endpoint plus a bucket policy with aws:SourceVpce
Private access to other AWS services
Interface endpoint in one subnet per Availability Zone
SQL injection or bad bots in HTTP requests
AWS WAF web ACL with managed rule groups on CloudFront, ALB or API Gateway
Per-IP request flood
AWS WAF rate-based rule
Large DDoS with expert help and cost protection
AWS Shield Advanced
Database password rotated automatically
Secrets Manager, managed rotation for RDS
Configuration values, AMI IDs
Parameter Store standard tier
App users sign in, then reach AWS resources
Cognito user pool, then identity pool
Admin shell without SSH ports or bastions
Systems Manager Session Manager
Encrypted connection to a data center, quick setup
Site-to-Site VPN
Private circuit that must be encrypted
Direct Connect with MACsec, or a VPN over Direct Connect
Detect compromised credentials or cryptomining
Amazon GuardDuty
Find personal data in S3
Amazon Macie
Find CVEs on EC2, ECR or Lambda
Amazon Inspector
1.3 Data security controls
Decision rule: Match each control to the failure it stops: a customer managed KMS key with a key policy you own for encryption, ACM certificates for TLS, Block Public Access before any bucket policy, Object Lock for deletes that must never happen, and AWS Backup with vault lock plus a copy in another Region or account for recovery.
Control or setting
Behavior
AWS managed key (aws/s3)
AWS owns the key policy and yearly rotation; you cannot change its properties or schedule its deletion; you cannot use it directly in cryptographic operations
Customer managed key
Your key policy; optional automatic rotation (default every year, custom 90 to 2,560 days); disable; deletion waits 7 to 30 days
KMS key policy
Must allow access; IAM policies alone cannot grant KMS access the key policy does not allow
KMS Encrypt payload
Up to 4,096 bytes; larger data uses envelope encryption with GenerateDataKey
Default bucket encryption
SSE-S3 is the base level for every bucket; SSE-KMS with no named key uses aws/s3
RDS encryption
Set only at creation; encrypt an unencrypted instance by copying its snapshot with encryption turned on, then restoring from the encrypted copy
ACM public certificate
No charge for non-exportable public certificates on integrated services; exportable certificates are charged at issuance and renewal (see ACM pricing); renews itself while DNS validation stays in place
ACM for CloudFront
Certificate must be requested in US East (N. Virginia)
ACM email validation
Being deprecated for new public certificates and renewals throughout 2027; move to DNS validation (ACM docs)
S3 Block Public Access
Four settings, at account or bucket level; refuses public policies and ACLs
S3 Object Lock governance mode
Blocks normal deletes; users with s3:BypassGovernanceRetention can remove it
S3 Object Lock compliance mode
No user, including root, can shorten or remove retention before it ends
S3 legal hold
No end date; independent of retention; stays until set to off
S3 Object Lock prerequisites
Versioning is required; replicating locked objects needs Object Lock on the destination
S3 Replication
Needs versioning on both buckets; copies overwrites and, if enabled, delete markers; version-specific deletes are not replicated, so it is not a backup on its own
AWS Backup vault lock compliance
Immutable after the grace time (cooling-off period you set, at least 72 hours); sets minimum and maximum retention for jobs into the vault
Multi-Region KMS key
Same key ID and key material in several Regions (KMS docs)
If the question says…
Reach for…
Disable the key or control its key policy
Customer managed KMS key
Another account must decrypt
Key policy allows the other account's role; its IAM policy allows the KMS actions
ACM public certificate, DNS validation, on the ALB or CloudFront
Backups that no admin can delete
AWS Backup vault lock, compliance mode, plus copies to another account
Key scheduled for deletion by mistake
KMS CancelKeyDeletion before the waiting period ends (the key is left Disabled), then EnableKey (KMS docs)
Domain 2: Design Resilient Architectures
2.1 Scalable and loosely coupled architectures
Decision rules:
Messaging (2.1a): Name the shape before the service: work done once is a queue (SQS), a copy for every subscriber is a topic (SNS), routing by content or AWS/SaaS events is an event bus (EventBridge), ordered steps with state are a workflow (Step Functions).
Compute (2.1b): Pick the most managed compute whose limits the workload fits: Lambda for short event-driven code, Fargate for containers without servers, EC2-backed ECS or EKS for GPUs or host control; EKS only when the stem says Kubernetes.
Scaling (2.1c): Scale out stateless copies behind the right load balancer, then cache whatever repeats: CloudFront at the edge, the API Gateway cache for repeated responses, ElastiCache for hot data and sessions, read replicas for read-heavy databases.
