Kerchunk, hvPlot, and Datashader: Visualizing datasets on-the-fly

Overview

This notebook will demonstrate how to use Kerchunk with hvPlot and Datashader to lazily visualize a reference dataset in a streaming fashion.

We will be building off the references generated through the notebook content from thePangeo_Forge notebook, so it’s encouraged you first go through that.

Prerequisites

Concepts	Importance	Notes
Kerchunk Basics	Required	Core
Introduction to Xarray	Required	IO
Introduction to hvPlot	Required	Data Visualization
Introduction to Datashader	Required	Big Data Visualization

• Time to learn: 10 minutes

---

Motivation

Using Kerchunk, we don’t have to create a copy of the data–instead we create a collection of reference files, so that the original data files can be read as if they were Zarr.

This enables visualization on-the-fly; simply pass in the URL to the dataset and use hvplot.

Getting to Know The Data

gridMET is a high-resolution daily meteorological dataset covering CONUS from 1979-2023. It is produced by the Climatology Lab at UC Merced. In this example, we are going to look create a virtual Zarr dataset of a derived variable, Burn Index.

Imports

import hvplot.xarray  # noqa
import xarray as xr

Opening the Kerchunk Dataset

Now, it’s a matter of opening the Kerchunk dataset and calling hvplot with the rasterize=True keyword argument.

If you’re running this notebook locally, try zooming around the map by hovering over the plot and scrolling; it should update fairly quickly. Note, it will not update if you’re viewing this on the docs page online as there is no backend server, but don’t fret because there’s a demo GIF below!

%%timeit -r 1 -n 1


storage_options = {
    "remote_protocol": "http",
    "skip_instance_cache": True,
}  # options passed to fsspec
open_dataset_options = {"chunks": {}, "decode_coords": "all"}  # opens passed to xarray

ds_kerchunk = xr.open_dataset(
    "references/Pangeo_Forge/reference.json",
    engine="kerchunk",
    storage_options=storage_options,
    **open_dataset_options,
)

display(ds_kerchunk.hvplot("lon", "lat", rasterize=True))  # noqa

35.4 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)

Comparing Against THREDDS

Now, we will be repeating the previous cell, but with THREDDS.

Note how the initial load is longer.

If you’re running the notebook locally (or a demo GIF below), zooming in/out also takes longer to finish buffering as well.

%%timeit -r 1 -n 1


def url_gen(year):
    return (
        f"http://thredds.northwestknowledge.net:8080/thredds/dodsC/MET/bi/bi_{year}.nc"
    )


years = list(range(1979, 1980))
urls_list = [url_gen(year) for year in years]
netcdf_ds = xr.open_mfdataset(urls_list, engine="netcdf4")
display(netcdf_ds.hvplot("lon", "lat", rasterize=True))  # noqa

6.12 s ± 0 ns per loop (mean ± std. dev. of 1 run, 1 loop each)