Load Balancing & Auto Scaling

ELB & Auto Scaling Quick Notes

Load Balancer Types

LB, Layer, Fit

ALB: 7 (HTTP/HTTPS), Path/host/header routing, content-based

NLB: 4 (TCP/UDP/TLS), Extreme performance, static IP, non-HTTP

GWLB: 3/4 gateway, Scale firewall / appliance fleets

CLB: Legacy, Prefer ALB/NLB

ALB Routing

• Path-based: /orders vs /products
• Host-based: test.example.com Host header
• Also: headers, query string, source IP
• SNI: Many TLS certs on one listener (multi-site behind one ALB)

Cross-Zone Load Balancing

Assume 1 instance in AZ-A and 4 in AZ-B:

Mode: Distribution

Enabled: Equal per instance → 20% each

Disabled: Equal per AZ → AZ-A 50%; each of AZ-B’s four gets 12.5%

Connection Draining

• Keep in-flight requests alive when a target deregisters / goes unhealthy.
• ALB/NLB name: deregistration delay.
• Sticky sessions / idle timeout / cross-zone do not solve dropped in-flight requests.

Health Check Failures (targets look fine)

1. Instance SG does not allow traffic from the load balancer SG
2. Health-check path/port misconfigured

NLB Targeting

• Routes to the instance’s primary private IP on the primary ENI.

Auto Scaling Groups

• ASG: When/where (min / desired / max, AZs).
• Launch template: What (AMI, type, user data, …).
• Mixed On-Demand + Spot across instance types: Requires a launch template (not launch configuration).

Scaling Policies

Policy: Idea

Manual: Change desired by hand

Scheduled: Known peaks — set desired before the spike

Simple / Step: Alarm-driven adds/removes (bands for step)

Target tracking: “Keep CPU at 50%”

Predictive: Learns daily/weekly patterns; scales before the surge

• Steady baseline: ASG min ≥ 2 across AZs + cover baseline with RIs / Savings Plans.

Maintenance Without Replacement

• Put instance in Standby, patch, exit Standby — or suspend ReplaceUnhealthy, patch, mark healthy, resume.

Health Checks

• EC2 status (default) vs ELB/ALB checks (app-aware).
• ASG and ALB health are independent — point ASG at ELB checks when ALB is source of truth.
• Grace period: Wait for boot before judging unhealthy.

Default Termination Policy

• When balancing an AZ: prefer the instance with the oldest launch configuration (classic exam answer).

Capacity Math

• Need ≥4 instances after losing one AZ → e.g. 3 AZs × 2 (not 2 AZs × 2).