Limit or behavior
Value
SQS message size
Up to 1 MiB
SQS retention
1 minute to 14 days, default 4 days
SQS visibility timeout
0 seconds to 12 hours, default 30 seconds; for Lambda, at least 6x the function timeout
SQS FIFO throughput (no high-throughput mode)
300 transactions per second per partition per API action without high-throughput mode; with batches of 10, up to 3,000 messages per second per API action
SQS FIFO deduplication interval
5 minutes
SQS long polling wait
Up to 20 seconds
EventBridge targets per rule
5
Step Functions duration
Standard up to 1 year (exactly-once); Express up to 5 minutes (async at-least-once, sync at-most-once)
API Gateway account throttle
10,000 requests per second with a 5,000 burst, per account per Region, shared across HTTP, REST and WebSocket APIs
API Gateway REST integration timeout
50 ms minimum; 29 seconds by default, which can be raised above 29 seconds (may lower the Region throttle quota)
API Gateway cache TTL
Default 300 seconds, maximum 3,600 seconds (REST APIs only)
Lambda timeout and memory
15 minutes; 128 MB to 10,240 MB
Lambda concurrency
1,000 per account per Region by default (adjustable)
Lambda payload
6 MB synchronous, 1 MB asynchronous
Lambda package
250 MB unzipped (.zip); 10 GB container image
CloudFront default TTL
86,400 seconds (1 day) when the origin sends no caching headers
If the question says…
Reach for…
Buffer spikes, retry later
SQS standard queue + DLQ
Strict order, no duplicates
SQS FIFO with a message group ID
Same event to several systems
SNS fan-out to one SQS queue each
Route AWS service or SaaS events by content
EventBridge rule
Multi-step process with waits and retries
Step Functions Standard
Existing JMS, AMQP or MQTT app
Amazon MQ
Partners send files over SFTP
AWS Transfer Family into S3
Over 15 minutes, containerized, no servers
ECS on Fargate
Existing Kubernetes or Helm tooling
Amazon EKS
GPUs
ECS or EKS on EC2 capacity
No cold starts on a Lambda API
Provisioned concurrency
HTTP path or host routing
Application Load Balancer
TCP or UDP, static IPs
Network Load Balancer
Global users, shared content
CloudFront
Sessions lost on scale-in
Sessions in ElastiCache or DynamoDB
Read-heavy database
Read replicas (details in 3.3)
2.2 High availability and disaster recovery
Decision rules:
Strategy (2.2): Match the DR strategy to the stated RPO and RTO and choose the lowest one that meets both. Multi-AZ covers a data center failure; a second Region is needed only for a Region-level requirement.
Failover: Use health-checked failover routing in Route 53 and Multi-AZ failover for databases. Keep failover on data plane operations; scaling and weighted changes are control plane actions.
Standby readiness: The recovery Region must be able to scale to production capacity, so raise its service quotas and keep its infrastructure defined as code.
DR strategy
RPO
RTO
What runs in the recovery Region
Backup and restore
Hours (minutes with PITR)
Up to 24 hours
Nothing; backups only
Pilot light
Minutes
Tens of minutes
Replicated data; app servers switched off
Warm standby
Seconds
Minutes
Full stack at reduced capacity
Multi-site active/active
Near zero
Potentially zero
Full stack serving traffic
Limit or behavior
Value
RDS Multi-AZ DB instance failover
Typically 60 to 120 seconds; standby is in another AZ and cannot serve reads
RDS Multi-AZ replication
Synchronous to the standby
Aurora global database failover
Replication lag typically under a second; secondary promoted typically under a minute per the FAQ (the user guide frames RTO in minutes)
20 per account per Region by default (adjustable through Service Quotas)
RDS Proxy endpoints
Default endpoint spans two AZs; up to 20 additional endpoints per proxy
RDS Proxy placement
Same VPC as the database; the proxy itself is not publicly accessible
If the question says…
Reach for…
Survive an AZ failure, no Region requirement
Multi-AZ (Auto Scaling across AZs, Multi-AZ database)
Lowest cost, RPO in hours
Backup and restore with cross-Region backup copies
RPO in minutes, core database always replicated
Pilot light
Standby must serve reduced traffic immediately
Warm standby
Every Region serves users, RTO near zero
Multi-site active/active
Cross-Region database failover, typically under a minute per the FAQ (user guide frames RTO in minutes)
Aurora global database
DNS failover between Regions with health checks
Route 53 failover routing
Same database endpoint after failover, no app change
RDS Multi-AZ
Connection spikes or "too many connections" during failover
RDS Proxy in front of the database
Recovery Region drifted after hand patches
Immutable infrastructure: replace from a tested image
Data must survive loss of one Availability Zone
S3 or a Regional EFS file system (stored across AZs)
Restore an EBS volume in another zone or Region
EBS snapshot, copied to the recovery Region
Domain 3: Design High-Performing Architectures
3.1 High-performing and scalable storage
Decision rule: choose the storage type from how the data is reached (object over an API, a shared file mount, or a disk on one instance), then tune that service to the number in the stem.
