Regions are physical world locations where multiple Availability Zones exist.
They are physically isolated and independent from one another.
Regions come at no charge.

Availability Zones are sets of one or more data centers, each with their own resources, housed in separate facilities.

Resources created in one Region do not exist in any other Region, unless explicitly using replication features offered by AWS services.
Some services like IAM do not have Regional resources.

Recommended using regional STS endpoints instead of the global one to reduce latency.
Session tokens from regional STS endpoints are valid in all AWS Regions. However, tokens from the global endpoint are only valid in enabled Regions.

Session tokens valid in all Regions are larger. If storing session tokens, these might affect one's systems.

Regions introduced before 2019-03-20 are enabled by default. Newer regions are now disabled by default.
Regions enabled by default cannot be enabled or disabled.

Disabling Regions disables IAM access to resources in those Region. It will not delete resources in the disabled region, and they will continue to be charged at the standard rate.

Disabling a Region can takes a few minutes to several hours to take effect. Services and Console will be visible until the region is completely disabled.

Enabling Regions takes a few minutes to several hours. They cannot be used until the preparation process is complete.

The API for some AWS services (e.g. EC2) are eventually consistent.
This means that the result of an API request that affects resources might not be immediately visible to the subsequent requests that API receives.

Networking

VPCs define isolated virtual networking environments.
AWS accounts include one default VPC for each AWS Region.

Every VPC will have at least one CIDR block.
Every new AWS account will have one default VPC in every region, all with the 172.31.0.0/16 CIDR block assigned.

VPCs can be peered to enable direct connectivity between them via private IP addresses.
The peer connection also requires exchanging route table entries between the VPCs.

Subnets are virtual networks, each of which carves out smaller range of IP addresses from their VPC's CIDR block.
Each subnet resides in a single Availability Zone.
Public subnets have a direct route to an Internet gateway. Resources in public subnets can access the public Internet.
Private subnets do not have a direct route to an Internet gateway. Resources in private subnets require a NAT device to access the public internet.

Security groups control the traffic in and out of the resources associated with them, like firewalls would do.
Security group rules are stateful, meaning that connections initiated from the security group will allow the corresponding answers to come back in (but not new connections).
The default value for Egress traffic is to allow all connections.

Network Access Control Lists also control the traffic in and out. However, they are associated with subnets, and affect all resources within that subnet.
NACLs are stateless, meaning that both the Inbound and Outbound rules must match traffic patterns to allow communications in any direction.
NACL rules allow all traffic by default. They also have a priority.

Gateways connect VPCs to other networks.
Internet gateways connect VPCs to the Internet.
NAT gateways allow resources in private subnets to connect to the Internet, other VPCs, or on-premises networks. They can communicate with services outside the VPC, but cannot receive unsolicited connection requests.
VPC endpoints connect VPCs to AWS services privately, without the need of Internet gateways or NAT devices.

Route tables control how traffic flows throughout, and in or out, a VPC.
They are associated with subnets, and affect all resources within those subnets.
By default, a VPC only comes with a single route table. It is referred to as the Main route table.

By default, connections to AWS services use the services' public endpoint.

Traffic from instances in public subnets is routed to the VPC's internet gateway, then forwarded to the requested AWS service.

graph LR
  i(Internet)
  subgraph Region
    as(AWS Service)
    subgraph VPC
      ig(Internet<br/>Gateway)
      subgraph Public Network
        ei(Instance)
      end
    end
  end
  ei --> ig
  ig --> i
  i --> as

Traffic from instances in private subnets is routed to a NAT gateway, which forwards it to the VPC's internet gateway, which then forwards it to the requested AWS service.
This traffic does not leave the AWS network, even if it does traverse the internet gateway.

graph LR
  i(Internet)
  subgraph Region
    as(AWS Service)
    subgraph VPC
      ig(Internet<br/>Gateway)
      subgraph Public Network
        ng(NAT Gateway)
      end
      subgraph Private Network
        ei(Instance)
      end
    end
  end
  ei --> ng
  ng --> ig
  ig --> i
  i --> as

PrivateLink leverages VPC endpoints to create a private and direct connection between a VPC and a service inside the AWS backbone (usually, AWS-provided services).
Gateway endpoints do the same thing PrivateLink does, but in a more convenient way that does not require Elastic Network Interfaces. These endpoints are only supported by specific AWS services (S3 and DynamoDB at the time of writing).

graph LR
  i(Internet)
  subgraph Region
    direction LR
    subgraph VPC
      subgraph Public Network
        ng(NAT Gateway)
      end
      subgraph Private Network
        ei(Instance)
      end
      ge(Gateway<br/>Endpoint)
      ig(Internet<br/>Gateway)
    end
    as(AWS Service)
  end
  ei --> ge
  ge --> as

Direct Connect creates a dedicated network connection between on-premises data centers or offices and AWS.

