16.5. SparkSQL

GeoMesa SparkSQL support builds upon the DataSet/DataFrame API present in the Spark SQL module to provide geospatial capabilities. This includes custom geospatial data types and functions, the ability to create a DataFrame from a GeoTools DataStore, and optimizations to improve SQL query performance.

GeoMesa SparkSQL code is provided by the geomesa-spark-sql module:

<dependency>
  <groupId>org.locationtech.geomesa</groupId>
  <artifactId>geomesa-spark-sql_2.11</artifactId>
  // version, etc.
</dependency>

16.5.1. Example

The following is a Scala example of connecting to GeoMesa Accumulo via SparkSQL:

// DataStore params to a hypothetical GeoMesa Accumulo table
val dsParams = Map(
  "accumulo.instance.id"   -> "instance",
  "accumulo.zookeepers"    -> "zoo1,zoo2,zoo3",
  "accumulo.user"          -> "user",
  "accumulo.password"      -> "*****",
  "accumulo.catalog"       -> "geomesa_catalog",
  "geomesa.security.auths" -> "USER,ADMIN")

// Create SparkSession
val sparkSession = SparkSession.builder()
  .appName("testSpark")
  .config("spark.sql.crossJoin.enabled", "true")
  .master("local[*]")
  .getOrCreate()

// Create DataFrame using the "geomesa" format
val dataFrame = sparkSession.read
  .format("geomesa")
  .options(dsParams)
  .option("geomesa.feature", "chicago")
  .load()
dataFrame.createOrReplaceTempView("chicago")

// Query against the "chicago" schema
val sqlQuery = "select * from chicago where st_contains(st_makeBBOX(0.0, 0.0, 90.0, 90.0), geom)"
val resultDataFrame = sparkSession.sql(sqlQuery)

resultDataFrame.show
/*
+-------+------+-----------+--------------------+-----------------+
|__fid__|arrest|case_number|                 dtg|             geom|
+-------+------+-----------+--------------------+-----------------+
|      4|  true|          4|2016-01-04 00:00:...|POINT (76.5 38.5)|
|      5|  true|          5|2016-01-05 00:00:...|    POINT (77 38)|
|      6|  true|          6|2016-01-06 00:00:...|    POINT (78 39)|
|      7|  true|          7|2016-01-07 00:00:...|    POINT (20 20)|
|      9|  true|          9|2016-01-09 00:00:...|    POINT (50 50)|
+-------+------+-----------+--------------------+-----------------+
*/

16.5.2. Configuration

Because GeoMesa SparkSQL stacks on top of the geomesa-spark-core module, one or more of the SpatialRDDProvider implementations must be included on the classpath. See Configuration for details on setting up the Spark classpath.

Note

In most cases, it is not necessary to set up the Kryo serialization described in Simple Feature Serialization when using SparkSQL. However, this may be required when using the GeoTools RDD Provider.

If you will be JOIN-ing multiple DataFrames together, it will be necessary to add the spark.sql.crossJoin.enabled property when creating the SparkSession object:

val spark = SparkSession.builder().
   // ...
   config("spark.sql.crossJoin.enabled", "true").
   // ...
   getOrCreate()

Warning

Cross-joins can be very, very inefficient. Take care to ensure that one or both sets of data joined are very small, and consider using the broadcast() method to ensure that at least one DataFrame joined is in memory.

16.5.3. Usage

To create a GeoMesa SparkSQL-enabled DataFrame with data corresponding to a particular feature type, do the following:

// dsParams contains the parameters to pass to the data store
val dataFrame = sparkSession.read
  .format("geomesa")
  .options(dsParams)
  .option("geomesa.feature", typeName)
  .load()

Specifically, invoking format("geomesa") registers the GeoMesa SparkSQL data source, and option("geomesa.feature", typeName) tells GeoMesa to use the feature type named typeName. This also registers the custom user-defined types and functions implemented in GeoMesa SparkSQL.

By registering a DataFrame as a temporary view, it is possible to access this data frame in subsequent SQL calls. For example:

dataFrame.createOrReplaceTempView("chicago")

makes it possible to call this data frame via the alias “chicago”:

val sqlQuery = "select * from chicago where st_contains(st_makeBBOX(0.0, 0.0, 90.0, 90.0), geom)"
val resultDataFrame = sparkSession.sql(sqlQuery)

Registering user-defined types and functions can also be done manually by invoking SQLTypes.init() on the SQLContext object of the Spark session:

SQLTypes.init(sparkSession.sqlContext)

It is also possible to write a Spark DataFrame to a GeoMesa table with

dataFrame.write.format("geomesa").options("dsParams").option("geomesa.feature", "featureName").save()

This will automatically convert the data frame’s underlying RDD[Row] into an RDD[SimpleFeature] and write to the data store in parallel. For this to work, the feature type featureName must already exist in the data store.

