Geometry and Geography Tables

WherobotsDB stores tables in Apache Iceberg. New tables in Wherobots-managed catalogs, such as org_catalog, are format-version 3 (V3), which has geometry and geography column types, and Wherobots adds a raster column type. Tables created before September 15, 2026 were created at format-version 2 (V2), and Wherobots doesn’t raise them to V3. Which column types a table can have, and whether WherobotsDB skips data files for spatial filters, depend on its format version; see the Iceberg Tables Overview. This page shows how to create, write, and query tables with geometry and geography columns. The examples use the catalog name org_catalog; replace it with the name of the catalog you write to.

Readable as geometry by other engines

The geometry type is recorded in the table schema and in the Parquet files, so an Apache Iceberg engine outside Wherobots that supports geospatial types sees a geometry column instead of a binary column it has to decode itself.

One CRS per column, enforced

A column carries exactly one Coordinate Reference System (CRS). A value whose Spatial Reference Identifier (SRID) disagrees with it is rejected instead of being stored under the wrong CRS.

Tables created before September 15, 2026, and tables created from the result of an ST_ function, have geometry columns with no declared CRS. To check a table and upgrade it, see Upgrade Existing Geometry Tables.

Creating a table with a geometry column

Apache Iceberg tables have a GEOMETRY data type to represent geospatial data. Declare the column with its CRS in parentheses, as an SRID:

Python
Scala
Java

sedona.sql("""
CREATE TABLE IF NOT EXISTS org_catalog.test_db.test_table (
    id bigint,
    data string,
    geom geometry(4326)
) USING iceberg
PARTITIONED BY (bucket(4, id))""")

sedona.sql("""
CREATE TABLE IF NOT EXISTS org_catalog.test_db.test_table (
    id bigint,
    data string,
    geom geometry(4326)
) USING iceberg
PARTITIONED BY (bucket(4, id))""")

sedona.sql(
    "CREATE TABLE IF NOT EXISTS org_catalog.test_db.test_table (" +
    "id bigint," +
    "data string," +
    "geom geometry(4326)" +
    ") USING iceberg " +
    "PARTITIONED BY (bucket(4, id))"
);

This will create an empty table partitioned by the bucketed value of id, with a geometry column in SRID 4326, which is World Geodetic System 1984 (WGS84) longitude and latitude. A geometry column requires a declared CRS: write geometry(4326), not geometry. Currently, a geometry(<srid>) column can declare SRID 0, 3857, or 4326, and a geography(<srid>) column can declare 4326. Declaring another SRID, such as 4269 or 32633, fails with ST_INVALID_SRID_VALUE. For data in another CRS, see Data in another CRS. USING iceberg names the Apache Iceberg data source. havasu.iceberg is an accepted alias for it, in USING and in format(...). A table with a geometry(<srid>) or geography(<srid>) column is Apache Iceberg format-version 3, which is the default for new tables, so leave 'format-version' out of TBLPROPERTIES.

Error: ... is not supported until v3

The table was created with TBLPROPERTIES ('format-version'='2'). Remove the property:

Invalid schema for v2: - Invalid type for geom: geometry(OGC:CRS84) is not supported until v3

Error: bare GEOMETRY in a column definition does not declare a CRS

The column was declared as geometry with no CRS in parentheses. Add the SRID:

Operation not allowed: bare GEOMETRY in a column definition does not declare a CRS:
with spark.sql.geospatial.enabled=true it would silently create a Havasu-legacy column
with no CRS enforcement. Declare a CRS (e.g. GEOMETRY(4326)) to create an Iceberg V3
native column, or write the table from a Sedona GeometryUDT DataFrame if a Havasu-legacy
column is intended.

Inspecting the schema

Inspect the table schema with DESCRIBE TABLE:

Python
Scala
Java

sedona.sql("DESCRIBE TABLE org_catalog.test_db.test_table").show()

sedona.sql("DESCRIBE TABLE org_catalog.test_db.test_table").show()

sedona.sql("DESCRIBE TABLE org_catalog.test_db.test_table").show();

+--------------+-------------+-------+
|      col_name|    data_type|comment|
+--------------+-------------+-------+
|            id|       bigint|   NULL|
|          data|       string|   NULL|
|          geom|     geometry|   NULL|
|              |             |       |
|# Partitioning|             |       |
|        Part 0|bucket(4, id)|       |
+--------------+-------------+-------+

DESCRIBE TABLE shows the type as geometry without the CRS. To see the CRS, run SHOW CREATE TABLE, which renders the column as geom GEOMETRY(4326):

Python
Scala
Java

print(sedona.sql("SHOW CREATE TABLE org_catalog.test_db.test_table").collect()[0][0])

println(sedona.sql("SHOW CREATE TABLE org_catalog.test_db.test_table").collect()(0).getString(0))

System.out.println(sedona.sql("SHOW CREATE TABLE org_catalog.test_db.test_table").collectAsList().get(0).getString(0));

CREATE TABLE org_catalog.test_db.test_table (
id BIGINT,
data STRING,
geom GEOMETRY(4326))
USING iceberg
PARTITIONED BY (bucket(4, id))
...

