A recent research study delved into the cost advantage offered by AWS Spot Instances in terms of an organization's overall cloud spending. By analyzing the preemption rates of t3/t4 Spot Instances within the research group, the data revealed a notable surge in demand for spot instances as a whole.
Eric Pauley, NSF Graduate Fellow at the University of Wisconsin-Madison, put forth the research in a blog post. Pauley said that as businesses seek to reduce expenses amid shifting macroeconomic conditions, cloud expenditures have become an important aspect of cost-cutting measures.
The research group has overseen the deployment of t3/t4 spot instances on AWS. Over the period spanning from October 2022 to May 2023, the group provisioned 5.5 million spot instances across various regions. These instances were utilized for in-depth research on the allocation of IP addresses by cloud providers and their impact on tenant security. Since the start of 2023, there has been a significant increase in spot ratios, particularly in the us-east-1 region, where it reached as high as 55%. Additionally, prices have experienced an upward trend in four of the largest AWS regions.
Continuing the research, the investigation delved into the aspect of instance preemption. To analyze the behavior, each server was operational for a duration of 10 minutes before being shut down, and the instance preemption was carefully monitored. When plotting the data for preemption rates of spot instances in regions with price spikes versus regions without, the following observations were made:
Source: Farewell to the Era of Cheap EC2 Spot Instances
Spot instance demand experienced a substantial increase, leading to a fourfold rise in preemptions over a few months. Instance families such as m3, a1, t3a, and g5g experienced significant price increases.
The tech community took notice of this research on Twitter, Reddit, and YCombinator. We saw an interesting comment by Ricardo Aravena on Twitter, "Looks like the law of diminishing returns is catching up when it comes to spot instances". A Reddit conversation had one of the Redditors workmakesmegrumpy commenting on the analysis, "Isn't this just a case of 'If more people are using spot at the same time, there's less spot instances available'?"
According to Pauley, we have entered a race to the bottom in terms of spot discounts. If your system is designed to accommodate instance preemptions and the cost of provisioning replacement instances is manageable, using spot instances may result in a better outcome than using on-demand instances. Considering the diminishing benefits, organizations should reevaluate their architecture decisions and prioritize approaches like Savings Plans for more guaranteed savings, especially when usage is predictable.
Community comments
Great article and spot on!
by Nico Coetzee,
Great article and spot on!
by Nico Coetzee,
Your message is awaiting moderation. Thank you for participating in the discussion.
We operate several EKS clusters mainly in "eu-central-1" and found more or less the same. Especially "az1" is almost always out of t-series spot capacity. But even m-series instances can be problematic. It was ok, since we never had issues in the remaining availability zones. However, since about two weeks ago another AZ have also become problematic for us.
It looks like there are a lot of people running spot instances only in ONE availability zone. And it seems these same people have now also realized the problems in one AZ and they are now starting to target another AZ, so the problem is starting to just move around.
Very frustrating, especially when AWS have bragged for a long time how they have virtually endless capacity. Turns out, they too have limits, and it's starting to affect everyone.
Just want to end off by saying for EKS we are still seeing huge savings, regardless of these events. So, at least for now, there is still great value in spot instances. But yes, I am concerned for how long this will last.