Parallel GeoPandas with Dask

This project is not in a functional state and should not be relied upon. No guarantee of support is provided.

This was was originally implemented to demonstrate speedups from parallelism alongside an experimental Cythonized branch of GeoPandas. That cythonized branch has since evolved to the point where the code here no longer works with the latest version.

If you really want to get this to work then you should checkout the geopandas-cython branch of geopandas at about 2017-09-21 and build from source (this may not be fun). But really the solution is probably to wait until everything settles. There is no known timeline for this.

If you would like to see this project in a more stable state then you might consider pitching in with developer time or with financial support from you or your company.

Given a GeoPandas dataframe

import geopandas as gpd
df = gpd.read_file('...')

We can repartition it into a Dask-GeoPandas dataframe either naively by rows. This does not provide a spatial partitioning and so won't gain the efficiencies of spatial reasoning, but will still provide basic multi-core parallelism.

import dask_geopandas as dg
ddf = dg.from_pandas(df, npartitions=4)

We can also repartition by a set of known regions. This suffers an upfront cost of a spatial join, but enables spatial-aware computations in the future to be faster.

regions = gpd.read_file('boundaries.shp')
ddf = dg.repartition(df, regions)

Additionally, if you have a distributed dask.dataframe you can pass columns of x-y points to the set_geometry method. Currently this only supports point data.

import dask.dataframe as dd
import dask_geopandas as dg

df = dd.read_csv('...')

df = df.set_geometry(df[['latitude', 'longitude']])