The statement SHOW CREATE TABLE returns recreates an equivalent table, so it is also the way to clone the schema of an existing table.

To see a table’s Apache Iceberg format version, run SHOW TBLPROPERTIES org_catalog.test_db.test_table ('format-version').

Columns with no declared CRS

A geometry column created without a CRS renders as geom GEOMETRY, with no parentheses, in SHOW CREATE TABLE; DESCRIBE TABLE shows geometry for both kinds of column. Tables created before geometry(<srid>) was available have such columns, and so does a table created from the result of an ST_ function, as described in Creating a table from a query or a DataFrame. These columns keep working: reads, writes, and every ST_ function behave the same. The SRID of written values isn’t checked, Iceberg engines outside Wherobots read them as binary, and WherobotsDB skips data files for spatial filters on them only in a format-version 1 or 2 table. In a format-version 3 table, it reads every data file for those filters; see Spatial file skipping. Wherobots will upgrade the read-only public tables in the wherobots_open_data and wherobots_pro_data catalogs. To upgrade your own tables, see Upgrade Existing Geometry Tables.

Data in another CRS

Until more SRIDs can be declared, there are three ways to store data in another CRS, such as a Universal Transverse Mercator (UTM) zone. Spatial filters skip data files with all three:

Option	Keeps the coordinates	Records the CRS	Format version
Reproject with `ST_Transform`, and declare `geometry(4326)` or `geometry(3857)`	No	Yes, in the column	3
Declare `geometry(0)`, and write the values with SRID 0	Yes	No	3
Keep a column with no declared CRS, in a table created at format-version 2	Yes	Yes, in each value	2

Don’t keep a column with no declared CRS in a format-version 3 table. That is what CREATE TABLE ... AS SELECT and writeTo(...).create() give you by default when the geometry comes from an ST_ function, and spatial filters on that column read every data file. Nothing reports it: queries return correct results, only more slowly.

To use geometry(0), create the table, then set each value’s SRID to 0 as you write it:

CREATE TABLE org_catalog.test_db.parcels_utm (
    id bigint,
    geom geometry(0)
) USING iceberg;

INSERT INTO org_catalog.test_db.parcels_utm
SELECT id, ST_SetSRID(geom, 0) FROM org_catalog.test_db.parcels_source;

Writing a value that still carries its SRID fails with Cannot write geometry with SRID 32633 into a column with SRID 0. Build query shapes with SRID 0 too, for example with ST_GeomFromText and no SRID argument. Distances and areas come out in the CRS’s units, meters for UTM. The table doesn’t record the CRS, so keep track of it yourself, and label the values before you reproject them: ST_Transform(ST_SetSRID(geom, 32633), 'EPSG:4326'). To use a format-version 2 table, create it from a query or a DataFrame with the format version set. CREATE TABLE with a geometry column and no SRID is rejected, even at format-version 2.

CREATE TABLE org_catalog.test_db.parcels_utm
USING iceberg
TBLPROPERTIES ('format-version'='2')
AS SELECT id, ST_Transform(geom, 'EPSG:32633') AS geom FROM org_catalog.test_db.parcels_source;

From a DataFrame, use df.writeTo("org_catalog.test_db.parcels_utm").tableProperty("format-version", "2").create(). Later writes with INSERT INTO or writeTo(...).append() keep the table at format-version 2. Each value keeps its own SRID, which the column doesn’t check. A format-version 2 table can’t use format-version 3 features or take a geometry(<srid>) column. When your SRID can be declared, copy the table as described in Upgrade Existing Geometry Tables.

Writing data

INSERT INTO

Insert rows with INSERT INTO ... VALUES. Every value has to carry the SRID of the column:

Python
Scala
Java

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table
VALUES (1, 'a', ST_SetSRID(ST_GeomFromText('POINT (1 2)'), 4326)),
       (2, 'b', ST_SetSRID(ST_Point(2, 3), 4326))
""")

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table
VALUES (1, 'a', ST_SetSRID(ST_GeomFromText('POINT (1 2)'), 4326)),
       (2, 'b', ST_SetSRID(ST_Point(2, 3), 4326))
""")

sedona.sql(
        "INSERT INTO org_catalog.test_db.test_table " +
        "VALUES (1, 'a', ST_SetSRID(ST_GeomFromText('POINT (1 2)'), 4326)), " +
        "(2, 'b', ST_SetSRID(ST_Point(2, 3), 4326))"
);

Or insert the result of a query with INSERT INTO ... SELECT:

Python
Scala
Java

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table
SELECT 3, 'c', ST_SetSRID(ST_Point(3, 4), 4326)
""")