Limit or behavior
Value
S3 request rate per partitioned prefix
Per prefix: 3,500 PUT/COPY/POST/DELETE and 5,500 GET/HEAD per second; above that, spread requests across prefixes
S3 single PUT
Up to 5 GB; multipart upload recommended from 100 MB
S3 multipart upload
Up to 10,000 parts of 5 MiB to 5 GiB
EBS gp3
3,000 IOPS and 125 MiB/s included; up to 80,000 IOPS and 2,000 MiB/s; 1 GiB to 64 TiB
EBS io2 Block Express
Up to 256,000 IOPS and 4,000 MiB/s; average latency under 500 microseconds
EBS st1 / sc1
Up to 500 / 250 MiB/s; HDD, cannot be boot volumes
EBS Multi-Attach
io1 and io2 only; up to 16 Nitro instances in one Availability Zone
EBS Elastic Volumes
Change size, type, IOPS, throughput while attached; size can only grow
EFS latency (Standard)
About 1 ms reads, 2.7 ms writes (Regional, General Purpose, best case); Windows EC2 cannot mount
EFS modes
General Purpose (default, recommended); Elastic (recommended for spiky or unpredictable workloads); Bursting (scales with storage); Max I/O previous gen
Decision guide
If the question says…
Reach for…
Because…
503 Slow Down from S3
More prefixes, retries
Request rate scales per prefix
Global users upload to one bucket
S3 Transfer Acceleration
Edge locations shorten the long-distance path
Single-digit ms object access next to compute
S3 Express One Zone
Lowest-latency S3 class, one AZ
One instance, a database, set IOPS
EBS gp3
IOPS and throughput set apart from size
Sub-ms latency or beyond gp3 limits
EBS io2 Block Express
Highest IOPS and lowest latency per volume
Rebuildable cache or scratch, fastest local disk
Instance store
Host-attached, lost on stop or terminate
Many Linux servers share files across AZs
Amazon EFS
Regional NFS file system that scales on its own
Windows SMB with Active Directory
FSx for Windows File Server
Native Windows file system
HPC or ML on an S3 dataset
FSx for Lustre linked to S3
Sub-ms latency, very high throughput
NFS, SMB and iSCSI from one system
FSx for NetApp ONTAP
Multiprotocol with ONTAP features
On-prem files should land in S3
S3 File Gateway
NFS or SMB in front, objects behind, local cache
On-prem iSCSI volumes backed by AWS
Volume Gateway (cached or stored)
Cached: primary in AWS. Stored: primary local
Tape backup software
Tape Gateway
Virtual tape library over iSCSI
Migrate a NAS once or on a schedule
AWS DataSync
Managed transfer, not ongoing access
Storage classes and lifecycle rules for cost: see 4.1.
3.2 High-performing and elastic compute
Decision rule: pick the compute shape from how the work arrives and the instance family from the resource that runs out first, then let Auto Scaling follow a metric that grows with load per instance, not a raw total.
