A word on scale—there is a tendency to reach out to packages or external programs, such as Spark, to solve the problem. Often they do solve the problem. But it's been my experience that ultimately, when doing things at scale, there is no one-size-fits-all solution. Therefore, it's good to learn the basics, so that when necessary, you may refer to the basics and extrapolate them to your situation.

Again on the topic of scale—both researchers and practitioners would do well to learn to plan projects. This is one thing that I am exceedingly bad at. Even with the help of multiple project managers, machine learning projects have a tendency to spiral out of control. It does take quite a bit of discipline to manage these. This is both on the implementor's part and on the stakeholder's part.

Last, learn to manage the expectations of stakeholders. Many of my projects fail. That I can say the projects fail is itself a qualifying statement. For most projects I enter into, I have defined success and failure criteria. If it's a more traditional statistics-based project, then these are your simple null hypotheses. Failing to reject the null hypothesis would then be a failure. Likewise, more complicated projects would have multiple hypotheses—these come in form of F-scores and the like. Learn these tools well, and communicate them to your stakeholders. You must be aware that a large majority of machine learning projects fail on their first few attempts.