Keywords: machine learning, artificial intelligence, open source, data science, Kubeflow, TensorFlow
Launching the first proof-of-concept version of a machine learning system is fairly straightforward, but when you productionize, all sorts of issues can arise. What was built as the prototype is only a small piece of what needs attention, and problems often show up when you try to scale out and keep a system in long-term continuous operation. Data cleaning and processing is hard at scale; training/serving skew is often a problem, as is model drift; there are new iteration, tracking/monitoring, and reproducibility requirements; and more. This talk will describe some "anti-patterns" that can make it hard to productionize, as well as some open-source libraries and frameworks that can help. In particular, we'll introduce Kubeflow (https://www.kubeflow.org/), an open-source project to build a machine learning stack on Kubernetes, and Kubeflow Pipelines.