Limit or behavior
Value
EC2 type name
Series letter + generation + options (g Graviton, n network, d instance store) + size
T family (standard mode)
Held at baseline CPU once credits are spent; T8i, T4g, T3a, T3 launch as unlimited by default
Default EC2 metrics
Every 5 minutes; detailed monitoring gives 1 minute; memory needs the CloudWatch agent
Simple scaling default cooldown
300 seconds; target tracking and step scaling use instance warmup instead
Default instance warmup
Optional group setting; AWS suggests starting at 300 seconds if unsure
Predictive scaling
Needs 24 hours of data, uses up to 14 days, forecasts 48 hours; scales out only
Warm pool instance states
Stopped, running or hibernated
Default health check
EC2 status checks only; turn on ELB health checks behind a load balancer
Cluster placement group
One AZ; no burstable instances; 10 Gbps single flow inside versus 5 Gbps outside
Partition placement group
Up to 7 partitions per AZ
Spread placement group
Up to 7 running instances per AZ per group
Lambda memory
128 MB to 10,240 MB; CPU scales with memory (1,769 MB = one vCPU); 15-minute maximum
Decision guide
If the question says…
Reach for…
Because…
CPU-bound
C family
Most vCPU per GiB of memory
Memory-bound
R or X family
Most memory per vCPU
Training, inference, graphics
P, G, Trn
GPUs or ML chips
High local disk IOPS
I or D family
Large instance store
Low average CPU with bursts
T family
Baseline plus CPU credits
Load follows a metric
Target tracking
Auto Scaling manages the alarms
SQS workers
Target tracking on backlog per instance
Falls as workers are added
Peak at a known time
Scheduled action
Capacity before the load
Daily or weekly pattern, slow boot
Predictive scaling plus a dynamic policy
Launches ahead; dynamic policy scales in
Long boot, unpredictable spikes
Warm pool
Pre-initialized instances join fast
Tightly coupled HPC
Cluster placement group, EFA
Lowest latency between nodes
HDFS, HBase, Cassandra
Partition placement group
Rack isolation per partition
A few critical instances
Spread placement group
Distinct hardware per instance
Queued jobs that run to completion
AWS Batch
Managed job queues and compute environments
CPU-bound Lambda function
More memory
CPU is allocated in proportion to memory
Lambda, Fargate, ECS and EKS: see 2.1b. Spot, Savings Plans and rightsizing for cost: see 4.2.
3.3 High-performing databases
Decision rule: choose the database by its access pattern, scale relational reads with replicas and a cache, and scale DynamoDB by spreading traffic across partition key values rather than by raising table capacity.
Limit or behavior
Value
RDS read replicas (MySQL, MariaDB)
Up to 15 per source in one Region; asynchronous; watch ReplicaLag
Aurora Replicas
Up to 15 per cluster; share the cluster volume; lag typically tens of milliseconds within a Region
Aurora reader endpoint
Balances connections, not individual queries; writes need the writer endpoint
Aurora storage
Grows in 10 GB steps up to 256 TiB on Aurora MySQL 3.10 and later (128 TiB before 3.10)
RDS storage autoscaling
Up to a maximum storage threshold; can't decrease storage; limited to 4 storage modifications per 24 hours
gp3 baseline (Db2, MariaDB, MySQL, PostgreSQL)
3,000 IOPS and 125 MiB/s below 400 GiB; 12,000 IOPS and 500 MiB/s from 400 GiB. Oracle: from 200 GiB. SQL Server: no threshold
io2 Block Express
Up to 256,000 IOPS (io2 Block Express storage)
DynamoDB per partition
3,000 read units and 1,000 write units per second
DynamoDB item size
400 KB maximum
DynamoDB GSI
Eventually consistent only; sparse; up to 20 per table
DynamoDB LSI
Defined at table creation; same partition key; strongly consistent reads; up to 5 per table
DynamoDB on-demand
Absorbs up to double the previous peak instantly; new tables start with 12,000 read and 4,000 write units per second; warm throughput is configurable
DAX
Item cache TTL default 5 minutes; up to 10 read replicas per cluster; strongly consistent reads not cached
ElastiCache replicas
Valkey or Redis OSS: up to 5 replicas per shard; Memcached: no replication
Decision guide
If the question says…
Reach for…
Because…
Reads slow the relational writer
RDS read replicas or Aurora Replicas
Reads move off the writer
Many read connections, no replica names
Aurora reader endpoint
Follows replicas as they change
Repeated identical query results
ElastiCache with lazy loading and TTL
Answers repeats; strategy in 2.1c
Sorted sets, lists, or replicated cache
Valkey or Redis OSS
Data structures and replicas
Microsecond DynamoDB reads, eventual is fine
DAX
Caches items and queries
One hot partition key value throttled
Higher-cardinality partition key or write sharding
Per-partition limit applies to each key
Query by a non-key attribute
Global secondary index
Own key; eventually consistent
Unannounced bursts, no forecast
DynamoDB on-demand
Scales with traffic
Storage fills during a large load
Allocate storage before the load
Autoscaling can't keep pace
gp3 IOPS baseline too low
Grow gp3 to 400 GiB or more (200 GiB on Oracle), or io2 Block Express
Baseline rises to 12,000 IOPS (not RDS for SQL Server)
Database must move to a different engine
SCT for schema and code, DMS for data
Heterogeneous: schema converted first
Same engine, new home
Homogeneous migration (native tools or DMS)
Schema and SQL carry over
Failover, Multi-AZ and RDS Proxy: see 2.2. Backups and encryption: see 1.3. Database cost and Aurora Serverless v2 pricing: see 4.3.
