6.5. Index Configuration

GeoMesa exposes a variety of configuration options that can be used to customize and optimize a given installation.

6.5.1. Setting Schema Options

Static properties of a SimpleFeatureType must be set when calling createSchema, and can’t be changed afterwards. Most properties are controlled through user-data values, either on the SimpleFeatureType or on a particular attribute. Setting the user data can be done in multiple ways.

If you are using a string to indicate your SimpleFeatureType (e.g. through the command line tools, or when using SimpleFeatureTypes.createType), you can append the type-level options to the end of the string, like so:

// append the user-data values to the end of the string, separated by a semi-colon
String spec = "name:String,dtg:Date,*geom:Point:srid=4326;option.one='foo',option.two='bar'";
SimpleFeatureType sft = SimpleFeatureTypes.createType("mySft", spec);

If you have an existing simple feature type, or you are not using SimpleFeatureTypes.createType, you may set the values directly in the feature type:

// set the hint directly
SimpleFeatureType sft = ...
sft.getUserData().put("option.one", "foo");

If you are using TypeSafe configuration files to define your simple feature type, you may include a ‘user-data’ key:

geomesa {
  sfts {
    "mySft" = {
      attributes = [
        { name = name, type = String             }
        { name = dtg,  type = Date               }
        { name = geom, type = Point, srid = 4326 }
      ]
      user-data = {
        option.one = "foo"
      }
    }
  }
}

6.5.2. Setting Attribute Options

In addition to schema-level user data, each attribute also has user data associated with it. Just like the schema options, attribute user data can be set in multiple ways.

If you are using a string to indicate your SimpleFeatureType (e.g. through the command line tools, or when using SimpleFeatureTypes.createType), you can append the attribute options after the attribute type, separated with a colon:

// append the user-data after the attribute type, separated by a colon
String spec = "name:String:index=true,dtg:Date,*geom:Point:srid=4326";
SimpleFeatureType sft = SimpleFeatureTypes.createType("mySft", spec);

If you have an existing simple feature type, or you are not using SimpleFeatureTypes.createType, you may set the user data directly in the attribute descriptor:

// set the hint directly
SimpleFeatureType sft = ...
sft.getDescriptor("name").getUserData().put("index", "true");

If you are using TypeSafe configuration files to define your simple feature type, you may add user data keys to the attribute elements:

geomesa {
  sfts {
    "mySft" = {
      attributes = [
        { name = name, type = String, index = true }
        { name = dtg,  type = Date                 }
        { name = geom, type = Point, srid = 4326   }
      ]
    }
  }
}

6.5.3. Setting the Indexed Date Attribute

For schemas that contain a date attribute, GeoMesa will use the attribute as part of the primary Z3/XZ3 index. If a schema contains more than one date attribute, you may specify which attribute to use through the user-data key geomesa.index.dtg. If you would prefer to not index any date, you may disable it through the key geomesa.ignore.dtg. If nothing is specified, the first declared date attribute will be used.

// specify the attribute 'myDate' as the indexed date
sft1.getUserData().put("geomesa.index.dtg", "myDate");

// disable indexing by date
sft2.getUserData().put("geomesa.ignore.dtg", true);

6.5.4. Customizing Index Creation

To speed up ingestion, or because you are only using certain query patterns, you may disable some indices. The indices are created when calling createSchema. If nothing is specified, the Z2/Z3 (or XZ2/XZ3 depending on geometry type) indices and record indices will all be created, as well as any attribute indices you have defined.

Warning

Certain queries may be much slower if you disable an index.

To enable only certain indices, you may set a user data value in your simple feature type. The user data key is geomesa.indices.enabled, and it should contain a comma-delimited list containing a subset of index identifiers, as specified in Index Overview.