Elastic IP addresses

Refer Elastic IP addresses.

Static, public IPv4 addresses allocated to one's AWS account until one releases it.
One can can rapidly remapping addresses to other instances in one's account and use them as targets in DNS records.

Proxying connections

AWS does not currently really offer a proxy-like service like HA Proxy or Squid would be.

Options:

Consider using PrivateLink, if the destination service is inside the AWS network.
Consider routing traffic to the destination directly through a NAT Gateway, if one can set static CIDRs in the subnets' Route Tables.
This is the most basic way to accomplish the goal, while keeping maintenance efforts at a minimum.

The idea is to:
1. Create, or reuse, NAT Gateways as the points of exit.
2. Create Route Tables accordingly.
3. Configure the Route Table to route traffic to the wanted CIDRs to the NAT Gateway serving the AZ.
Route Tables require defining sets of CIDRs.
If the destination is inside a cloud provider and do not use static IPs, one might need to add all of the cloud provider's CIDRs for the Region, which change without notice and require updating.
Even doing that, the sheer number of CIDRs might too high for a single Route Table to manage.
Route all traffic through an AWS Network Firewall, then re-route traffic to the destination service through a NAT Gateway.
Network Firewalls can re-route traffic based on its URL instead of based on its CIRDs.
This solves the issue requires Route Tables to use CIDRs, but requires all traffic in a Subnet to go through the Firewall's endpoint for its AZ.

The idea is to:
1. Create, or reuse, NAT Gateways as the points of exit.
2. Create Network Firewall Endpoints accordingly.
3. Configure the Network Firewalls to intercept all traffic from the subnets in the AZ each Endpoint covers, and re-route the packets for the destination to the NAT Gateway for the AZ.
Warning

Network Firewalls might be way too expensive to be used only for proxying purposes, with a baseline estimation just shy of €900/month only for 3 Endpoints existing (for AZs A to C, for instance). Add to this the charges for data processing for the entire traffic exiting or entering every Subnet served by a Network Firewall Endpoint.
Route traffic to the destination service through a custom proxy instance.

This will require clients to configure their tools to use the proxy.
This could be an EC2 instance, or some containerized solution.

Services

Service	Summary
Billing and Cost Management	Cost management
CloudFront	Content delivery
CloudWatch	Observability (logging, monitoring, alerting)
Config	Compliance
Detective	Behaviour anomalies
Direct Connect	Private on-premise to AWS connection
EC2	Managed virtual machines
ECR	Container registry
ECS	Run containers as a service
EFS	Serverless file storage
EKS	Managed Kubernetes clusters
EventBridge	Stream real time data
GuardDuty	Threat detection
IAM	Access control
Image Builder	Build custom AMIs
Inspector	Security vulnerability assessment
Kinesis	Video or data streams
KMS	Key management
OpenSearch	ELK, logging
PrivateLink	Private VPC to AWS service connection
RDS	Databases
Route53	DNS
S3	Storage
Sagemaker	Machine learning
Secrets Manager	Secrets management
Security Hub	Aggregator for security findings
SNS	Pub/sub message delivery
SQS	Queues
Step Functions	Task orchestration

Service icons are publicly available for diagrams and such. Public service IP address ranges are available in JSON form at https://ip-ranges.amazonaws.com/ip-ranges.json.

Billing and Cost Management

Costs can be grouped by Tags applied on resources.
Tags to use for this kind of grouping need to be activated in the Cost allocation tags section.
New tags might take 24 or 48 hours to appear there.

Config

Compliance service for assessing and auditing AWS resources.

Provides an inventory of resources.
Records and monitors resource configurations and their changes.
Allows for automatic remediation for non-compliant resources by leveraging Systems Manager Automation documents.