3.4 High-performing and scalable network architectures
Decision rule: pick the network path by where the traffic starts and where it must go: Global Accelerator for global users who need fixed IPs, a transit gateway once VPCs outnumber a handful, Direct Connect for steady high bandwidth, and an endpoint to keep AWS service traffic on a short private path.
Limit or behavior
Value
Global Accelerator static IPs
Two IPv4 addresses per accelerator (four with dual-stack); endpoints are ALB, NLB, EC2 or Elastic IP
Transit gateway attachments
Up to 5,000 per transit gateway (adjustable); one attachment per VPC
Transit gateway bandwidth
Up to 100 Gbps per VPC attachment per AZ, each direction
Transit gateway MTU
8500 bytes between VPCs, Direct Connect and peering attachments; VPN attachments accept packets up to 1500 bytes, but the usable Site-to-Site VPN tunnel MTU is 1446 bytes (MSS 1406); no jumbo frames
VPC peering
One-to-one, not transitive; no overlapping CIDR blocks; 9001 MTU in a Region, 8500 across Regions
Site-to-Site VPN tunnel
Standard up to 1.25 Gbps; Large Bandwidth Tunnel up to 5 Gbps on a transit gateway or Cloud WAN only, not in five Regions (ap-southeast-4, ca-west-1, eu-central-2, il-central-1, me-central-1)
VPN aggregate
ECMP needs dynamic routing; two Large Bandwidth connections with four tunnels reach 20 Gbps
Accelerated VPN
Transit gateway attachments only; chosen at creation, not switched on later
Direct Connect dedicated ports
1, 10, 100 or 400 Gbps
Direct Connect hosted ports
50 Mbps to 25 Gbps (partner-provisioned); traffic policing drops bursts above the rate
Gateway endpoints
S3 and DynamoDB only; route table entry to the prefix list
Interface endpoint throughput
10 Gbps per AZ by default, scaling to 100 Gbps per AZ; MTU 8500; endpoints don't send the ICMP messages Path MTU Discovery needs
Single-flow bandwidth (EC2)
5 Gbps outside a cluster placement group; 10 Gbps inside; ENA Express 25 Gbps single flow on supported instances
Internet multi-flow (EC2)
Through an internet or local gateway: fewer than 32 vCPUs capped at 5 Gbps; more than 32 vCPUs capped at 50 percent of the instance bandwidth
ENA and EFA
ENA up to 100 Gbps on supported types (for example C5n); EFA adds OS-bypass for HPC and ML
Decision guide
If the question says…
Reach for…
Because…
Global users, fixed IPs, TCP or UDP, fast regional failover
Global Accelerator
Two static IPs, edge entry and health-based endpoint choice
Cacheable HTTP content, global users
CloudFront (see 2.1c)
Caching at the edge; Global Accelerator never caches
Two or three VPCs, direct link only
VPC peering
Direct and cheap, but not transitive
Many VPCs, or VPCs plus on-premises through one hub
Transit gateway
Transitive routing with route tables per segment
Steady high bandwidth, consistent latency
Direct Connect (dedicated, or a LAG)
Private circuit at 1 to 400 Gbps per port
Fast setup over the internet, modest bandwidth
Site-to-Site VPN
Two tunnels per connection, 1.25 Gbps each
VPN above 1.25 Gbps on a transit gateway
Large Bandwidth Tunnels, then ECMP
5 Gbps per tunnel; dynamic routing adds connections
Private S3 or DynamoDB from the VPC
Gateway endpoint
Route table entry, no interface to size
Other services, or on-premises or peered callers
Interface endpoint
Private IP in your subnets, reachable over VPN and Direct Connect
One flow above 5 Gbps in one Region
ENA Express
25 Gbps single flow on supported instances
Tightly coupled HPC or ML nodes
EFA in a cluster placement group
OS-bypass with Libfabric for MPI and NCCL traffic
Route 53 routing is in 2.2; placement groups are in 3.2; data transfer and NAT cost are in 4.4.
3.5 High-performing data ingestion and transformation
Decision rule: match the ingestion path to the arrival pattern and the consumer: Kinesis Data Streams for ordered real-time reads, Firehose for buffered delivery to S3, MSK when the team already speaks Kafka, and Glue, EMR or Athena for transforms, with Parquet and partitions deciding how fast the data is read later.