See Setting Schema Options for details on setting user data. If you are using the GeoMesa SchemaBuilder, you may instead call the indexes methods:

import org.locationtech.geomesa.utils.geotools.SchemaBuilder

val sft = SchemaBuilder.builder()
    .addString("name")
    .addDate("dtg")
    .addPoint("geom", default = true)
    .userData
    .indices(List("id", "z3", "attr"))
    .build("mySft")

6.5.5. Configuring Feature ID Encoding

While feature IDs can be any string, a common use case is to use UUIDs. A UUID is a globally unique, specially formatted string composed of hex characters in the format {8}-{4}-{4}-{4}-{12}, for example 28a12c18-e5ae-4c04-ae7b-bf7cdbfaf234. A UUID can also be considered as a 128 bit number, which can be serialized in a smaller size.

You can indicate that feature IDs are UUIDs using the user data key geomesa.fid.uuid. If set before calling createSchema, then feature IDs will be serialized as 16 byte numbers instead of 36 byte strings, saving some overhead:

sft.getUserData().put("geomesa.fid.uuid", "true");
datastore.createSchema(sft);

If the schema is already created, you may still retroactively indicate that feature IDs are UUIDs, but you must also indicate that they are not serialized that way using geomesa.fid.uuid-encoded. This may still provide some benefit when exporting data in certain formats (e.g. Arrow):

SimpleFeatureType existing = datastore.getSchema("existing");
existing.getUserData().put("geomesa.fid.uuid", "true");
existing.getUserData().put("geomesa.fid.uuid-encoded", "false");
datastore.updateSchema("existing", existing);

Warning

Ensure that you use valid UUIDs if you indicate that you are using them. Otherwise you will experience exceptions writing and/or reading data.

6.5.6. Configuring Geometry Serialization

By default, geometries are serialized using a modified version of the well-known binary (WKB) format. Alternatively, geometries may be serialized using the tiny well-known binary (TWKB) format. TWKB will be smaller on disk, but does not allow full double floating point precision. For point geometries, TWKB will take 4-12 bytes (depending on the precision specified), compared to 18 bytes for WKB. For line strings, polygons, or other geometries with multiple coordinates, the space savings will be greater due to TWKB’s delta encoding scheme.

For any geometry type attribute, TWKB serialization can be enabled by setting the floating point precision through the precision user-data key. Precision indicates the number of decimal places that will be stored, and must be between -7 and 7, inclusive. A negative precision can be used to indicate rounding of whole numbers to the left of the decimal place. For reference, 6 digits of latitude/longitude precision can store a resolution of approximately 10cm.

For geometries with more than two dimensions, the precision of the Z and M dimensions may be specified separately. Generally these dimensions do not need to be stored with the same resolution as X/Y. By default, Z will be stored with precision 1, and M with precision 0. To change this, specify the additional precisions after the X/Y precision, separated with commas. For example, 6,1,0 would set the X/Y precision to 6, the Z precision to 1 and the M precision to 0. Z and M precisions must be between 0 and 7, inclusive.

TWKB serialization can be set when creating a new schema, but can also be enabled at any time through the updateSchema method. If modifying an existing schema, any data already written will not be updated.

SimpleFeatureType sft = ...
sft.getDescriptor("geom").getUserData().put("precision", "4");

See Setting Attribute Options for details on how to set attribute options.

6.5.7. Configuring Column Groups

For back-ends that support it (currently HBase and Accumulo), subsets of attributes may be replicated into separate column groups. When possible, only the reduced column groups will be scanned for a query, which avoids having to read unused data from disk. For schemas with a large number of attributes, this can speed up some queries, at the cost of writing more data to disk.

Column groups are specified per attribute, using attribute-level user data. An attribute may belong to multiple column groups, in which case it will be replicated multiple times. All attributes will belong to the default column group without having to specify anything. See Setting Attribute Options for details on how to set attribute options.

Column groups are specified using the user data key column-groups, with the value being a comma-delimited list of groups for that attribute. It is recommended to keep column group names short (ideally a single character), in order to minimize disk usage. If a column group conflicts with one of the default groups used by GeoMesa, it will throw an exception when creating the schema. Currently, the reserved groups are d for HBase and F, A, I, and B for Accumulo.

