Commonly, it is impossible to mimic a public cloud solution because it relies on services unique to the service providers, such as GCP (Borg, Jupiter, Colossus). Similar services exist in AWS and Azure but with less impact. Those service providers lay down critical components, such as networking, storage subsystems, and file systems, which can’t be replicated. This happens when a “custom build” solution becomes a service without having an intervening production phase.
Kubeflow, one of the most critical components of our platform, is an excellent example of such a service. Despite working well in a GCP environment, we faced numerous issues adopting it to typical data center needs. Some examples of products that can be successfully adopted in a private data center are the Hadoop ecosystem (Apache HBase®, Hadoop Distributed File System [HDFS]), and Kubernetes.
Easily integrating a machine learning platform into a data center ecosystem has many benefits. It naturally fills the gap between the data processing platform and the serving environment, and it can be the only solution the client would ever need.