A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.
Python artificial-intelligencedagdata-sciencedata-visualizationdataflowdeveloper-toolsmachine-learningnotebookspipelinepythonreactivesqlweb-app
---
title: Dataflow
marimo-version: 0.0.0
header: |-
# Copyright 2026 Marimo. All rights reserved.
# /// script
# requires-python = ">=3.12"
# dependencies = [
# "marimo",
# "matplotlib==3.10.9",
# "numpy==2.2.4",
# ]
# ///
---
# How marimo notebooks run
Reactive execution is based on a single rule: when a cell is run, all other
cells that reference any of the global variables it defines run
automatically.
To provide reactive execution, marimo creates a dataflow graph out of your
cells.
<!---->
**Tip: disabling automatic execution.**
marimo lets you disable automatic execution: in the notebook
footer, change "On Cell Change" to "lazy".
When the runtime is lazy, after running a cell, marimo marks its
descendants as stale instead of automatically running them. The lazy
runtime puts you in control over when cells are run, while still giving
guarantees about the notebook state.
<!---->
## References and definitions
A marimo notebook is a directed acyclic graph in which nodes represent
cells and edges represent data dependencies. marimo creates this graph by
analyzing each cell (without running it) to determine its
- references ("refs"), the global variables it reads but doesn't define;
- definitions ("defs"), the global variables it defines.
There is an edge from one cell to another if the latter cell references any
global variables defined by the former cell.
The rule for reactive execution can be restated in terms of the graph: when
a cell is run, its descendants are run automatically.
<!---->
### Example
The next four cells plot a sine wave with a given period and amplitude.
Each cell is labeled with its refs and defs.
```python {.marimo hide_code="true"}
mo.accordion(
{
"Tip: inspecting refs and defs": f"""
Use `mo.refs()` and `mo.defs()` to inspect the refs and defs of any
given cell. This can help with debugging complex notebooks.
For example, here are the refs and defs of this cell:
{mo.as_html({"refs": mo.refs(), "defs": mo.defs()})}
"""
}
)
```
```python {.marimo}
mo.md(f"""
{mo.as_html(plot_wave(amplitude, period))}
- `refs: {mo.refs()}`
- `defs: {mo.defs()}`
""")
```
```python {.marimo}
period = 2 * 3.14159
mo.md(
f"""
- `refs: {mo.refs()}`
- `defs: {mo.defs()}`
"""
)
```
```python {.marimo}
amplitude = 1
mo.md(
f"""
- `refs: {mo.refs()}`
- `defs: {mo.defs()}`
"""
)
```
```python {.marimo}
def plot_wave(amplitude, period):
x = np.linspace(0, 2 * np.pi, 256)
plt.plot(x, amplitude * np.sin(2 * np.pi / period * x))
plt.xlim(0, 2 * np.pi)
plt.ylim(-2, 2)
plt.xticks(
[0, np.pi / 2, np.pi, 3 * np.pi / 2, 2 * np.pi],
[0, r"$\pi/2$", r"$\pi$", r"$3\pi/2$", r"$2\pi$"],
)
plt.yticks([-2, -1, 0, 1, 2])
plt.gcf().set_size_inches(6.5, 2.4)
return plt.gca()
mo.md(
f"""
- `refs: {mo.refs()}`
- `defs: {mo.defs()}`
"""
)
```
🌊 **Try it!** In the above cells, try changing the value `period` or
`amplitude`, then click the run button ( ▷ ) to register your changes.
See what happens to the sine wave.
<!---->
Here is the dataflow graph for the cells that make the sine wave plot, plus
the cells that import libraries. Each cell is labeled with its defs.
```
+------+ +-----------+
+-----------| {mo} |-----------+ | {np, plt} |
| +---+--+ | +----+------+
| | | |
| | | |
v v v v
+----------+ +-------------+ +--+----------+
| {period} | | {amplitude} | | {plot_wave} |
+---+------+ +-----+-------+ +------+------+
| | |
| v |
| +----+ |
+------------> | {} | <-------------+
+----+
```
The last cell, which doesn't define anything, produces the plot.
<!---->
## Dataflow programming
marimo's runtime rule has some important consequences that may seem
surprising if you are not used to dataflow programming. We list these
below.
<!---->
### Execution order is not cell order
The order in which cells are executed is determined entirely by the
dataflow graph. This makes marimo notebooks more reproducible than
traditional notebooks. It also lets you place boilerplate, like
imports or long markdown strings, at the bottom of the editor.
<!---->
### Global variable names must be unique
Every global variable can be defined by only one cell. Without this
constraint, there would be no way for marimo to know which order to
execute cells in.
If you violate this constraint, marimo provides a helpful
error message, like below:
```python {.marimo}
planet = "Mars"
planet
```
```python {.marimo}
planet = "Earth"
planet
```
**🌊 Try it!** In the previous cell, change the name `planet` to `home`,
then run the cell.
<!---->
Because defs must be unique, global variables cannot be modified with
operators like `+=` or `-=` in cells other than the one that created
them; these operators count as redefinitions of a name.