The service's data is stored in an S3 bucket.
The bucket is named config-bucket-{aws-account-number} by default and created upon service's activation.

The changes logs can be streamed to 1! SNS topic for notification purposes.

Uses rules to evaluate whether the resources configurations comply.
Rule evaluation is done either once every time a configuration changes, or periodically.
Resources are marked with the evaluation result (compliant, non-compliant).

Custom rules can be used to evaluate for uncommon requirements.
Custom rules leverage lambda functions.

Conformance packs are set of rules bundled together as a deployable, single, immutable entity.
Defined as YAML templates.
Users cannot make changes without updating the whole rule package.
Sample templates for compliance standards and benchmarks are available.

Detective

Uses ML and graphs to try and identify the root cause of security issues.
Creates visualizations with details and context by leveraging events from VPC Flow Logs, CloudTrail and GuardDuty.

Direct Connect

Provides a dedicated, private network connection from on-premise to the AWS network.

One needs to establish a dedicated connection from a on-premise location to an AWS Direct Connect location.
This typically involves a physical connection between the on-premises network and the AWS infrastructure.

Once the connection is established, one can create one or more virtual interfaces to act as logical connections within the physical Direct Connect connection.
Virtual interfaces can be configured with specific routing details, virtual private gateway associations, and bandwidth settings.

Direct Connect uses link layer encryption over the connection to secure the traffic.

Global Accelerator

Global service creating accelerators to improve the performance of applications.
Supports endpoints in multiple Regions.

Standard accelerators improve availability of Internet applications used by a global audience.
Global Accelerator directs traffic over AWS' global network to endpoints in the nearest Region to the client.
Endpoints for standard accelerators can be Network Load Balancers, Application Load Balancers, Amazon EC2 instances, or Elastic IP addresses located in one or more Regions.

Custom routing accelerators map one or more users to a specific destination among many.

Global Accelerator provides 2 static IPv4 addresses and 2 static IPv6 addresses (for dual-stack VPCs) that are associated with accelerators.
Those static IP addresses are anycast addresses from AWS' edge network and remain assigned to accelerators for as long as they exist, even if disabled and no longer accepting or routeing traffic.
When deleting accelerators, the static IP addresses assigned to it are lost.

Global Accelerator also assigns each accelerator a default DNS name, similar to a1234567890abcdef.awsglobalaccelerator.com for single-stack ones or similar to a1234567890abcdef.dualstack.awsglobalaccelerator.com for dual-stack ones, that points to the static IP addresses assigned to the same accelerator.

The static IP addresses provided by Global Accelerator serve as single fixed entry points for your clients.
They accept incoming traffic onto AWS' global network from the edge location that is closest to the users.
From there, traffic is routed based on the type of accelerator configured:

Standard accelerators route traffic to the optimal endpoint based on several factors including the user's location, the health of the endpoint, and the endpoint weights one configures.
Custom routing accelerators route each client to a specific EC2 instance and port in a subnet based on the external static IP address and listener port that one provided.

Global Accelerator terminates TCP connections from clients at AWS' edge locations and establishes a new TCP connection to one's endpoints.

Client IP addresses are preserved for endpoints on custom routing accelerators.
Standard accelerators have the option to preserve and access the client IP address for some endpoint types.

Global Accelerator continuously monitors the health of all standard accelerators' endpoints, and reroutes traffic for all new connections automatically.
Health checks are not used with custom routing accelerators and there is no failover, because one specifies the destination to route traffic to.

One can configure weights for one's endpoints in standard accelerators.
In addition, one can use the traffic dial in Global Accelerator to increase (dial up) or decrease (dial down) the percentage of traffic to specific endpoint groups.

Global Accelerator sets an idle timeout to its connections.
If no data has been sent nor received by the time that the idle timeout period elapses, it closes the connection.

Idle timeout periods are not customizable.

To prevent connection timeout, one must send a packet with a minimum of one byte of data in either direction within the TCP connection timeout window. One cannot use TCP keep-alive packets to maintain a connection open.

Idle timeouts are set to 340 seconds for TCP connections and 30 seconds for UDP connections.

Global Accelerator continues to direct traffic for established connections to endpoints until the idle timeout is met, even if the endpoint is marked as unhealthy or it is removed from the accelerator.
It selects a new endpoint, if needed, only when a new connection starts or after an idle timeout.

