Amazon EC2

Amazon EC2 (Elastic Compute Cloud) is the virtual-machine service AWS launched in 2006, and it still underpins a large share of all AWS workloads. You pick a size, an operating system, a Region, and an Availability Zone; AWS hands you a running server in seconds. You connect over SSH or RDP and use it like any other machine.

What changes in the cloud is everything around the instance. You start ten or ten thousand with one command, stop them when traffic drops, and replace a failed one in a minute — paying by the second for what runs. EC2 is the default when you need full control of the operating system or run software that does not fit a managed service.

Instance Types

The instance type fixes how much CPU, memory, network, and storage an instance gets. A name like m5.large has three parts: a family letter (what it is good at), a generation number (higher is usually newer and cheaper per unit of work), and a size (nano through 48xlarge).

Family	Best for	Examples
T	Small, bursty workloads	t3.micro, t4g.small
M	General-purpose work	m6i.large, m7g.xlarge
C	CPU-heavy work	c7g.large, c6i.4xlarge
R	Memory-heavy work	r6i.large, r7g.2xlarge
I	Fast local NVMe SSDs	i4i.large
G / P	GPU work (graphics, ML)	g5.xlarge, p4d.24xlarge

The pattern beats memorization: start with an M-family instance, move to C if you are CPU-bound, move to R if you are memory-bound. Stay on the latest generation — newer generations are usually faster and cheaper for the same workload.

AMIs and Storage

An AMI (Amazon Machine Image) is the template an instance boots from — operating system, pre-installed software, and configuration. A common pattern bakes your software into a custom AMI so startup is fast and predictable; tools like EC2 Image Builder and Packer build them repeatably.

Almost every instance boots from an EBS volume — network-attached storage that persists independently of the instance, so it survives a stop or terminate and can be detached and reattached elsewhere. Instance store is local NVMe that is far faster but disappears when the instance stops; use it only for caches and rebuildable scratch data.

Networking and Security

Every instance runs inside a VPC with a private IP, and optionally a public IP to be reachable from the internet. A security group is a stateful firewall around the instance: inbound is deny-all (you open ports explicitly), outbound is allow-all by default, and return traffic for an allowed connection is permitted automatically.

A key pair handles the first login — AWS holds the public key, you keep the private key. For access without managing keys at all, use Systems Manager Session Manager instead of opening SSH.

Pricing Models

Choosing the pricing model is one of the largest cost levers on AWS. On-Demand pays by the second with no commitment. Reserved Instances and Savings Plans trade a one- or three-year commitment for up to ~70% off; Savings Plans are usually the easier choice because the discount spans instance types and services. Spot buys spare capacity at up to ~90% off, with the catch that AWS can reclaim it on two minutes' notice.

Model	Discount	Commitment	Best for
On-Demand	none	none	Spiky, short-lived workloads
Reserved	up to ~70%	1 or 3 years	Steady production workloads
Savings Plans	up to ~70%	1 or 3 years	Steady spend, flexible usage
Spot	up to ~90%	none — reclaimable	Fault-tolerant batch work

Most real systems mix models: cover the steady baseline with a Savings Plan, absorb daily spikes On-Demand, and run interruptible batch work on Spot.

Scaling with Auto Scaling Groups

An Auto Scaling Group manages a fleet of instances from a launch template, holding a desired capacity between a minimum and maximum and moving it with scaling policies tied to CloudWatch metrics like CPU or ALB request count. Unhealthy instances are replaced automatically; behind a load balancer this is the classic high-availability pattern.

Scaling is not instant — each new instance launches, configures, and passes health checks, and each action respects a cooldown. The instance-count curve always lags the traffic curve by a few minutes. Design for the lag with a minimum capacity that absorbs the first spike.