**🌊 Try it!** Get rid of the following errors by merging the next two
cells into a single cell.
```python {.marimo}
count = 0
```
```python {.marimo}
count += 1
```
### Underscore-prefixed variables are local to cells
Global variables prefixed with an underscore are "private" to the cells
that define them. This means that multiple cells can define the same
underscore-prefixed name, and one cell's private variables won't be
made available to other cells.
**Example**.
```python {.marimo}
_private_variable, _ = 1, 2
_private_variable, _
```
```python {.marimo}
_private_variable, _ = 3, 4
_private_variable, _
```
```python {.marimo}
# `_private_variable` and `_` are not defined in this cell
_private_variable, _
```
### Deleting a cell deletes its variables
Deleting a cell deletes its global variables and
then runs all cells that reference them. This prevents severe bugs that
can arise when state has been deleted from the editor but not from the
program memory.
```python {.marimo}
to_be_deleted = "variable still exists"
mo.md(
"""
🌊 **Try it!**
Delete this cell by clicking the trash bin icon.
"""
)
```
```python {.marimo}
to_be_deleted
```
### Cycles are not allowed
Cycles among cells are not allowed. For example:
```python {.marimo}
one = two - 1
```
```python {.marimo}
two = one + 1
```
### marimo doesn't track attributes
marimo only tracks global variables. Writing object attributes does not
trigger reactive execution.
**🌊 Example**. Change the value of `state.number` in the next cell, then
run the cell. Notice how the subsequent cell isn't updated.
```python {.marimo}
state.number = 1
```
```python {.marimo}
state.number
```
```python {.marimo}
class namespace:
pass
state = namespace()
state.number = 0
```
```python {.marimo hide_code="true"}
mo.accordion(
{
"Why not track attributes?": """
marimo can't reliably trace attributes
to cells that define them. For example, attributes are routinely
created or modified by library code.
"""
}
)
```
### marimo doesn't track mutations
In Python, it's impossible to know whether code will
mutate an object without running it. So: mutations (such as
appending to a list) will not trigger reactive execution.
```python {.marimo hide_code="true"}
mo.accordion(
{
"Tip (advanced): mutable state": (
"""
You can use the fact that marimo does not track attributes or
mutations to implement mutable state in marimo. An example of
this is shown in the `ui` tutorial.
"""
)
}
)
```
## Best practices
The constraints marimo puts on your notebooks are all natural consequences
of the fact that marimo programs are directed acyclic graphs. As long as
you keep this fact in mind, you'll quickly adapt to the marimo way of
writing notebooks.
Ultimately, these constraints will enable you to create powerful notebooks
and apps, and they'll encourage you to write clean, reproducible code.
Follow these tips to stay on the marimo way:
```python {.marimo}
mo.accordion(tips)
```
## What's next?
Check out the tutorial on interactivity for a tour of UI elements:
```
marimo tutorial ui
```
```python {.marimo}
import matplotlib.pyplot as plt
import numpy as np
```
````python {.marimo hide_code="true"}
tips = {
"Use global variables sparingly": (
"""
Keep the number of global variables in your program small to avoid
name collisions across cells. Keep the number of global variables
defined by any one cell small to make sure that the units of
reactive execution are small.
"""
),
"Use descriptive names": (
"""
Use descriptive variable names, especially for global variables.
This will help you minimize name clashes, and will also result in
better code.
"""
),
"Use functions": (
"""
Encapsulate logic into functions to avoid polluting the global
namespace with temporary or intermediate variables.
"""
),
"Minimize mutations": (
"""
We saw earlier that marimo cannot track object mutations. So try
to only mutate an object in the cell that creates it, or create
new objects instead of mutating existing ones.
For example, don't do this:
```python3
# a cell
numbers = [1, 2, 3]
```
```python3
# another cell
numbers.append(4)
```
Instead, prefer
```python3
# a cell
numbers = [1, 2, 3]
numbers.append(4)
```
or
```python3
# a cell
numbers = [1, 2, 3]
```
```python3
# another cell
more_numbers = numbers + [4]
```
"""
),
"Write idempotent cells": (
"""
Write cells whose outputs and behavior are the same when given
the same inputs (refs); such cells are called _idempotent_. This will
help you avoid bugs, and let you cache expensive intermediate
computations (see the next tip).
"""
),
"Cache intermediate computations with `@mo.cache`": (
"""
Use `mo.cache` to cache the return value of expensive functions.
You can do this if you abstract complex logic into idempotent
functions, following earlier tips.
For example:
```python3
import marimo as mo
@mo.cache
def compute_prediction(problem_parameters):
...
```
Whenever `compute_predictions` is called with a value of
`problem_parameters` it has not seen, it will compute the predictions
and store them in a cache. The next time it is called with the same
parameters, instead of recomputing the predictions, it will just
fetch the previously computed ones from the cache.
If you are familiar with `functools.cache`, `mo.cache` is
similar but more robust, with the cache persisting even
if the cell defining the function is re-run.
"""
),
}
````
```python {.marimo}
import marimo as mo
```