Refer How AWS Global Accelerator works for more and updated details.
Also see Using Amazon CloudWatch with AWS Global Accelerator.

GuardDuty

Threat detection service.

It continuously monitors accounts and workloads for malicious activity and delivers security findings for visibility and remediation.
Done by pulling streams of data from CloudTrail, VPC Flow Logs or EKS.

Member accounts can administer GuardDuty by delegation if given the permissions to do so.

Findings are potential security issues for malicious events.
Those are also sent to EventBridge for notification (leveraging SNS).
Each is assigned a severity value (0.1 to 8+).

Trusted IP List is a whitelist of public IPs that will be ignored by the rules.
Threat IP List is a blacklist of public IPs and CIDRs that will be used by the rules.

EventBridge

TODO

Inspector

TODO

Kinesis

TODO

KMS

Key material is the cryptographic secret of Keys that is used in encryption operations.

Enabling automatic key rotation for a KMS key makes the service generate new cryptographic material for the key every year by default.
Specify a custom rotation period to customize that time frame.

Perform on-demand rotation should you need to immediately initiate key material rotation.
This works regardless of whether the automatic key rotation is enabled or not. On-demand rotations do not change existing automatic rotation schedules.

KMS saves all previous versions of the cryptographic material in perpetuity to allow decryption of any data encrypted with keys.
Rotated key material is not deleted until the key itself is deleted.

Track the rotation of key material CloudWatch, CloudTrail, and the KMS console.
Alternatively, use the GetKeyRotationStatus operation to verify whether automatic rotation is enabled for a key and identify any in progress on-demand rotations. Use the ListKeyRotations operation to view the details of completed rotations.

When using a rotated KMS key to encrypt data, KMS uses the current key material.
When using the same rotated KMS key to decrypt ciphertext, KMS uses the version of the key material that was used for encryption.
One cannot select a particular version of key materials for decrypt operations. This automation allows to safely use rotated KMS keys in applications and AWS services without code changes.

Automatic key rotation has no effect on the data that KMS keys protect: it does not rotate the data generated by rotated keys, re-encrypts any data protected by the keys, nor it will mitigate the effect of compromised data keys.

KMS supports automatic and on-demand key rotation only for symmetric encryption keys with key material that KMS itself creates.
Automatic rotation is optional for customer managed KMS keys. KMS rotates the key material for AWS managed keys on an yearly basis. Rotation of AWS owned KMS keys is managed by the AWS service that owns the key.

Key rotation only changes the key material, not the key's properties.
The key is considered the same logical resource, regardless of whether or how many times its key material changes.

Creating a new key and using it in place of the original one has the same effect as rotating the key material in an existing key.
This is considered a manual key rotation and is a good choice to rotate keys that are not eligible for automatic key rotation.

AWS charges a monthly fee for the first and second rotation of key material maintained for each key.
This price increase is capped at the second rotation. Any subsequent rotations will not be billed.

Each key counts as one when calculating key resource quotas, regardless of the number of rotated key material versions.

PrivateLink

Allows private access to AWS services from within a VPC without the need of Internet Gateways.

It needs one to create:

One interface VPC endpoint for each desired AWS service.
Each interface endpoint establishes a connection between the subnet they are in and their specific AWS service by means of Elastic Network Interfaces.
A route in the subnet's route table for each interface endpoint.
This will route all traffic destined to the AWS service referred by the interface endpoint to the endpoint.

The service's endpoint is resolved with the private IP addresses of the relative endpoint's network interface, instead of the default public endpoint.
This allows traffic to be sent to the requested AWS service using the private connection between the VPC endpoint and the AWS service.

graph
  subgraph Region
    as(AWS Service)
    subgraph VPC
      e(VPC Endpoint)
      subgraph Private Network
        ei(Instance)
        ie(Interface<br/>Endpoint)
      end
    end
  end
  ei --> ie
  ie --> e
  e --> as

Services will not be able to initiate requests to resources through their VPC endpoint.

Bills per each VPC endpoint provisioned, per each Availability Zone, per hour, per GB of data processed.
Refer AWS PrivateLink Pricing.

Security Hub

Aggregator of findings for security auditing.

Uses Config to check resources' configuration by leveraging compliancy rules.

Security standards are offered as ret of rules for Config.