Java

spec

config

SchemaBuilder

SimpleFeatureType sft = ...
sft.getDescriptor("name").getUserData().put("column-groups", "a,b");

import org.locationtech.geomesa.utils.geotools.SimpleFeatureTypes
// for java, use org.locationtech.geomesa.utils.interop.SimpleFeatureTypes

val spec = "name:String:column-groups=a,dtg:Date:column-groups='a,b',*geom:Point:srid=4326:column-groups='a,b'"
SimpleFeatureTypes.createType("mySft", spec)

geomesa {
  sfts {
    "mySft" = {
      attributes = [
        { name = "name", type = "String", column-groups = "a"                }
        { name = "dtg",  type = "Date",   column-groups = "a,b"              }
        { name = "geom", type = "Point",  column-groups = "a,b", srid = 4326 }
      ]
    }
  }
}

import org.locationtech.geomesa.utils.geotools.SchemaBuilder

val sft = SchemaBuilder.builder()
    .addString("name").withColumnGroups("a")
    .addDate("dtg").withColumnGroups("a", "b")
    .addPoint("geom", default = true).withColumnGroups("a", "b")
    .build("mySft")

6.5.8. Configuring Z-Index Shards

GeoMesa allows configuration of the number of shards (or splits) into which the Z2/Z3/XZ2/XZ3 indices are divided. This parameter may be changed individually for each SimpleFeatureType. If nothing is specified, GeoMesa will default to 4 shards. The number of shards must be between 1 and 127.

Shards allow us to pre-split tables, which provides some initial parallelism for reads and writes. As more data is written, tables will generally split based on size, thus obviating the need for explicit shards. For small data sets, shards are more important as the tables might never split due to size. Setting the number of shards too high can reduce performance, as it requires more calculations to be performed per query.

The number of shards is set when calling createSchema. It may be specified through the simple feature type user data using the key geomesa.z.splits. See Setting Schema Options for details on setting user data.

sft.getUserData().put("geomesa.z.splits", "4");

6.5.9. Configuring Z-Index Time Interval

GeoMesa uses a z-curve index for time-based queries. By default, time is split into week-long chunks and indexed per week. If your queries are typically much larger or smaller than one week, you may wish to partition at a different interval. GeoMesa provides four intervals - day, week, month or year. As the interval gets larger, fewer partitions must be examined for a query, but the precision of each interval will go down.

If you typically query months of data at a time, then indexing per month may provide better performance. Alternatively, if you typically query minutes of data at a time, indexing per day may be faster. The default per week partitioning tends to provides a good balance for most scenarios. Note that the optimal partitioning depends on query patterns, not the distribution of data.

The time interval is set when calling createSchema. It may be specified through the simple feature type user data using the key geomesa.z3.interval. See Setting Schema Options for details on setting user data.

sft.getUserData().put("geomesa.z3.interval", "month");

6.5.10. Configuring XZ-Index Precision

GeoMesa uses an extended z-curve index for storing geometries with extents. The index can be customized by specifying the resolution level used to store geometries. By default, the resolution level is 12. If you have very large geometries, you may want to lower this value. Conversely, if you have very small geometries, you may want to raise it.

The resolution level for an index is set when calling createSchema. It may be specified through the simple feature type user data using the key geomesa.xz.precision. See Setting Schema Options for details on setting user data.

sft.getUserData().put("geomesa.xz.precision", 12);

For more information on resolution level (g), see “XZ-Ordering: A Space-Filling Curve for Objects with Spatial Extension” by Böhm, Klump and Kriegel.

6.5.11. Configuring Attribute Index Shards

GeoMesa allows configuration of the number of shards (or splits) into which the attribute indices are divided. This parameter may be changed individually for each SimpleFeatureType. If nothing is specified, GeoMesa will default to 4 shards. The number of shards must be between 1 and 127.

See Configuring Z-Index Shards for more background on shards.

The number of shards is set when calling createSchema. It may be specified through the simple feature type user data using the key geomesa.attr.splits. See Setting Schema Options for details on setting user data.

sft.getUserData().put("geomesa.attr.splits", "4");

6.5.12. Configuring Attribute Cardinality

GeoMesa allows attributes to be marked as either high or low cardinality. If set, this hint will be used in query planning. For more information, see Cardinality Hints.

