GeoLake Parquet

Compare the performance of GeoLake Parquet with other spatial parquet data formats:

The experiments run on 3 datasets:

	Geometry Type	Records	Points	APER
Porto Taxi	MultiLinePoint	1.7M	83M	49
TIGER2018/Roads	MultiLineString	18M	350M	19
MSBuildings	Polygon	125M	753M	6

APER: Average points in each record.

In these experiments, we use two spatial predicates(one large, one small) to hit different numbers of row groups respectively. Note that the predicates constructed for different datasets are different, as below figure shows:

Build & Run

./gradlew build
java -jar build/libs/geolake-benchmark-1.0-SNAPSHOT-all.jar

By default, it runs on a very small dataset which is too small to reflect the efficiency of GeoLake Parquet.

In order to run the 3 large datasets mentioned previously, you need to download it first.

wget https://star.cs.ucr.edu/datasets/portotaxi/download.geojson.gz -O - | gzip -d > build/resources/main/source/portotaxi.geojson
wget https://star.cs.ucr.edu/datasets/TIGER2018/ROADS/download.geojson.gz -O - | gzip -d > build/resources/main/source/tiger_2018_roads.geojson
wget https://star.cs.ucr.edu/datasets/MSBuildings/download.geojson.gz -O - | gzip -d > build/resources/main/source/ucr_msbuildings.geojson

After downloading, you can run experiment on each dataset with the following commands:

java -jar build/libs/geolake-benchmark-1.0-SNAPSHOT-all.jar portotaxi
java -jar build/libs/geolake-benchmark-1.0-SNAPSHOT-all.jar tiger
java -jar build/libs/geolake-benchmark-1.0-SNAPSHOT-all.jar msbuildings

The source code of GeoLake Parquet will be released in the GeoLake project.