Data can be aggregated from different regions.
If the integration is enabled, findings from AWS services (GuardDuty) are used too within 5 minutes on average, while ones from 3rd parties can take longer.

Data can be imported from or exported to 3rd parties if the integration is enabled.
Kinda acts as a middle layer for AWS accounts.

Findings are consumed in AWS Security Finding Format (ASFF).
Those are automatically updated and deleted. Findings after 90 days are automatically deleted even if not resolved.

Can use custom insights.

Custom actions can be sent to EventBridge for automation.

Member accounts can administer Security Hub by delegation if given the permissions to do so.

Step Functions

Refer What is Step Functions?.

Workflows (A.K.A. state machines) for building applications, automating processes, orchestrating microservices, and creating pipelines.
Can also be long-running and require human interaction.

Step Functions call AWS services or external workers to perform tasks.
They can also call other Step Functions in various ways (wait for finish, just start, …). See Start a new AWS Step Functions state machine from a running execution.

In the context of Step Functions:

State machines are called workflows.
Workflows are a series of event-driven steps.
Each step in a workflow is known as state.
Task states represent units of work performed by AWS services, like calling another service.
Instances of running workflows performing tasks are called executions in Step Functions.
Activities represent units of work executed by workers that exist outside of Step Functions.

Workflows can be:

Standard, if they run each step exactly once, for long time.
They can run for up to 1y, are auditable, and show execution history and visual debugging.

Step Functions counts a state transition each time a step in a standard workflow is executed.
It charges for the total number of state transitions across all one's state machines, including retries.
4000 state transitions per month are free, then they are charged $0.025 per every 1000 transitions.
Charges are metered daily and billed monthly.
Express, if they run each step at least once, for up to 5 minutes.
They are ideal for high-event-rate workloads like streaming data processing and IoT data ingestion.

Pricing is based off of number of requests and duration.

Step Functions counts a request each time it starts executing an express workflow.
It charges for the total number of requests across all the workflows, including tests from the console.

Duration is calculated:
1. From the time a workflow begins executing until it completes or otherwise terminates.
2. Rounded up to the nearest 100ms.
The amount of memory used in the execution of a workflow is billed in 64MB chunks.
Memory consumption is based on the size of a workflow definition, the use of map or parallel states, and the execution (payload) data size.

Step functions require to assume an IAM Role during execution.
Such roles need to allow being assumed by the states.amazonaws.com Principal.

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Effect": "Allow",
            "Principal": {
                "Service": [
                    "states.amazonaws.com"
                ],
            },
            "Action": "sts:AssumeRole"
        }
    ]
}

If wanting to send logs to CloudWatch, the execution role must be able to access the log group.

Example: get an RDS DB instance's information and pass it as a specific attribute

{
  "Comment": "Get an RDS DB instance's information and pass it as a specific attribute",
  "StartAt": "Get RDS DB Instance information",
  "States": {
    "Get RDS DB Instance information": {
      "Type": "Task",
      "Arguments": {
        "DbInstanceIdentifier": "some-rds-db-instance-identifier"
      },
      "Resource": "arn:aws:states:::aws-sdk:rds:describeDBInstances",
      "End": true,
      "Output": {
        "RdsDbInstanceInformation": "{% $states.result.DbInstances[0] %}"
      }
    }
  },
  "QueryLanguage": "JSONata"
}

Unless one knows exactly what one is doing, prefer setting arguments from the service's Console to have some level of suggestions.

FIXME - using variables in DescribeDBSnapshots

{
  "DbInstanceIdentifier": "{% $states.input.ProdDBInstanceInfo.DbInstanceIdentifier %}"
}

{
  "AvailableDBSnapshots": "{% $states.result.DbSnapshots[Status='available'] %}"
}

