Talk in English - US at Incontro DevOps Italia 2024
Track Name:
Sala 1
View Slides: https://univr-my.sharepoint.com/:b:/g/personal/francesco_lumpp_univr_it/EedPtP9eh65Ng_-YOa_NovsBr7F1umSt4U3ny8B7RMz7BQ?e=fclHmO
Short URL: https://joind.in/talk/4fe64
(QR-Code (opens in new window))
Explore the convergence of cost management and development with FinOps in the context of Kubernetes. This talk presents an integrated solution to effortlessly collect costs from service providers' APIs, convert them to the FinOps Foundation FOCUS specification, and store them in a central data lake. Leveraging advanced algorithms, including machine learning, the collected data is analyzed to uncover opportunities for cost and quality of service optimization. This solution includes a closed-loop feedback system where insights gleaned from the analysis are fed back into Kubernetes to enhance integrated auto-scaling features. This talk will unravel the complexities of FinOps within Kubernetes, gaining practical insights into transforming cloud cost management and bolstering DevOps efficiency.
Comments
Comments are closed.
Probably too many tech details. Some slides have too much text
Very interesting, I appreciated the introduction of LLMs in the analisys context, even if it has currently restrictions
The idea to take the FOCUS specification as baseline is indeed good.
The talk's topic is original, and the exposition clear. However the approach explained seems over engineering the solution to the talks's problem.
First of all, cloud costs managements of a multi-cloud initiative is a super-niche problem: costs control for a cloud vendor is a niche, GCP / Oracle are a niche, so costs control for other than AWS or Azure is a super-niche.
Secondly, the fact a technology is "exploding" it is not a sufficient condition to use it like the salt. LLMs are not appropriate to perform a conversion from one schema to another schema: ontologies for instance could make a better job and costs for development, maintenance are surely lower. To not mention the costs of inference of a 70B neural network used to convert one field in another with an accuracy that in the best case is 90%. And it is not neither discussed the impact of wrong conversion given by the LLM. And the costs for guard railing and monitor the results of LLM.
In my opinion, it should be revisited the approach involving LLMs as totally unnecessary, but the idea of collecting all the measures and the costs in order to take data-driven decisions is really interesting.
Lots of interesting info. Just a remark about the implementation details, which sometimes felt really too much and hard to follow.
FinOps is something we need to take care in Cloud and microservices / orchestration contexts.
It's a very complex topic and the starting point with the FOCUS specification was a good theoric part to know.
The speech is a deep dive into tooling and LLM integrations to manage the complexity of FinOps data gathered from very different sources and consolidation into the FOCUS standardized data format.
perhaps explained at too high a level