polars-gw

Native Polars computation engine for Graphic Walker.

Translates Graphic Walker IDataQueryPayload workflow steps directly into Polars operations — no DuckDB, no SQL intermediate.

Installation

pip install polars-gw              # core translator only
pip install 'polars-gw[viz]'       # core + built-in walk() UI

The core install only pulls in polars. The [viz] extra adds FastAPI, uvicorn and pydantic so you can launch a local Graphic Walker against a DataFrame in one line.

Interactive UI

import polars as pl
from polars_gw import walk

df = pl.read_parquet("sales.parquet")
handle = walk(df)          # opens http://127.0.0.1:<free-port> in your browser
# ...
handle.stop()

walk() starts a FastAPI server in a background daemon thread, serves the Graphic Walker UI bundled inside the wheel (no CDN, no network required), and wires its computation callback to execute_workflow.

Export / Import chart specs

Charts you build in the UI can be saved and restored across sessions:

# Export the current charts to a JSON file
handle.export("my_charts.json")

# Or just grab the spec as a Python list
spec = handle.export()

Next time, pass the file back to walk() to restore exactly where you left off:

handle = walk(df, spec_file="my_charts.json")

The spec file is a plain JSON array of Graphic Walker IChart objects — safe to version-control or share with collaborators.

Logs

By default walk() prints one line per compute call so you can see what GW is asking the backend for:

13:42:11  INFO    polars_gw.viz: Graphic Walker running on http://127.0.0.1:54221 — 12 000 000 rows x 19 cols, max_rows=1000000
13:42:18  INFO    polars_gw.viz: compute: 2 step(s) → 847 row(s) in 142.3 ms
13:42:24  INFO    polars_gw.viz: compute: 1 step(s) → 1000000 row(s) in 1284.7 ms [CAPPED]

[CAPPED] means the result hit max_rows and was truncated — Graphic Walker shows its "Data Limit Reached" toast in the UI when this happens. Default cap is 1 000 000 rows; tune it with max_rows= or disable with max_rows=None.

Quiet things down with log_level="warning", or get more detail with log_level="debug":

walk(df, log_level="warning")   # only warnings and errors
walk(df, max_rows=None)         # no row cap (use with care on big data)

If you've already configured Python logging yourself (logging.basicConfig, a framework, pytest's caplog, etc.), walk() won't overwrite it — it only attaches a console handler when nothing else owns logging.

Usage

import polars as pl
from polars_gw import execute_workflow, get_fields

df = pl.read_csv("data.csv")

# Get field definitions for Graphic Walker
fields = get_fields(df)
# [{"fid": "city", "name": "city", "semanticType": "nominal", "analyticType": "dimension"}, ...]

# Execute a Graphic Walker computation payload
payload = {
    "workflow": [
        {"type": "filter", "filters": [
            {"fid": "age", "rule": {"type": "range", "value": [18, 65]}}
        ]},
        {"type": "view", "query": [
            {"op": "aggregate", "groupBy": ["city"], "measures": [
                {"field": "salary", "agg": "mean", "asFieldKey": "avg_salary"}
            ]}
        ]},
        {"type": "sort", "sort": "descending", "by": ["avg_salary"]}
    ],
    "limit": 100
}
results = execute_workflow(df, payload)
# Returns list[dict] — ready to return as IRow[] to Graphic Walker

Supported Operations

Workflow Steps

Field Inference

get_fields() maps Polars dtypes to Graphic Walker field types:

Integer classification (classify_integers)

An integer column can be a dimension (a discrete key you group by) or a measure (a quantity you aggregate) depending on how many distinct values it holds. That count is cheap on a DataFrame but on a LazyFrame means running the plan — so the cost is yours to choose:

fields = get_fields(lf, classify_integers="scan")     # accurate, one pass
fields = get_fields(lf, classify_integers="measure")  # zero data access

Floats and Decimal are always measures; semanticType stays quantitative for every integer regardless of role.