Limit or behavior
Value
Kinesis shard, provisioned writes
1 MB per second or 1,000 records per second per shard
Kinesis shard, reads
2 MB per second and 5 GetRecords calls per second per shard, shared by standard consumers
Enhanced fan-out
2 MB per second per registered consumer per shard; up to 20 consumers per stream (50 on On-demand Advantage, in listed Regions only)
Kinesis on-demand scaling
Up to double the peak write rate of the previous 30 days; a larger spike can still throttle
Kinesis capacity mode switch
Twice per 24 hours, between on-demand and provisioned
Kinesis PutRecords
Up to 500 records per call; HTTP 200 can still carry failed entries, each with an ErrorCode to retry
Firehose buffer to S3
Size 1 to 128 MB (default 5); interval 0 to 900 seconds (default 300); first condition met triggers
Firehose record format conversion
JSON to Parquet or ORC; S3 is the only destination; buffer size at least 64 MB; Glue table is the schema
Glue worker G.1X
1 DPU: 4 vCPUs and 16 GB memory; Auto Scaling needs Glue 3.0 or later and does not support Standard DPUs
Athena partition projection
Computes partitions from table properties, avoiding a Glue Data Catalog partition lookup per query
DataSync multiple agents
Up to four Basic and four Enhanced mode agents per location; no high availability, all agents must be online
Decision guide
If the question says…
Reach for…
Because…
Several real-time consumers, one needing dedicated read throughput
Kinesis Data Streams with enhanced fan-out
2 MB per second per shard for each registered consumer
Unpredictable traffic, no shard planning
Kinesis Data Streams on-demand
Scales to double the 30-day peak; still size for known spikes
Files in S3 with no consumer code
Firehose to S3, buffer tuned for file size
Batching and retries handled; larger buffer, larger files
JSON stream that analysts query with SQL
Firehose Parquet conversion, Glue schema
Columnar, compressed files land in S3
Existing Apache Kafka code, topics or clients
Amazon MSK
Kafka APIs kept; no client rewrite
Managed Spark ETL that reprocesses only new files
Glue ETL with job bookmarks
Managed workers, Auto Scaling, skip processed data
Custom big data frameworks or cluster control
Amazon EMR on EC2 or EKS
You choose the cluster shape and frameworks
SQL over S3 data with no ETL
Athena with Parquet and partition projection
Column pruning and partition pruning cut bytes scanned
One place to restrict tables, columns or S3 paths
Lake Formation with a registered location
Table and column permissions across Athena, Glue and EMR
Applications write files over NFS or SMB, and the files must reach S3
AWS Storage Gateway (S3 File Gateway)
Keeps the file share on premises and copies each file into S3
Dashboards and charts over the S3 lake, with no code
Amazon Quick, with Athena as the source
Visualizes queried data; it does not ingest or transform it
Producers in a VPC write to Kinesis privately
Interface VPC endpoint, with an IAM role per producer
PrivateLink keeps traffic on the Amazon network; the role limits the streams
Domain 4: Design Cost-Optimized Architectures
4.1 Cost-optimized storage
Decision rule: pay for the access pattern you actually have: match each dataset's read frequency and tolerable retrieval time to the cheapest tier that still meets them, then let lifecycle rules move and expire data for you.
S3 storage class
Reads about
First byte
Min duration
Min billable size
S3 Standard
More than monthly
Milliseconds
None
None
S3 Intelligent-Tiering
Unknown or changing
Milliseconds (archive tiers: restore)
None
Under 128 KB not tiered
S3 Standard-IA
Monthly
Milliseconds
30 days
128 KB
S3 One Zone-IA (one AZ)
Monthly, re-creatable
Milliseconds
30 days
128 KB
S3 Glacier Instant Retrieval
Quarterly
Milliseconds
90 days
128 KB
S3 Glacier Flexible Retrieval
Yearly
Restore: 1–5 min, 3–5 h, 5–12 h
90 days
About 40 KB added
S3 Glacier Deep Archive
Less than yearly
Restore: within 12 h or 48 h
180 days
About 40 KB added
Limit or behavior
Value
Lifecycle default for small objects
Objects under 128 KB are not transitioned
Lifecycle direction
Down the waterfall only; leaving Flexible or Deep Archive needs restore + copy
Intelligent-Tiering moves
Infrequent Access after 30 days, Archive Instant Access after 90 days without access
Requester pays requests and downloads; owner pays storage
gp3 baseline (in the storage price)
3,000 IOPS and 125 MiB/s at any size; gp2 to gp3 online with Elastic Volumes
EBS volume size
Can grow, can't shrink
st1 / sc1
Cheaper HDD for sequential or cold data; never a boot volume
EBS Snapshots Archive
90-day minimum; restore up to 72 hours
EFS Archive storage class
90-day minimum; Regional file systems with Elastic throughput
EFS One Zone
One AZ, reduced cost compared to Standard; only for data you can re-create
FSx for Windows File Server levers
HDD storage for general shares; deduplication typically saves 50–60% on general-purpose file shares
FSx for NetApp ONTAP levers
Auto tiering policy moves cold data to the capacity pool; size SSD for active data
FSx for Lustre levers
Scratch for temporary processing; not replicated, keep the durable copy in S3
AWS Backup cold storage
Warm for at least one week; at least 90 days in cold; not every resource type supports it
DataSync vs Transfer Family
DataSync copies NFS/SMB/object data over your link; Transfer Family is SFTP/FTPS/FTP/AS2 for partners
Cost allocation tags
Must be activated before they appear in Cost Explorer
Cost Explorer / Budgets / CUR
Explorer to explore, Budgets to alert (actual or forecast), CUR to analyze (S3 + Athena)
4.2 Cost-optimized compute
Decision rule: rightsize first, then buy for the shape of the usage: Spot for anything that can be interrupted, a Savings Plan or Reserved Instances for the steady baseline, On-Demand for the rest, and nothing at all for capacity that is idle.
