Spot instance recommendations

[maksimov]

Add spot instance pricing recommendations for suitable workloads.

Currently the only pricing recommendation is on-demand → Reserved Instance.
Spot instances can save 60-90% over on-demand for interruptible workloads.

Core

• Identify workloads suitable for spot (training jobs, batch inference, dev instances)
• Exclude persistent endpoints and production inference from spot recommendations
• Factor in spot pricing when calculating potential savings
• Adjust idle signal severity for spot instances (idle spot is much cheaper than idle on-demand)

Spot Fleet diversification

• Suggest mixed-instance Spot Fleets (e.g. G4ad + G5 + G6) to minimize total outage risk
• Correlate GPU-specific interruption rates — g5.xlarge (A10G) tends to have higher stability than p4d (A100) due to larger capacity pool

Interruption data

• AWS EC2 Capacity Manager (launched 2026) provides native Spot interruption metrics
• Use this to give concrete interruption risk percentages per instance type per region
• Example recommendation: "Switch from g5.12xlarge On-Demand to g5.2xlarge Spot fleet — saves 72% with <5% interruption risk in us-east-1"