Croquis: plot graphs 100x faster on Jupyter Notebook

Croquis is a lightweight Python library for drawing interactive graphs really fast on Jupyter Notebook. It lets you effortlessly browse and examine much larger data than other similar libraries.

Install croquis by running:

pip install croquis

👉 Please look at the tutorial for more example.

You might be wondering: there are already many mature and feature-rich plotting libraries — what's the appeal of a new, experimental library? Well, croquis is built with the following goals in mind:

• Fast: Croquis contains a multi-threaded C++ plotting engine which automatically parallelizes graph generation, easily handling gigabytes of data.
• Simple: Croquis provides a unified, simple API, regardless of data size. Whether your data contains ten data points or fifty million, you can make the same three function calls (or as many calls as you wish in a loop, if you prefer). The library will handle the rest.
• Interactive: Since the C++ engine is alive as long as your figure is visible, at any moment you can zoom in, move around, or select a different subset of your data, and the figure will update accordingly.

As an example, here's hourly ground temperature data of 2020 from the world's weather stations, downloaded from NOAA website. The data set contains 127 million points.

The word "croquis" means a quick, sketchy drawing - it's from French croquis which simply means "sketch." (The final -s is silent: it's French, after all.)

Requirements

• 64-bit Linux/Mac OS running on x86 with AVX2 instruction set support. (Intel: Haswell (2013) or later; AMD: Excavator (2015) or later.)
  • Windows support is under work.
  • Sorry, other architectures aren't supported yet.
• Python 3.6 or later.
• Jupyter Notebook.
• A modern browser (if it can run Jupyter Notebook, it's probably fine).

How to install

pip install croquis

For building from the source, see DEVELOPMENT.md.

To test if it's working correctly, try this inside Jupyter Notebook:

# Paste into a Jupyter cell.

import croquis
import numpy as np

N = 1000000
X = np.random.normal(size=(N, 1))
Y = np.random.normal(size=(N, 1))
labels=['pt %d' % i for i in range(N)]

fig = croquis.plot()
fig.add(X, Y, marker_size=3, labels=labels)
fig.show()

It should generate a plot like this:

For documentation, see the tutorial and the (very short) reference.

By the way, this library is of course open source (MIT License) and totally free to use, but just in case you really liked it for some reason, the author could use a cup of coffee or two... :)

Limitations

Croquis is still experimental: as of version 0.1, we only support the absolute bare minimum functionality. In particular:

• Only line plots are supported, nothing else: no bars, pie charts, heatmaps, etc.
• All lines are solid: no dotted/dashed lines.
• All markers are solid circles: no other shapes are currently supported.
• No subplots: each Jupyter cell can contain only one graph.
• Very few options to customize the plot. No titles, axis labels, or secondary axes.
• No support for mobile browsers.
• No dark mode.
• As you can see, the UI is rather primitive.

If croquis seems useful to you, but some features are missing for your use case, then please feel free to file an issue. (Of course I can't guarantee anything, but it will be nice to know that someone's interested.)

FAQ

Is it really 100 times faster?

With large data, croquis can be several hundred times faster than other popular libraries. With very small data, there's less difference, as fixed-size overheads start to dominate.

Can we use it outside of Jupyter Notebook?

No, croquis is currently tied to Jupyter's message passing architecture, and all computation is done in the backend, so it needs an active Jupyter Python kernel.

How does it work?

Unlike most other similar libraries, croquis works by running a C++ "tile server," which computes fixed-sized "tiles" which is then sent back to the browser. If you have used Google Maps, the idea should be familiar. This has an important advantage:

• The browser only has to know about tiles. Hence, the size of the data the browser needs to know is independent of the data set size.

As a result, the browser stays lean and "snappy" even with massive data. (As explained in the reference, we support copy_data=False option that even eliminates data copies altogether.)

Moreover, unlike the browser's single-threaded javascript code, the C++-based tile server can draw multiple tiles in parallel, which allows even more speedup.

(On the other hand, there are drawbacks - we have to basically re-implement every graph drawing algorithm inside this tile server, not being able to use any javascript API, except for very trivial things like coordinate grids.)