Purchase option
Max discount
Commitment
Flexibility
Reserves capacity
On-Demand
None
None
Anything
No
Compute Savings Plans
Up to 66%
USD per hour, 1 or 3 years
Any family, size, Region, OS, tenancy; Fargate, Lambda
No
EC2 Instance Savings Plans
Up to 72%
USD per hour, 1 or 3 years
Size, OS, tenancy within one family in one Region
No
Standard RI
Up to 72%
One configuration, 1 or 3 years
Can be sold on the RI Marketplace; not exchangeable
Only if zonal
Convertible RI
Up to 66%
One configuration, 1 or 3 years
Exchangeable; can't be sold
Only if zonal
Spot Instances
Up to 90%
None
Interrupted with a two-minute notice
No
Capacity Reservation
None itself
None for immediate start; future-dated has a commitment duration and can carry a cancellation charge
One instance type in one AZ
Yes, billed if unused
Limit or behavior
Value
Savings Plans and Spot
Savings Plans don't discount Spot, and Spot spend doesn't count toward the commitment
Savings Plans and capacity
Never reserve capacity; they do discount On-Demand Capacity Reservations
Spot reliability
Many instance types and AZs, price-capacity-optimized, Capacity Rebalancing
Spot is wrong for
Stateful, fault-intolerant or tightly coupled work; workloads that need guaranteed capacity
Compute Optimizer
Opt-in; 14-day lookback (93 with paid enhanced metrics); 30 hours of EC2 data minimum
Compute Optimizer memory
Only with the CloudWatch agent (or a supported external tool); it never resizes on its own
Hibernation
Enabled at launch only; encrypted EBS root sized for RAM; Linux RAM under 150 GiB
Non-production off-hours
Scheduled actions set desired capacity (and min) to 0, then back in the morning
Lambda / Fargate / EC2
Lambda for spiky or mostly idle; Fargate bills for the task CPU and memory you specify; EC2 for steady use
Discount sharing (Organizations)
Commitments apply to the buying account first, then other accounts; management account controls it
Savings Plans budget
AWS Budgets alerts on Savings Plans utilization or coverage below a target
Savings Plan return
Within 7 days of purchase, same calendar month, hourly commitment $100 or less; yearly return quota applies
4.3 Cost-optimized databases
Decision rule: price the database by its load shape and its data's lifetime: commit to the steady baseline, let on-demand or Serverless v2 carry the unpredictable part, expire data you no longer need, and cache repeated reads so the database can stay smaller.