Resource constraints

Data type	Component	Summary	Description	Type	Length	Pattern	Required
Statement ID	Value	Optional identifier for a policy statement	The element supports only ASCII uppercase letters (A-Z), lowercase letters (a-z), and numbers (0-9).	String	FIXME	`[A-Za-z0-9]`	No
Tag	Key	Required name of the tag	The string value can be Unicode characters and cannot be prefixed with "aws:". The string can contain only the set of Unicode letters, digits, white-space, `_`,' `.`, `/`, `=`, `+`, `-`, `:`, `@` (Java regex: `^([\\p{L}\\p{Z}\\p{N}_.:/=+\\-]*)$`)	String	1 to 128	`^([\p{L}\p{Z}\p{N}_.:/=+\-@]*)$`	Yes
Tag	Value	The optional value of the tag	The string value can be Unicode characters. The string can contain only the set of Unicode letters, digits, white-space, `_`, `.`, `/`, `=`, `+`, `-`, `:`, `@` (Java regex: `^([\\p{L}\\p{Z}\\p{N}_.:/=+\\-])$"`, `[\p{L}\p{Z}\p{N}_.:\/=+\-@]` on AWS)	String	0 to 256	`^([\p{L}\p{Z}\p{N}_.:/=+\-@]*)$`	Yes

Access control

Refer IAM.

Costs

One pays for data transfer between instances and services in the same region but different availability zone.
See Understanding data transfer charges.

One pays for sending logs and metrics to CloudWatch.

Available discount options:

Discount type	Discount range	Commitment length	Flexibility	Applies to	Limitations
Free tier	100%	1 year	None	Selected services (EC2, S3, Lambda)	Low usage limits Overuse is billed Only available to new accounts for 1y
Spot instances	Up to 90%	None	High (for stateless/batch)	EC2, EMR, ECS, EKS, Batch	Can be terminated anytime Avoid for critical or long-running workloads
Savings plans	Up to 72%	1 or 3 years	Medium to high	EC2, Lambda, Fargate	Must commit to a $/hour spend Not cancellable Unused commitment is wasted
Reserved instances	Up to 75%	1 or 3 years	Low (standard) or Medium (convertible)	EC2, RDS, Redshift, ElastiCache	Specific instance type/region Harder to manage Non-refundable
Tiered pricing	Various	None	Automatic depending on usage	S3, CloudFront, Lambda, DynamoDB	Requires high volume Tiers and availability vary by service
Enterprise discount program	Custom (10 to >30%)	Custom (1 to 3 years typically)	High (custom contract)	All AWS	Requires large spend Enterprise-only Contract-based

Use case	Best discount options
Long-term predictable workloads	Savings plans or Reserved instances
Short-term batch or flexible tasks	Spot instances
New to AWS / testing a service	Free tier
High-volume services (e.g., S3 storage)	Tiered pricing
Large-scale enterprise planning to stay on AWS for some time	Enterprise discount program + Savings plans

Order of application: reserved instances -> Savings plans (EC2 instances -> Compute)

Free Tier

New AWS customers get 1 year of free tier access to selected services only.

Only allows for limited monthly usage (E.G., up to 750 hours of t2.micro EC2, 5GB S3 per month).

Free tier is only available in specific regions.
Usage in multiple regions counts as a whole. FIXME: check

Automatically charges standard rates when one exceeds one's account limits.

Spot Instances

Lends spare EC2 instance capacity at up to 90% discount.
Prices vary based on regional supply and demand.

Instances can be interrupted by AWS with a 2-minute warning.
Not suitable for workloads needing guaranteed uptime or long-term execution.

Works great for stateless, fault-tolerant, or batch workloads.

Savings plans

Refer Savings Plans user guide.
See also Understanding how Savings Plans apply to your usage.

Pricing models offering lower prices compared to On-Demand prices. They require specific usage commitments ($/hour) for 1-year or 3-years terms.

Dedicated Instances, Spot Instances and Reserved Instances are not discounted by Savings Plans.

Savings Plan	Included resources	Up to
Compute	EC2 instances regardless of family, size, AZ, region, OS or tenancy Lambda Fargate	66%
EC2 Instance	Individual EC2 instance families in a specific region (e.g. M5 usage in N. Virginia) regardless of AZ, size, OS or tenancy	72%
Amazon SageMaker	Eligible SageMaker ML instances, including SageMaker Studio Notebook, SageMaker On-Demand Notebook, SageMaker Processing, SageMaker Data Wrangler, SageMaker Training, SageMaker Real-Time Inference, and SageMaker Batch Transform regardless of instance family, size, or region	64%

Both Compute and EC2 Instance plan types apply to EC2 instances that are a part of Amazon EMR, Amazon EKS, and Amazon ECS clusters. They do not apply to RDS instances.
Charges for the EKS service itself will not be covered by Savings Plans, but the underlying EC2 instances will be.

