EC2

Amazon EC2 Quick Notes

Scaling Vocabulary

• Vertical (scale up/down): Bigger/smaller instance type (e.g. t2.nano → metal). Needs reboot; has a size ceiling.
• Horizontal (scale out/in): More/fewer instances behind a load balancer.

Storage — EBS vs Instance Store

EBS, Instance store

Attached how: Network volume, Physically on the host

Persistence: Survives stop; detachable, Tied to instance lifecycle

Use when: Persistence / snapshots, Extreme local I/O; app can rebuild

• Bootable: gp2/gp3, io1/io2 (and supported instance store). st1 / sc1 cannot be boot volumes.
• Cost tip: io1 dominating the bill without needing provisioned IOPS → move to gp2/gp3.
• DeleteOnTermination: Root EBS is deleted on terminate by default — set to false to keep it.
• Encryption: At rest on the volume, on snapshots, and in transit between instance ↔ volume.
• Recycle Bin: Retention rules can recover accidentally deleted EBS snapshots/AMIs.

Purchasing Options

Option: Idea

On-Demand: Pay by the second; no commitment

Reserved / Savings Plans: Commit 1–3 years for discount (steady 24×7)

Spot: Spare capacity; interruptible; deep discount

Spot Fleet: Target capacity across types/AZs

Dedicated Host: Physical host visibility (licensing)

Dedicated Instance: Single-tenant hardware, no host visibility (cheaper single-tenant path)

• Spot trap: Canceling a Spot request does not terminate a running Spot instance.
• Persistent Spot request: Can reopen and replace after interruption.
• Tenancy after launch: Can change dedicated ↔ host. Cannot freely flip default ↔ dedicated/host.

AMI Lifecycle

1. Launch → configure → Create Image
2. Copy AMI to another Region → creates AMI + snapshot there
3. Launch from that AMI → destination has instance + AMI + snapshot

• Sharing: AMIs can be shared across accounts (encrypted AMIs need correct KMS grants).
• Golden AMI: Bake static installs; leave dynamic bits in user data (long Beanstalk bootstraps).

User Data

• Runs only on first launch by default (not every restart).
• Executes as root.
• Cannot update user data on a running instance to re-run it by default.

Hibernation

• Saves RAM to the root EBS volume so stop/start resumes in-memory state.
• Use when bootstrap takes minutes and you need fast wake.

Placement Groups

Type: Fit

Cluster: Low latency / HPC — one AZ

Spread: Separate racks — max ~7 instances per AZ

Partition: Large distributed ETL / Hadoop-style isolation

• EFA (Elastic Fabric Adapter): OS-bypass HPC networking for tightly coupled MPI workloads.

Instance Access & Ops

• Instance profile / role: Prefer over access keys on disk.
• EC2 Instance Connect: Inject a temporary SSH public key.
• CloudWatch recover: System-status failure → recover keeps instance ID / private IPs / EIPs (auto public IPv4 may change).
• Detailed monitoring: 1-minute metrics (vs basic 5-minute).
• AZ IDs: us-west-2a is randomized per account — use AZ ID to align physical AZs across accounts.