Field overrides

Force any per-column key with field_overrides (typed as IMutFieldOverride):

fields = get_fields(df, field_overrides={
    "user_id": {"analyticType": "dimension"},   # an ID, not a quantity
    "rating":  {"semanticType": "ordinal"},
    "revenue": {"aggName": "mean"},
})

An override that pins analyticType also skips the distinct-count pass for that column, so overriding is free.

How It Differs from PyGWalker

Field-inference parity

The default classify_integers="sample" mirrors PyGWalker's field inference: it classifies from the first 1000 rows (df[:1000]) and treats an integer with ≤16 distinct values as a dimension, otherwise a measure. The distinct count uses Polars' approx_n_unique rather than an exact count, so classification matches PyGWalker in practice but is not guaranteed bit-for-bit at the ≤16 boundary. Deliberate differences: "scan" and "measure" are polars-gw additions PyGWalker has no equivalent for, and polars-gw does not force geo-named columns (lat/long/latitude/longitude) to dimensions the way PyGWalker does.

Compute path

PyGWalker (and panel-graphic-walker) always route through DuckDB — even for Polars DataFrames:

DataFrame → DuckDB → SQL → execute → dicts

polars-gw translates directly to Polars operations:

DataFrame → Polars expressions → execute → dicts

No DuckDB dependency. No SQL intermediate. Just Polars.

polars-gw vs PyGWalker — which should you use?

Different tools, different jobs. polars-gw is a focused compute engine

• a standalone browser UI. PyGWalker is a broader product with richer UX surface area.

Use polars-gw when

• You're already in the Polars ecosystem (e.g. LazyFrames, streaming, or tools like Flowfile) and don't want a DuckDB round-trip in the middle.
• You need type fidelity for Categorical, Decimal, Duration, List, Struct — they pass through as-is instead of degrading through SQL.
• You want a lean install (~40 MB core, no DuckDB).
• You're building your own frontend and just need the translator (execute_workflow + get_fields).
• You want debuggable query plans (Polars expressions, not generated SQL).
• You need a standalone browser UI (walk(df)) without notebook dependencies.

Use PyGWalker when

• You want the Jupyter inline widget — PyGWalker renders the UI inside a notebook cell via anywidget. polars_gw.walk() pops a browser tab.
• You need first-party framework integrations (Streamlit, Gradio, Dash). PyGWalker has these; polars-gw does not.
• You rely on chart persistence — saving/loading chart specs, exporting HTML/PNG, vis_spec round-tripping.
• You want heterogeneous input support (pandas + Polars + parquet + SQL tables under one DuckDB layer).
• Your data is too large to collect into memory and you need DuckDB's battle-tested out-of-core execution.
• You're using Kanaries cloud features (sharing, cloud chat, etc.).

At a glance

Short version: if you're all-in on Polars and want the fast, native path, use polars-gw. If you want inline-notebook, Streamlit, or chart persistence out of the box, use PyGWalker.

Development

Python

uv sync --extra viz           # runtime + viz + dev deps
uv run pytest                 # 60+ tests
uv run ruff check .

Bundling the viz assets

The walk() UI ships a pre-built JS/CSS bundle under polars_gw/viz_assets/ (committed to the repo) so end users don't need Node to pip install polars-gw[viz].

Maintainers rebuild when bumping Graphic Walker:

cd js
npm install
npm run build                 # one-shot production build

Or iterate with watch mode (rebuilds JS + CSS on save, source maps on):

npm run dev                   # in one shell
uv run python example/walk_demo.py   # in another — refresh browser to pick up changes

Bundle layout:

• graphic-walker.js — Graphic Walker + React 19.2.0, minified IIFE (~4.4 MB)
• graphic-walker.css — Tailwind-compiled stylesheet (~57 KB)
• versions.json — pinned npm versions + build mode + timestamp

See js/README.md for details (including why React is pinned to exactly 19.2.0).

License

MIT