Savings Plans are available in the following payment options:

No Upfront: no upfront payments, commitment charged purely on a monthly basis.
Partial Upfront: lower prices, at least half of one's commitment upfront, remainder charged on a monthly basis.
All Upfront: lowest prices, entire commitment charged in one payment at the start.

Savings Plans can be purchased in any account within an AWS Organization/Consolidated Billing family.
By default, the benefits of the Plans are applicable to usage across all accounts. One can choose to restrict the benefit of the Plans to only the account that purchased them.

One account can have multiple Savings Plans active at the same time.

Plans cannot be cancelled during their term.
Plans can be returned only if:

They consist in an hourly commitment of $100 or less.
They have been purchased in the past 7 days and in the same calendar month.

Once returned, one will receive a 100% refund for any upfront charges for the Savings Plan.
Refunds will be reflected in one's bill within 24 hours of return.

Any usage covered by the plan will be charged at On-Demand rates, or get covered by a different Savings Plans if applicable.

Plans do not provide capacity reservations.
One can however reserve capacity with On Demand Capacity Reservations and pay lower prices on them with Savings Plans.

Savings Plans are applied after Reserved Instances.
Furthermore, EC2 Instance Savings Plans are applied before Compute Savings Plans.

Savings Plans are applied to the highest savings percentage first. If there are multiple usages with equal savings percentages, Savings Plans are applied to the first usage with the lowest Savings Plans rate.
Savings Plans continue to apply until there are no more remaining usages, or one's commitment is exhausted. Any remaining usage is then charged at the On-Demand rates.

Reserved instances

Gives discounts of up to 75% compared to On-Demand pricing for EC2, RDS, Redshift and ElastiCache instances in exchange for an advance payment for either 1 or 3 years.

Available as follows:

Standard: higher discounts, but very little flexibility.
Limits to specific instance types, regions, OS, etc for the whole duration of the term.
Convertible: lower discounts, but can switch instance families, OS, or tenancy during the term.

Reserved Instances are available in the following payment options:

No Upfront: no upfront payments, commitment charged purely on a monthly basis.
Partial Upfront: lower prices, at least half of one's commitment upfront, remainder charged on a monthly basis.
All Upfront: lowest prices, entire commitment charged in one payment at the start.

Reserved instances plans are usually hard to manage at scale, and can lead to unused capacity if your usage changes.

Tiered pricing

Selected services like S3, Lambda, CloudFront, and DynamoDB offer automatic tiered pricing.
Tiered pricing lowers the per-unit cost of resources the more one uses them (E.G., S3 gets cheaper per GB as one stores more and more data).

Tiered pricing requires large usage volumes to see meaningful savings.

Enterprise discount program

Large customers, with high level of committed AWS spend (typically hundreds of thousands to millions per year), can negotiate custom discounts and support terms.

These negotiations are only available to large enterprises, and require long-term contractual commitment.

Other tools

AWS offers tools that can help optimize cost:

Cost Explorer: analyzes past usage and helps forecast costs and savings.
Trusted Advisor: provides recommendations for RIs, underutilized resources, etc.

Resource tagging

Refer What are tags?, Tagging best practices and strategies, Best Practices for Tagging AWS Resources and Tag naming limits and requirements.

Tags are labels consisting of a key and an optional value.
One can apply them to resources in order to add metadata to them.

Most, but not all, AWS services and resource types currently support tags. See Services that support the Resource Groups Tagging API for a list of which of them do.
Services other then the ones in the list may support tags via their own APIs.

Note

Tags are not encrypted.
They should not be used to store sensitive data like personally identifiable information (PII).

Tags that users create are known as user-defined tags.

Several AWS services automatically assign tags to those resources that they create and manage.
These keys are known as AWS generated, tags and are usually prefixed with aws:.
As such, the aws: prefix cannot be used in user-defined tag keys.

User-defined tags have usage requirements, and there are limits on the number that can be added to any AWS resource.
AWS generated tags do not count against these limits.

AWS generated tags use a namespace format, e.g. aws:cloudformation:some-stack.
User-defined tags can also use that format.

Using an organizational identifier as a prefix in tags is recommended to help identify where tags come from.