To set the cardinality of an attribute, use the key cardinality on the attribute, with a value of high or low.

Java

SchemaBuilder

SimpleFeatureType sft = ...
sft.getDescriptor("name").getUserData().put("index", "true");
sft.getDescriptor("name").getUserData().put("cardinality", "high");

import org.locationtech.geomesa.utils.geotools.SchemaBuilder
import org.locationtech.geomesa.utils.stats.Cardinality

val sft = SchemaBuilder.builder()
    .addString("name").withIndex(Cardinality.HIGH)
    .addDate("dtg")
    .addPoint("geom", default = true)
    .build("mySft")

6.5.13. Configuring Partitioned Indices

To help with large data sets, GeoMesa can partition each index into separate tables, based on the attributes of each feature. Having multiple tables for a single index can make it simpler to manage a cluster, for example by making it trivial to delete old data. This functionality is currently supported in HBase, Accumulo and Cassandra.

Partitioning must be specified through user data when creating a simple feature type, before calling createSchema. To indicate a partitioning scheme, use the key geomesa.table.partition. Currently the only valid value is time, to indicate time-based partitioning:

Java

SchemaBuilder

sft.getUserData().put("geomesa.table.partition", "time");

import org.locationtech.geomesa.utils.geotools.SchemaBuilder

val sft = SchemaBuilder.builder()
    .addString("name")
    .addDate("dtg")
    .addPoint("geom", default = true)
    .userData
    .partitioned()
    .build("mySft")

Note that to enable partitioning the schema must contain a default date field.

When partitioning is enabled, each index will consist of multiple physical tables. The tables are partitioned based on the Z-interval (see Configuring Z-Index Time Interval). Tables are created dynamically when needed.

Partitioned tables can still be pre-split, as described in Configuring Index Splits. For Z3 splits, the min/max date configurations are automatically determined by the partition, and do not need to be specified.

When a query must scan multiple tables, by default the tables will be scanned sequentially. To instead scan the tables in parallel, set the sytem property geomesa.partition.scan.parallel=true. Note that when enabled, queries that span many partitions may place a large load on the system.

The GeoMesa command line tools provide functions for managing partitions; see manage-partitions for details.

6.5.14. Configuring Index Splits

When planning to ingest large amounts of data, if the distribution is known up front, it can be useful to pre-split tables before writing. This provides parallelism across a cluster from the start, and doesn’t depend on implementation triggers (which typically split tables based on size).

Splits are managed through implementations of the org.locationtech.geomesa.index.conf.TableSplitter interface.

6.5.14.1. Specifying a Table Splitter

A table splitter may be specified through user data when creating a simple feature type, before calling createSchema.

To indicate the table splitter class, use the key table.splitter.class:

sft.getUserData().put("table.splitter.class", "org.example.CustomSplitter");

To indicate any options for the given table splitter, use the key table.splitter.options:

sft.getUserData().put("table.splitter.options", "foo,bar,baz");

See Setting Schema Options for details on setting user data.

6.5.14.2. The Default Table Splitter

Generally, table.splitter.class can be omitted. If so, GeoMesa will use a default implementation that allows for a flexible configuration using table.splitter.options. If no options are specified, then all tables will generally create 4 split (based on the number of shards). The default ID index splits assume that feature IDs are randomly distributed UUIDs.

For the default splitter, table.splitter.options should consist of comma-separated entries, in the form key1:value1,key2:value2. Entries related to a given index should start with the index identifier, e.g. one of id, z3, z2 or attr (xz3 and xz2 indices use z3 and z2, respectively).

Index	Option	Description
Z3/XZ3	z3.min	The minimum date for the data
	z3.max	The maximum date for the data
	z3.bits	The number of leading bits to split on
Z2/XZ2	z2.bits	The number of leading bits to split on
ID	id.pattern	Split pattern
Attribute	attr.<attribute>.pattern	Split pattern for a given attribute

6.5.14.2.1. Z3/XZ3 Splits

Dates are used to split based on the Z3 time prefix (typically weeks). They are specified in the form yyyy-MM-dd. If the minimum date is specified, but the maximum date is not, it will default to the current date. After the dates, the Z value can be split based on a number of bits (note that due to the index format, bits can not be specified without dates). For example, specifying two bits would create splits 00, 01, 10 and 11. The total number of splits created will be <number of z shards> * <number of time periods> * 2 ^ <number of bits>.

