After a year of running KubeFlow across multiple client deployments, I can confidently say: the platform that promises end-to-end ML on Kubernetes delivers end-to-end frustration instead. Here's the post-mortem.
The Infrastructure as Code wars are over. Terraform won. And yet enterprise teams are still clinging to vendor-native tooling like ARM templates and CloudFormation because 'it's supported by the vendor.' Supported doesn't mean good.
Vendor consultants optimise for showcasing their platform's capabilities, not for your team's ability to maintain it. They build reference architectures, not products. The demo passes. Production doesn't.
Not everyone can 'just use the cloud.' Healthcare data sovereignty, air-gapped networks, public sector paranoia — sometimes you're deploying ML inference on a rack server in a basement. Here's what that actually looks like with MicroK8s.
Every organisation wanted an MLOps platform circa 2020. Most had data scientists running Jupyter notebooks on their laptops and calling it a pipeline. Building an airport for one flight a week.
Nobody is running workloads portably across clouds. They're running different things on different clouds and calling it a strategy. Multi-cloud is what happens when nobody's in charge.