EC2 Fleet

TL;DR: EC2 Fleet (2018) launches a fleet across multiple purchasing options in one request — Spot, On-Demand, and Reserved together — where Spot Fleet is Spot-only. That lets you say "80% Spot for cost, 20% On-Demand as a safety net," getting massive Spot savings plus a guaranteed baseline that doesn't collapse when Spot tightens. It's a capacity-fulfillment tool (best price for the compute you need), while ASGs are scaling-policy tools — and modern ASGs have absorbed most of Fleet's mixing features, so for web apps an ASG usually wins.

The numbers

• Purchasing options blend: Spot (up to 90% off), On-Demand (full price, always available), Reserved (pre-purchased) — typically 50–80% savings with a small On-Demand baseline.
• Allocation strategies (the key lever): capacityOptimized / capacityOptimizedPrioritized use AWS interruption data to pick the safest Spot pools — the production go-to; lowestPrice concentrates risk; diversified spreads evenly.
• Instance weighting lets a target of 100 units mix sizes (m5.large=2, m5.xlarge=4…) so Fleet packs cheaper/more-available pools.
• Fleet types: maintain (replaces interruptions — production default), request (one-time), instant (synchronous one-shot).
• Field examples: video encoding on a 10-instance On-Demand base + Spot (c5/c5n/c5a/c6i, capacity-optimized, 30s checkpoints) ran ~70% cheaper than pure On-Demand; a hedge fund's GPU Monte Carlo (2 p3.8xlarge On-Demand + up to 20 Spot) cut 75%.

Do this

1. Set an On-Demand base for critical workloads (base + Spot percentage), so you always have core capacity even if Spot dries up entirely.
2. Use capacityOptimized or capacityOptimizedPrioritized for Spot — dramatically lower interruption rates than lowestPrice at similar cost.
3. Diversify aggressively — 10+ current-gen instance types across multiple AZs; teams that pinned to 3 types hit interruption problems.
4. Keep the config clean with launch templates + overrides — common settings in the template, instance-type/AZ variation in overrides.
5. Monitor interruption rates per instance type in CloudWatch and prune the bad actors; design the app to handle the 2-minute warning gracefully.

Gotchas

• Over-complicated configs (50 types + elaborate weighting) are a common trap — start with 5–10 types and iterate.
• Launch-template updates don't replace running instances — you must recycle instances or create a new fleet.
• Spot isn't guaranteed even diversified — always keep an On-Demand fallback for critical capacity.
• No built-in load balancer integration — Fleet doesn't auto-register with ALB/NLB the way an ASG does.

Skip this if

• It's a web app or service needing metric-based scaling and ALB/NLB integration — use an Auto Scaling Group with a mixed-instances policy instead (better CloudWatch/LB integration, same Spot economics).
• It's fault-tolerant batch wanting maximum savings with no reliability floor — Spot Fleet is simpler; for the underlying model see Spot Instances. Cover the always-on baseline with Reserved Instances.