flowview

Visual data pipeline debugger for Polars. Stop print-debugging your pipelines.

Install

pip install flowview

or with uv:

uv add flowview

Quick Start

Add @fv.trace to any function that transforms a Polars DataFrame. flowview traces every method call and renders a visual flow in your terminal.

import polars as pl
import flowview as fv

@fv.trace
def process(df: pl.DataFrame) -> pl.DataFrame:
    return (
        df.filter(pl.col("status") == "active")
          .with_columns((pl.col("price") * pl.col("quantity")).alias("revenue"))
          .group_by("category")
          .agg(pl.col("revenue").sum().alias("total_revenue"))
          .sort("total_revenue", descending=True)
    )

df = pl.DataFrame({
    "status": ["active", "inactive", "active"],
    "category": ["Books", "Books", "Electronics"],
    "price": [14.99, 599.99, 299.99],
    "quantity": [5, 1, 2],
})

result = process(df)

.pipe() chains work too:

@fv.trace
def process(df: pl.DataFrame) -> pl.DataFrame:
    return df.pipe(clean).pipe(filter_active).pipe(add_revenue)

What You See

Each step in your pipeline is displayed as a box showing:

• Row count with diff from the previous step (e.g., 700 rows x 4 cols (-300 rows))
• Schema changes — columns added or removed (e.g., +cols: revenue -cols: status)
• Sample data — first N rows at each transformation
• Execution time per step

Steps are connected with arrows to show the flow. A summary footer shows the total step count and wall-clock time.

Supported Operations

flowview traces any DataFrame method that returns a new DataFrame. These methods get human-readable step names:

Other methods (e.g., explode, melt, unpivot) are traced with a fallback name like explode('tags').

Options

@fv.trace(sample_rows=3, show_sample=True, show_schema=True)
def process(df):
    ...

How It Works

The @fv.trace decorator wraps the first DataFrame argument in a lightweight proxy before calling your function. The proxy intercepts every method call, captures a snapshot of the result (row count, schema, sample rows, timing), and delegates to the real Polars DataFrame underneath. When your function returns, the proxy is unwrapped and you get back a regular pl.DataFrame.

There is no monkey-patching and no global state. Each decorated call is fully isolated.

Limitations

• LazyFrame is not supported — df.lazy() exits the proxy. Only eager DataFrames are traced.
• GroupBy shortcuts like .count(), .sum(), .first() on a GroupBy object are not traced — use .agg() instead.
• Pipe internals are not individually traced — df.pipe(fn) produces a single step named after fn, not one step per operation inside fn.
• IDE autocomplete may not show DataFrame methods inside the decorated function body.
• type(df) returns TracedDataFrame inside the decorated function. isinstance(df, pl.DataFrame) works correctly.
• Only the first DataFrame argument is wrapped when a function takes multiple DataFrames.

License

MIT