There are many that try to bridge the gap between the iterative, incremental work done in Research and the integrated, routinely automated workflows of modern software development. This manifests itself in the DevOps pipelining tools that ML practitioners try to “jam” for their research needs.
From our experience working with groundbreaking companies, we can surmise that a bottom-up approach is far superior to a top-down one – for building pipelines and other automation. We propose a structured process for such design within an experiment-tracking platform with an orchestration component.
|[shared] Are You Using Airflow_ Dont - MLLifecycleConf Ariel Biller, Updated.pdf|