Skip to content

“Usage Patterns and the Economics of the Public Cloud”

The paper (PDF) is, to say it in the words of Sascha Konietzko, eine ausgesprochene Verbindung von Schlau und Dumm (“a very special combination of smart and stupid”)

The site is not related to the corporation of the same name, but the site of one of the authors, R. Preston McAfee.

The paper looks at the utilization data from a number of public clouds, and tries to apply some dynamic price finding logic to it. The authors are surprised by the level of stability in the cloud purchase and actual usage, and try to hypothesize why is is the case. They claim that a more dynamic price finding model might help to improve yield and utilization at the same time (but in the conclusion discover why in reality that has not happened).

Being AI people with not much Ops experience, they make a number of weird assumptions. For example, they are looking at CPU usage data in 5 min aggregation buckets and do math against 100% CPU utilization. Even if you assume at all or most workloads are actually CPU bound, 5 minute buckets of CPU usage are way to coarse to use for 100% as a utilization target. That is, because all the actual spikes are shorter and invisible.

(Workloads are rarely CPU bound, and especially not in the cloud, which is most often network bound.) (They do look at 5 min bucket maxima = 100th percentiles in some cases. But even that does not tell you for how long in a 5 min bucket you hit the roof. But hitting the roof even for a short time creates latency outliers, which in an actual cloud native microservice framework propagate and multiply, turning most of the requests into SLO violations. Jitter kills).

They are also for the most part ignore the cost of the cognitive load of dynamic pricing, or the cost of coding for variable deployment sizes in traditional workloads.

Finally, they are looking at a VM based IaaS deployment. VM based deployments are likely to see traditional bare metal workloads being forklifted into the cloud. Such deployments do not scale, because they weren’t built for dynamic scaling. Cloud Native stuff is more likely to be found in environments such as Amazon Lambda and Athena. This is a lot like looking at the childrens pool only and coming to the conclusion that most people can’t swim, i.e. there may be selection bias.

There are useful thoughts in the paper, too. For example, they find that they can model and predict many workloads and use this for scaling in advance.

In fact, typical reactive autoscalers do not work properly, because by the time they trigger a scaling action, demand is already there and latencies lag in violation of the SLO. The systems created as a reaction to a reactive autoscaler triggering are slow, because caches are cold, and will have trouble keeping up, right in the middle of a spike action.

Predictive scalers work better and may be safer. Here you establish an absolute size limit (which usually exists due to architectural constraints and locks on shared state), and size down from it to meet predicted demand with a comfortable margin of error. Then you scale up and down following predicted demand, correcting the model in time to keep the margin. This should create capacity in advance, and give it time to warm up.

Also, reading the paper you could come to the conclusion that the market is stable and static, because it is stable and static: Because nobody rocks the demand pressure up and down, the prices are stable and nobody has a reason to implement saving by changing the size of the deployment needlessly. Of course, once a sufficient number of people start doing this, demand pressure will vary sufficiently to affect short term pricing, so that everybody eventually needs to react to the changed environmental conditions. Noise breeding complexity, breeding even more noise and complexity, until everything drowns in chaos and the system needs downtime to clean the noise from the system.

TL;DR: Bunch of objectivist hipsters with belief in market forces and no Ops experience discover in the final paragraphs of their paper that stability and simplicity are tangible, priceable assets in themselves and maybe all the dynamic market shit is not really helpful in all cases, because stability savings outweigh wins from leveraging dynamic market forces.

Published inComputer ScienceContainers and KubernetesPerformance

Be First to Comment

Leave a Reply

Your email address will not be published. Required fields are marked *