Billing or storage choice
Use when
Key limit or behavior
DynamoDB on-demand
Traffic is unpredictable or new
Pays per request; costs more per request than well-used provisioned capacity
DynamoDB provisioned with auto scaling
Traffic is steady and forecastable
Pays for provisioned units; reserved capacity lowers the rate on the floor
DynamoDB Standard-IA
Storage exceeds 50% of the read and write cost and reads are rare
Storage about 60% lower; reads and writes about 25% higher; no reserved capacity
DynamoDB TTL
Items expire at a known time
Epoch-seconds number attribute; deletion is background and can lag; filter expired items from reads
Aurora Standard vs I/O-Optimized
Compare both bills; I/O-heavy favors I/O-Optimized
I/O-Optimized has no per-request I/O charge but higher instance and storage rates; pays off when I/O spend exceeds 25% of the bill; switch once every 30 days
Aurora Serverless v2
Idle or spiky load
Minimum 0 ACUs enables auto-pause; storage and backups still billed
RDS storage autoscaling
Growth is hard to forecast
Raises allocated storage; can't decrease it; allocate before very large loads
Free backup storage
Keep retention to what the requirement names
RDS: 100% of provisioned storage per Region per account; Aurora: latest cluster volume size, for automated backups
Multi-AZ standby
Availability requirement (see task 2.2)
Standby instance and storage are billed; it is not a backup
Database Savings Plans
Committed spend across Aurora, RDS, DynamoDB, ElastiCache
Up to 35% off; one-year term; flexible across those services
DynamoDB reserved capacity
Steady provisioned baseline, one or three years
Blocks of 100 capacity units; up to 54% off (1 year) or 77% (3 years) in select Regions; provisioned Standard table class only
RDS reserved DB instances
Fixed database configuration for a fixed term
One- or three-year term; one instance type, engine and Region; does not reserve capacity
ElastiCache reserved nodes
Cache cluster runs at a steady size
Discounted rate on pre-purchased nodes; Valkey priced below Redis OSS
Redshift Serverless
Bursty warehouse queries
Billed per RPU-hour while queries run; set usage limits; steady warehouse fits provisioned
ElastiCache in front of the database
The same reads repeat
Only repeated reads are saved; the cache is itself a running cost
4.4 Cost-optimized network
Decision rule: data transfer is priced by direction and path: keep traffic in one Availability Zone where you can, send private AWS traffic through gateway endpoints, use NAT gateways only for low-volume egress, and move high-volume internet traffic onto CloudFront or Direct Connect.
Path
Hourly charge
Per-GB charge
Cost note
Into AWS from the internet
None
None
Free
Out to the internet
None
Tiered per GB
Most expensive per-GB path; put high volume behind CloudFront
Same AZ, private IP
None
None
Usually free over private IPs; co-locate chatty tiers
Between AZs in one Region
None
Both directions
Billed for inbound and outbound
Between Regions
None
Outbound from the source Region only
Destination inbound has no charge
Gateway endpoint (S3, DynamoDB)
None
None
Always use before NAT for these services
Interface endpoint
Per endpoint per AZ
Per GB processed
Cheaper than NAT only at volume
NAT gateway, zonal
Per gateway per AZ
Per GB processed through the gateway
One per AZ with workloads; NAT plus cross-AZ adds cost
NAT gateway, regional
Per AZ-hour it is configured in
Per GB processed through the gateway
Expands across AZs with workloads; HA by default; not for private connectivity
NAT instance
EC2 instance cost
Normal data transfer, no processing fee
Self-managed; not the production answer
VPC peering
None
Data transfer rate for the path
Cheapest link for two or a few VPCs
Transit gateway
Per VPC attachment, per hour
Per GB processed; none on data from a peering attachment
Pays off when the peering mesh becomes hard to run
Site-to-Site VPN
Per connection
Internet data transfer rate
Low volume and backup
Direct Connect
Port-hour charge
Data transfer out at the DX rate; in is $0.00/GB
Steady high volume to on-premises
CloudFront
None; usage-based pricing
Out to viewers; AWS origin to edge is waived
Best path for high-volume public content; AWS origin to edge transfer is free
Global Accelerator
Per accelerator, per hour
Data transfer plus an incremental charge on top of standard rates
Buy it for latency, static IPs or cross-Region failover, not to save money
Route 53 alias to an AWS resource
Hosted zone per month
None for alias queries to a load balancer, CloudFront, Global Accelerator or S3 website
A CNAME to a Route 53 record is billed as two queries per lookup
Public IPv4 address
Per address, in use or idle
None
Share with a load balancer or NAT; check the CUR line
Trap
Correct answer
One NAT gateway for workloads in three AZs
One zonal NAT gateway per AZ with workloads, or a regional NAT gateway
NAT gateway carrying S3 or DynamoDB traffic
Gateway endpoint for S3 or DynamoDB (free)
Interface endpoint for S3 to save money
Gateway endpoint for S3; interface endpoints charge per hour and per GB
NAT instance in production to save money
Managed NAT gateway per AZ
Site-to-Site VPN for terabytes a month
Direct Connect for steady high-volume hybrid traffic
Transit gateway for two VPCs to save money
VPC peering (no hourly charge)
Ignoring public IPv4 charges
Every public IPv4 address is billed per hour, in use or idle