Suggested tags from AWS:

Tag	Value example	Notes
`Name`	`GitlabRunner`, `Prometheus Server`	Shows as human-friendly name in the AWS console
`Owner`	`SecurityLead`, `SecOps`, `Workload-1-Development-team`
`BusinessUnitId`	`Finance`, `Retail`, `API-1`, `DevOps`
`Environment`, `example-org:devops:environment`	`Sandbox`, `Dev`, `PreProd`, `QA`, `Prod`, `Testing`
`CostCenter`, `company-b:CostCenter`	`FIN123`, `Retail-123`, `Sales-248`, `HR-333`
`FinancialOwner`	`HR`, `SecurityLead`, `DevOps-3`, `Workload-1-Development-team`
`ComplianceRequirement`	`NIST`, `HIPAA`, `GDPR`

Create tag policies to enforce values, and to prevent the creation of non-compliant resources.

Note

Once created, tags cannot be deleted.

If unused for some time, they will not show up in services like Cost Explorer.
Deletion aside, one can manage tags in the Tag Editor.

API

Refer Tools to Build on AWS.

Python

Refer Boto3 documentation.
Also see Difference in Boto3 between resource, client, and session?.

Clients and Resources are different abstractions for service requests within the Boto3 SDK.
When making API calls to an AWS service with Boto3, one does so via a Client or a Resource.

Sessions are fundamental to both Clients and Resources and how both get access to AWS credentials.

Client

Provides low-level access to AWS services by exposing the botocore client to the developer.

Typically maps 1:1 with the related service's API and supports all operations for the called service.
Exposes Python-fashioned method names (e.g. ListBuckets API => list_buckets method).

Typically yields primitive, non-marshalled AWS data.
E.g. DynamoDB attributes are dictionaries representing primitive DynamoDB values.

Limited to listing at most 1000 objects, requiring the developer to deal with result pagination in code.
Use a paginator or implement one's own loop.

Example

import boto3

client = boto3.client('s3')
response = client.list_objects_v2(Bucket='mybucket')
for content in response['Contents']:
    obj_dict = client.get_object(Bucket='mybucket', Key=content['Key'])
    print(content['Key'], obj_dict['LastModified'])

Resource

Refer Boto3 resources.

Provides high-level, object-oriented code.

Does not provide 100% API coverage of AWS services.

Uses identifiers and attributes, has actions (operations on resources), and exposes sub-resources and collections of AWS resources.

Typically yields marshalled data, not primitive AWS data.
E.g. DynamoDB attributes are native Python values representing primitive DynamoDB values.

Takes care of result pagination.
The resulting collections of sub-resources are lazily-loaded.

Resources are not thread safe and should not be shared across threads or processes.
Create a new Resource for each thread or process instead.

Since January 2023 the AWS Python SDK team stopped adding new features to the resources interface in Boto3.
Newer service features can be accessed through the Client interface.
Refer More info about resource deprecation? for more information.

Example

import boto3

s3 = boto3.resource('s3')
bucket = s3.Bucket('mybucket')
for obj in bucket.objects.all():
    print(obj.key, obj.last_modified)

Session

Refer Boto3 sessions.

Stores configuration information (primarily credentials and selected AWS Region).
Initiates the connectivity to AWS services.

Leveraged by service Clients and Resources.
Boto3 creates a default session automatically when needed, using the default credential profile.
The default credentials profile uses the ~/.aws/credentials file if found, or tries assuming the role of the executing machine if not.

Container images

Amazon Linux

Refer Pulling the Amazon Linux container image.

Amazon Linux container images are infamous for having issues when connecting to their package repositories from outside of AWS' network.
While it can connect to them sometimes™ when running locally, one can get much easier and more consistent results by just running it from inside AWS.

Disconnect from the VPN, start the container, and reconnect to the VPN before installing packages when running the container locally.
If one can, prefer just build the image from an EC2 instance.

README.md

Amazon Web Services

TL;DR

Networking

Elastic IP addresses

Proxying connections

Services

Billing and Cost Management

Config

Detective

Direct Connect

Global Accelerator

GuardDuty

EventBridge

Inspector

Kinesis

KMS

PrivateLink

Security Hub

Step Functions

Resource constraints

Access control

Costs

Free Tier

Spot Instances

Savings plans

Reserved instances

Tiered pricing

Enterprise discount program

Other tools

Resource tagging

API

Python

Container images

Amazon Linux

Further readings

Sources