Spot Interruption Rates by Instance Family: What the Data Actually Shows

When engineering teams evaluate AWS Spot Instances, the primary concern is always interruption rate. The fear of sudden, cascading capacity loss often prevents organizations from utilizing spot pricing, opting instead for the expensive safety of on-demand instances. However, spot interruption behavior is not a monolith. Analyzing six months of spot interruption data across over forty instance families reveals that interruption patterns vary wildly depending on the specific hardware, the Availability Zone, and broader market demand.

The data demonstrates a sharp dichotomy between specialized hardware and general-purpose compute. For GPU instances—specifically the p3, p4d, and g5 families—demand heavily outstrips supply across nearly all regions. Our analysis indicates these instance types experience aggressive interruption rates hovering between 15% and 25%. If your machine learning training pipelines are not robustly configured for frequent checkpointing and rapid resumption, running on spot GPUs will result in severe operational friction.

Conversely, the story is entirely different for standard, general-purpose instances. The m5 and m6i families, which represent the bulk of standard web and API workloads, demonstrate remarkably stable behavior. Across most major regions, these families experience interruption rates of just 3% to 8%. For stateless microservices running in Kubernetes, an 8% interruption rate is easily mitigated by standard replica configurations and basic pod anti-affinity rules.

The strategy for successfully leveraging spot pricing relies entirely on diversification. A single-minded reliance on one instance type in one AZ is a fragile configuration. Instead, teams must implement a capacity-optimized allocation strategy. By configuring Auto Scaling Groups or node provisioners like Karpenter to select from a broad pool of compatible instance families (e.g., mixing c5, c6i, m5, and m6i), the risk of sudden capacity depletion is mathematically distributed.

Furthermore, monitoring is critical. Relying on anecdotal experience is insufficient for production workloads. Teams should actively track their specific interruption rates using CloudTrail events (BidEvictedEvent) and leverage Spot placement scores to identify AZs with deep capacity pools.

(Note: The data figures cited in this post are approximate estimates based on aggregate observations and will vary significantly based on region and time of day. Always consult current AWS Spot Instance Advisor metrics for baseline planning.)