6.5.14.2.2. Z2/XZ2 Splits

If any options are given, the number of bits must be specified. For example, specifying two bits would create splits 00, 01, 10 and 11. The total number of splits created will be <number of z shards> * 2 ^ <number of bits>.

6.5.14.2.3. ID and Attribute Splits

Splits are defined by patterns. For an ID index, the pattern(s) are applied to the single feature ID. For an attribute index, each attribute that is indexed can be configured separately, by specifying the attribute name as part of the option. For example, given the schema name:String:index=true,*geom:Point:srid=4326, the name attribute splits can be configured with attr.name.pattern.

Patterns consist of one or more single characters or ranges enclosed in square brackets. Valid characters can be any of the numbers 0 to 9, or any letter a to z, in upper or lower case. Ranges are two characters separated by a dash. Each set of brackets corresponds to a single character, allowing for nested splits.

For example, the pattern [0-9] would create 10 splits, based on the numbers 0 through 9. The pattern [0-9][0-9] would create 100 splits. The pattern [0-9a-f] would create 16 splits based on lower-case hex characters. The pattern [0-9A-F] would do the same with upper-case characters.

For data hot-spots, multiple patterns can be specified by adding additional options with a 2, 3, etc appended to the key. For example, if most of the name values start with the letter f and t, splits could be specified as attr.name.pattern:[a-z],attr.name.pattern2:[f][a-z],attr.name.pattern3:[t][a-z]

For number-type attributes, only numbers are considered valid characters. Due to lexicoding, normal number prefixing will not work correctly. E.g., if the data has numbers in the range 8000-9000, specifying [8-9][0-9] will not split the data properly. Instead, trailing zeros should be added to reach the appropriate length, e.g. [8-9][0-9][0][0].

6.5.14.2.4. Full Example

import org.locationtech.geomesa.utils.interop.SimpleFeatureTypes;

String spec = "name:String:index=true,age:Int:index=true,dtg:Date,*geom:Point:srid=4326";
SimpleFeatureType sft = SimpleFeatureTypes.createType("foo", "spec");
sft.getUserData().put("table.splitter.options",
    "id.pattern:[0-9a-f],attr.name.pattern:[a-z],z3.min:2018-01-01,z3.max:2018-01-31,z3.bits:2,z2.bits:4");

6.5.15. Configuring Cached Statistics

GeoMesa allows for collecting summary statistics for attributes during ingest, which are then stored and available for instant querying. Hints are set on individual attributes using the key keep-stats, as described in Setting Attribute Options.

Note

Cached statistics are currently only implemented for the Accumulo data store

Stats are always collected for the default geometry, default date and any indexed attributes. See Cost-Based Strategy for more details. In addition, any other attribute can be flagged for stats. This will cause the following stats to be collected for those attributes:

• Min/max (bounds)
• Top-k

Only attributes of the following types can be flagged for stats: String, Integer, Long, Float, Double, Date and Geometry.

For example:

// set the hint directly
SimpleFeatureType sft = ...
sft.getDescriptor("name").getUserData().put("keep-stats", "true");

See Analytic Commands and Accessing Stats through the GeoMesa API for information on reading cached stats.

6.5.16. Mixed Geometry Types

A common pitfall is to unnecessarily specify a generic geometry type when creating a schema. Because GeoMesa relies on the geometry type for indexing decisions, this can negatively impact performance.

If the default geometry type is Geometry (i.e. supporting both point and non-point features), you must explicitly enable “mixed” indexing mode. All other geometry types (Point, LineString, Polygon, etc) are not affected.

Mixed geometries must be declared when calling createSchema. It may be specified through the simple feature type user data using the key geomesa.mixed.geometries. See Setting Schema Options for details on setting user data.

sft.getUserData().put("geomesa.mixed.geometries", "true");