16.5.4. Geospatial User-defined Types and Functions

The GeoMesa SparkSQL module takes several classes representing geometry objects (as described by the OGC OpenGIS Simple feature access common architecture specification and implemented by the Java Topology Suite) and registers them as user-defined types (UDTs) in SparkSQL. For example the Geometry class is registered as GeometryUDT. In GeoMesa SparkSQL the following types are registered:

  • GeometryUDT
  • PointUDT
  • LineStringUDT
  • PolygonUDT
  • MultiPointUDT
  • MultiLineStringUDT
  • MultiPolygonUDT
  • GeometryCollectionUDT

GeoMesa SparkSQL also implements a subset of the functions described in the OGC OpenGIS Simple feature access SQL option specification as SparkSQL user-defined functions (UDFs). These include functions for creating geometries, accessing properties of geometries, casting Geometry objects to more specific subclasses, outputting geometries in other formats, measuring spatial relationships between geometries, and processing geometries.

For example, the following SQL query

select * from chicago where st_contains(st_makeBBOX(0.0, 0.0, 90.0, 90.0), geom)

uses two UDFs–st_contains and st_makeBBOX–to find the rows in the chicago DataFrame where column geom is contained within the specified bounding box.

A complete list of the implemented UDFs is given in the next section (SparkSQL Functions).

16.5.5. In-memory Indexing

If your data is small enough to fit in the memory of your executors, you can tell GeoMesa SparkSQL to persist RDDs in memory and leverage the use of CQEngine as an in-memory indexed data store. To do this, add the option option("cache", "true") when creating your data frame. This will place an index on every attribute excluding the fid and the geometry. To index based on geometry, add the option option("indexGeom", "true"). Queries to this relation will automatically hit the cached RDD and query the in-memory data store that lives on each partition, which can yield significant speedups.

Given some knowledge of your data, it is also possible to ensure that the data will fit in memory by applying an initial query. This can be done with the query option. For example, option("query", "dtg AFTER 2016-12-31T23:59:59Z")

16.5.6. Spatial Partitioning and Faster Joins

Additional speedups can be attained by also spatially partitioning your data. Adding the option option("spatial", "true") will ensure that data that are spatially near each other will be placed on the same partition. By default, your data will be partitioned into an NxN grid, but there exist 4 total partitioning strategies, and each can be specified by name with option("strategy", strategyName)

EQUAL - Computes the bounds of your data and divides it into an NxN grid of equal size, where N = sqrt(numPartitions)

WEIGHTED - Like EQUAL, but ensures that equal proportions of the data along each axis are in each grid cell.

EARTH - Like EQUAL, but uses the whole earth as the bounds instead of computing them based on the data.

RTREE - Constructs an R-Tree based on a sample of the data, and uses a subset of the bounding rectangles as partition envelopes.

The advantages to spatially partitioning are two fold:

1) Queries with a spatial predicate that lies wholly in one partition can go directly to that partition, skipping the overhead of scanning partitions that will be certain to not include the desired data.

2) If two data sets are partitioned by the same scheme, resulting in the same partition envelopes for both relations, then spatial joins can use the partition envelope as a key in the join. This dramatically reduces the number of comparisons required to complete the join.

Additional data frame options allow for greater control over how partitions are created. For strategies that require a sample of the data (WEIGHTED and RTREE), sampleSize and thresholdMultiplier can be used to control how much of the underlying data is used in the decision process and how many items to allow in an RTree envelope.

Other useful options are as follows:

option("partitions", "n") - Specifies the number of partitions that the underlying RDDs will be (overrides default parallelism)

option("bounds", "POLYGON in WellKnownText") - Limits the bounds of the grid that WEIGHTED and EQUAL strategies use. All data that do not lie in these bounds will be placed in a separate partition

option("cover", "true") - Since only the EQUAL and EARTH partition strategies can guarantee that partition envelopes will be identical across relations, data frames with this option set will force the partitioning scheme of data frames that they are joined with to match its own.