Data science applications tend to be built by composing \emph{tasks}:
discrete manipulations of data. These tasks are arranged in directed
acyclic graphs, and many frameworks exist within the data science
community supporting such a structure, which is called a \emph{workflow}.
In realistic applications, we want to be able to both \emph{analyze}
a workflow in the absence of data, and to \emph{execute} the workflow
with data.

This paper combines effect handlers with arrow-like structures to abstract out
data science tasks. This combination of techniques enables a modular design of
workflows. Additionally, these workflows can both be analyzed prior to running (e.g., to
provide early failure) and run conveniently. Our work is directly motivated by
real-world scenarios, and we believe that our approach is applicable to new
data science and machine learning applications and frameworks.