plotly-resampler 0.11.1

plotly_resampler: visualize large sequential data by adding resampling functionality to Plotly figures

plotly-resampler improves the scalability of Plotly for visualizing large time series datasets. Specifically, our library dynamically aggregates time-series data respective to the current graph view, ensuring efficient and responsive updates during user interactions like panning or zooming via callbacks.

This core aggregation functionality is achieved by utilizing by time-series data point selection algorithms, for which plotly-resampler leverages the highly optimized implementations available in tsdownsample. Our default data aggregation method is MinMaxLTTB (and selects 1000 data points for plotting). For a deeper understanding of this method, you can consult to the algorithm's dedicated MinMaxLTTB repository and the associated research paper.

In this Plotly-Resampler demo over 110,000,000 data points are visualized!

🛠️ Installation

📋 Features

• Convenient to use:
  • just add either
    • register_plotly_resampler function to your notebook with the best suited mode argument.
    • FigureResampler decorator around a plotly Figure and call .show_dash()
    • FigureWidgetResampler decorator around a plotly Figure and output the instance in a cell
  • allows all other plotly figure construction flexibility to be used!
• Environment-independent
  • can be used in Jupyter, vscode-notebooks, Pycharm-notebooks, Google Colab, DataSpell, and even as application (on a server)
• Interface for various aggregation algorithms:
  • ability to develop or select your preferred sequence aggregation method

🚀 Usage

Add dynamic aggregation to your plotly Figure (unfold your fitting use case)

• 🤖 Automatically (minimal code overhead):

  Use the register_plotly_resampler function

  
  

  1. Import and call the register_plotly_resampler method
  2. Just use your regular graph construction code

  • code example:

    import plotly.graph_objects as go; import numpy as np
    from plotly_resampler import register_plotly_resampler
    
    # Call the register function once and all Figures/FigureWidgets will be wrapped
    # according to the register_plotly_resampler its `mode` argument
    register_plotly_resampler(mode='auto')
    
    x = np.arange(1_000_000)
    noisy_sin = (3 + np.sin(x / 200) + np.random.randn(len(x)) / 10) * x / 1_000
    
    
    # auto mode: when working in an IPython environment, this will automatically be a 
    # FigureWidgetResampler else, this will be an FigureResampler
    f = go.Figure()
    f.add_trace({"y": noisy_sin + 2, "name": "yp2"})
    f
    

  Note: This wraps all plotly graph object figures with a FigureResampler | FigureWidgetResampler. This can thus also be used for the plotly.express interface. 🎉

• 👷 Manually (higher data aggregation configurability, more speedup possibilities):

  • Within a jupyter environment without creating a web application
    1. wrap the plotly Figure with FigureWidgetResampler
    2. output the FigureWidgetResampler instance in a cell

    import plotly.graph_objects as go; import numpy as np
    from plotly_resampler import FigureResampler, FigureWidgetResampler
    
    x = np.arange(1_000_000)
    noisy_sin = (3 + np.sin(x / 200) + np.random.randn(len(x)) / 10) * x / 1_000
    
    # OPTION 1 - FigureWidgetResampler: dynamic aggregation via `FigureWidget.layout.on_change`
    fig = FigureWidgetResampler(go.Figure())
    fig.add_trace(go.Scattergl(name='noisy sine', showlegend=True), hf_x=x, hf_y=noisy_sin)
    
    fig
    

  • Using a web-application with dash callbacks
    1. wrap the plotly Figure with FigureResampler
    2. call .show_dash() on the Figure

    import plotly.graph_objects as go; import numpy as np
    from plotly_resampler import FigureResampler, FigureWidgetResampler
    
    x = np.arange(1_000_000)
    noisy_sin = (3 + np.sin(x / 200) + np.random.randn(len(x)) / 10) * x / 1_000
    
    # OPTION 2 - FigureResampler: dynamic aggregation via a Dash web-app
    fig = FigureResampler(go.Figure())
    fig.add_trace(go.Scattergl(name='noisy sine', showlegend=True), hf_x=x, hf_y=noisy_sin)
    
    fig.show_dash(mode='inline')
    

  Tip 💡: For significant faster initial loading of the Figure, we advise to wrap the constructor of the plotly Figure and add the trace data as hf_x and hf_y

Note: Any plotly Figure can be wrapped with FigureResampler and FigureWidgetResampler! 🎉 But only the go.Scatter/go.Scattergl traces are resampled.

🔭 The successor: FlexViz

plotly-resampler does one thing well: it makes a line trace fast inside an existing Plotly figure. The features asked for most often over the years are the ones this design cannot deliver: linked charts, cross-filtering, histograms and heatmaps that re-aggregate on zoom, and reading straight from Parquet instead of from an in-memory array.

FlexViz is our answer, built by the same authors. Every zoom, pan, and brush is answered by a lazy Polars aggregation with Rust kernels behind it, so a whole dashboard stays interactive at 100M+ rows.

If plotly-resampler covers your needs, keep using it. If you have hit its ceiling, FlexViz is where our new development goes: live demo · GitHub · docs

💭 Important considerations & tips

• When running the code on a server, you should forward the port of the FigureResampler.show_dash() method to your local machine.
  Note that you can add dynamic aggregation to plotly figures with the FigureWidgetResampler wrapper without needing to forward a port!
• The FigureWidgetResampler uses the IPython main thread for its data aggregation functionality, so when this main thread is occupied, no resampling logic can be executed. For example; if you perform long computations within your notebook, the kernel will be occupied during these computations, and will only execute the resampling operations that take place during these computations after finishing that computation.
• In general, when using downsampling one should be aware of (possible) aliasing effects. The [R] in the legend indicates when the corresponding trace is being resampled (and thus possibly distorted) or not. Additionally, the ~<range> suffix represent the mean aggregation bin size in terms of the sequence index.
• The plotly autoscale event (triggered by the autoscale button or a double-click within the graph), does not reset the axes but autoscales the current graph-view of plotly-resampler figures. This design choice was made as it seemed more intuitive for the developers to support this behavior with double-click than the default axes-reset behavior. The graph axes can ofcourse be resetted by using the reset_axis button. If you want to give feedback and discuss this further with the developers, see issue #49.

📜 Citation and papers

The paper about the plotly-resampler toolkit itself (preprint): https://arxiv.org/abs/2206.08703

@inproceedings{van2022plotly,
  title={Plotly-resampler: Effective visual analytics for large time series},
  author={Van Der Donckt, Jonas and Van Der Donckt, Jeroen and Deprost, Emiel and Van Hoecke, Sofie},
  booktitle={2022 IEEE Visualization and Visual Analytics (VIS)},
  pages={21--25},
  year={2022},
  organization={IEEE}
}

Related papers:

• Visual representativeness of time series data point selection algorithms (preprint): https://arxiv.org/abs/2304.00900
  code: https://github.com/predict-idlab/ts-datapoint-selection-vis
• MinMaxLTTB - an efficient data point selection algorithm (preprint): https://arxiv.org/abs/2305.00332
  code: https://github.com/predict-idlab/MinMaxLTTB

👤 Jonas Van Der Donckt, Jeroen Van Der Donckt, Emiel Deprost