GeoSpark: Bring sf to spark

Introduction & Philosophy

Goal: make traditional GISer handle geospatial big data easier.

The origin idea comes from Uber, which proposed a ESRI Hive UDF + Presto solution to solve large-scale geospatial data processing problem with spatial index in production.

However, The Uber solution is not open source yet and Presto is not popular than Spark.

In that, geospark R package aims at bringing local sf functions to distributed spark mode with GeoSpark scala package.

Currently, geospark support the most of important sf functions in spark, here is a summary comparison. And the geospark R package is keeping close with geospatial and big data community, which powered by sparklyr, sf, dplyr and dbplyr.

Installation

This package requires Apache Spark 2.X which you can install using sparklyr::install_spark("2.3"), and spark2.4 is not supported yet; in addition, you can install geospark as follows:

pak::pkg_install("harryprince/geospark")

Getting Started

In this example we will join spatial data using quadrad tree indexing. First, we will initialize the geospark extension and connect to Spark using sparklyr:

library(sparklyr)
library(geospark)

sc <- spark_connect(master = "local")
register_gis(sc)

Next we will load some spatial dataset containing as polygons and points.

polygons <- read.table(system.file(package="geospark","examples/polygons.txt"), sep="|", col.names=c("area","geom"))
points <- read.table(system.file(package="geospark","examples/points.txt"), sep="|", col.names=c("city","state","geom"))

polygons_wkt <- copy_to(sc, polygons)
points_wkt <- copy_to(sc, points)

And we can quickly visulize the dataset by mapview and sf.

M1 = polygons %>%
sf::st_as_sf(wkt="geom") %>% mapview::mapview()


M2 = points %>%
sf::st_as_sf(wkt="geom") %>% mapview::mapview()

M1+M2

The SQL Mode

Now we can perform a GeoSpatial join using the st_contains which converts wkt into geometry object. To get the original data from wkt format, we will use the st_geomfromwkt functions. We can execute this spatial query using DBI:

DBI::dbGetQuery(sc, "
  SELECT area, state, count(*) cnt FROM
    (SELECT area, ST_GeomFromWKT(polygons.geom) as y FROM polygons) polygons
  INNER JOIN
    (SELECT ST_GeomFromWKT (points.geom) as x, state, city FROM points) points
  WHERE ST_Contains(polygons.y,points.x) GROUP BY area, state")

             area state cnt
1      texas area    TX  10
2     dakota area    SD   1
3     dakota area    ND  10
4 california area    CA  10
5   new york area    NY   9

The Tidyverse Mode

You can also perform this query using dplyr as follows:

library(dplyr)
polygons_wkt <- mutate(polygons_wkt, y = st_geomfromwkt(geom))
points_wkt <- mutate(points_wkt, x = st_geomfromwkt(geom))

sc_res <- inner_join(polygons_wkt,
                     points_wkt,
                     sql_on = sql("st_contains(y,x)")) %>% 
  group_by(area, state) %>%
  summarise(cnt = n()) 
  
sc_res %>%
  head()

# Source: spark<?> [?? x 3]
# Groups: area
  area            state   cnt
  <chr>           <chr> <dbl>
1 texas area      TX       10
2 dakota area     SD        1
3 dakota area     ND       10
4 california area CA       10
5 new york area   NY        9

The final result can be present by leaflet.

Idx_df = collect(sc_res) %>% 
right_join(polygons,by = (c("area"="area"))) %>% 
sf::st_as_sf(wkt="geom")

Idx_df %>% 
leaflet::leaflet() %>% 
leaflet::addTiles() %>% 
leaflet::addPolygons(popup = ~as.character(cnt),color=~colormap::colormap_pal()(cnt)) 

Finally, we can disconnect:

spark_disconnect_all()

Performance

Configuration

To improve performance, it is recommended to use the KryoSerializer and the GeoSparkKryoRegistrator before connecting as follows:

conf <- spark_config()
conf$spark.serializer <- "org.apache.spark.serializer.KryoSerializer"
conf$spark.kryo.registrator <- "org.datasyslab.geospark.serde.GeoSparkKryoRegistrator"

Benchmarks

This performance comparison is an extract from the original GeoSpark: A Cluster Computing Framework for Processing Spatial Data paper:

Query performance(ms),

According to this paper, the Geospark SQL definitely outperforms PG and ESRI UDF under a very large data set.

If you are wondering how the spatial index accelerate the query process, here is a good Uber example: Unwinding Uber’s Most Efficient Service and the Chinese translation version

Functions

Constructor

Geometry Measurement

Spatial Join

Distance join

ST_Distance:

Spark GIS SQL mode example:

SELECT *
FROM pointdf1, pointdf2
WHERE ST_Distance(pointdf1.pointshape1,pointdf2.pointshape2) <= 2

Tidyverse style example:

st_join(x = pointdf1,
           y = pointdf2,
           join = sql("ST_Distance(pointshape1, pointshape2) <= 2"))

Aggregation

More Advacned Functions

Architecture