A notebook is a dataflow graph

marimo reads which names each cell defines and which it reads. An edge runs from the cell that defines a name to every cell that reads it, and the edges form a directed acyclic graph. When you move a slider, marimo reruns the cells downstream of it, in an order where every edge points forward.

examples/dataflow.py runs that analysis on a small notebook of eleven cells. The scene draws the cells in the order the file lists them, slides them into topological order until every reference points forward, then lights up the cells the chosen slider reruns.

Notebookscene.pyExported page

Open the notebook and its view in Studio from a clone of the repository:

pixi run edit dataflow

The notebook computes the graph

A marimo file lists each cell as a function. Its parameters are the names it reads and its return value names what it defines, so the notebook builds the graph with Python's ast module alone:

for _node in ast.parse(source).body:
    _returned = _node.body[-1].value
    _names = _returned.elts if isinstance(_returned, ast.Tuple) else [_returned]
    cells.append(
        {
            "name": _node.name,
            "reads": [argument.arg for argument in _node.args.args],
            "defines": [name.id for name in _names],
        }
    )

Kahn's algorithm then repeatedly takes the earliest cell whose inputs have all run. A walk from the cells that read the chosen slider finds everything downstream of it. The scene reads four results:

Value	Holds
nodes	Each cell's name, its position in the file, and its topological rank
edges	One [source, target] pair per reference between cells
backward	How many references point back up the file
rerun	The slider's cell and the cells it reruns, in topological order

The scene places the cells, draws an arc for each reference, and moves the cells to their ranks. It sorts nothing and walks no graph itself.

Slider moved	Cells that rerun
noise	6: sample, fit, predict, chart, error, report
points	8: grid, truth, sample, fit, predict, chart, error, report
degree	5: fit, predict, chart, error, report

The page

The view's page places the notebook's controls and summary cells under the movie. A zero-python export prepares one movie per slider, and the Exported page link under the movie opens it.

<marimo-output render="scene.py" aria-label="The notebook's cells, sorted and rerun">...</marimo-output>
<marimo-cell name="controls"></marimo-cell>
<marimo-cell name="summary"></marimo-cell>