FlowSpecand implementing steps as methods. Besides steps, a flow can define other attributes relevant for scheduling, such as parameters and data triggers.
@stepdecorator in a flow class.
self.x, used in the step code become data artifacts that are persisted automatically. Stack variables, e.g.
x, are not persisted. This dichotomy allows the user to control the overhead of checkpointing by explicitly choosing between persistent vs. non-persistent variables in the step code.
python myflow.py runon the command line.
metaflow.client. A typical way to use
metaflow.clientis to access data artifacts of past runs in a Jupyter notebook. It is extremely convenient to be able to examine the internal state of production runs or perform further ad-hoc analysis of the results in a notebook.