Do not use Kubeflow!
This post is just my opinion about the Kubeflow solution, and I in addition to reading this post, run tests and create my own vision of this solution. For those who disagree with the points presented below, I wonder if they will publish a reply!
Kubeflow is a solution that they see with the proposal of being an “End-to-end” platform for data scientists; that is, using it the scientist will be able to do all the steps for creation, training, testing, and availability of data models of artificial intelligence.
This solution is gaining more users and contributors every day in the most diverse business and academic environments, in addition to having a very active maintenance community, making Kubeflow one of the most complete solutions, with many integrations and completely open source.
With the hype of this solution, I observed many considering it as a solution to all problems, a true silver bullet, and I made this list with some of the points that may influence its non-use:
Every open-source solution also generates cost; it is not because it can be downloaded and used by everyone that it will not be necessary to maintain an infrastructure, and depending on the projects, it will be necessary to use computer graphics; it is a platform that needs a container environment which in turn needs an infrastructure.
Remember that Kubeflow requires significant upfront investment and ongoing costs, which may only be feasible for some organizations.
Operating Cost 🙎
To use Kubeflow, it is necessary to install it on a Kubernetes infrastructure, and after installation, in addition to the Kubernetes cluster, it will be necessary to maintain and monitor its dependencies:
In many cases, it is necessary to maintain some extra facilities:
- Istio( Loadbalce allows calling models via API)
- KServe(Enables accessing…