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table
SELECT 3, 'c', ST_SetSRID(ST_Point(3, 4), 4326)
""")

sedona.sql(
        "INSERT INTO org_catalog.test_db.test_table " +
        "SELECT 3, 'c', ST_SetSRID(ST_Point(3, 4), 4326)"
);

Matching the column CRS

Geometry constructors such as ST_GeomFromText and ST_Point return SRID 0, which is a distinct CRS rather than “unset”, so a geometry(4326) column rejects it. Which function you need depends on the SRID your values already carry:

Your values	Apply	Effect
SRID 0, from a constructor, from shapely, or from a file with no CRS metadata	`ST_SetSRID(geom, 4326)`	Labels the values. Coordinates are unchanged.
A different CRS, such as SRID 3857	`ST_Transform(geom, 'EPSG:3857', 'EPSG:4326')`	Reprojects the coordinates.
Already SRID 4326	Nothing	Insert as-is.
A `geometry(4326)` column on another table	Nothing	The CRS carries through the read.

ST_SetSRID relabels without touching the coordinates. Applying it to metre coordinates declares them as longitude and latitude, which puts the data in the wrong place. Use ST_Transform when the coordinates are in another CRS.

Error: Cannot write geometry with SRID ...

A value disagrees with the column CRS. The message names the fix:

java.lang.IllegalArgumentException: Cannot write geometry with SRID 0 into a column
with SRID 4326: an Iceberg geometry column stores one CRS for every value. Use
ST_SetSRID to label an unset SRID, or ST_Transform to reproject the values to SRID 4326.

Writing a DataFrame to a table

Append a DataFrame to a table with writeTo(...).append(). The same SRID rule applies to the DataFrame’s values:

Python
Scala
Java

from pyspark.sql.functions import expr

# Create a DataFrame with geometry data
df = sedona.createDataFrame([
    (1, "a", "POINT (1 2)"),
    (2, "b", "POINT (2 3)")
], ["id", "data", "geom"]) \
    .withColumn("geom", expr("ST_SetSRID(ST_GeomFromText(geom), 4326)"))

# Write the DataFrame to the table
df.writeTo("org_catalog.test_db.test_table").append()

// Create a DataFrame with geometry data
val df = Seq(
    (1, "a", "POINT (1 2)"),
    (2, "b", "POINT (2 3)")
).toDF("id", "data", "geom")
.withColumn("geom", expr("ST_SetSRID(ST_GeomFromText(geom), 4326)"))

// Write the DataFrame to the table
df.writeTo("org_catalog.test_db.test_table").append()

List<Row> data = Arrays.asList(
    RowFactory.create(1, "a", "POINT (1 2)"),
    RowFactory.create(2, "b", "POINT (2 3)")
);

List<StructField> fields = Arrays.asList(
    DataTypes.createStructField("id", DataTypes.IntegerType, true),
    DataTypes.createStructField("data", DataTypes.StringType, true),
    DataTypes.createStructField("geom", DataTypes.StringType, true)
);

StructType schema = DataTypes.createStructType(fields);

Dataset<Row> df = sedona.createDataFrame(data, schema)
    .withColumn("geom", expr("ST_SetSRID(ST_GeomFromText(geom), 4326)"));

df.writeTo("org_catalog.test_db.test_table").append();

This is equivalent to INSERT INTO ... SELECT.

Creating a table from a query or a DataFrame

CREATE TABLE ... AS SELECT, writeTo(...).create(), and writeTo(...).createOrReplace() create a table from the result of a query or a DataFrame. When the geometry comes from a geometry(<srid>) column read from a table, the new column has the same declared CRS. When the geometry is the result of an ST_ function, such as ST_Point, ST_SetSRID, or ST_Buffer, the new column has no declared CRS. In a Wherobots-managed catalog, such as org_catalog, the new table is format-version 3 unless you set 'format-version', and in a format-version 3 table, spatial filters on a column with no declared CRS read every data file. To declare the CRS of the new column, create the table with CREATE TABLE first and then write to it with INSERT INTO ... SELECT or writeTo(...).append(), as shown in Working with Geometry Data.

Updating data

UPDATE changes the rows that match a filter, which can be a spatial filter:

Python
Scala
Java

sedona.sql("UPDATE org_catalog.test_db.test_table SET data = 'd' WHERE id = 1")

# Use ST_Intersects to update rows that intersect with a polygon
sedona.sql("""
UPDATE org_catalog.test_db.test_table
SET data = 'd'
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

sedona.sql("UPDATE org_catalog.test_db.test_table SET data = 'd' WHERE id = 1")

// Use ST_Intersects to update rows that intersect with a polygon
sedona.sql("""
UPDATE org_catalog.test_db.test_table
SET data = 'd'
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

sedona.sql("UPDATE org_catalog.test_db.test_table SET data = 'd' WHERE id = 1");

// Use ST_Intersects to update rows that intersect with a polygon
sedona.sql(
        "UPDATE org_catalog.test_db.test_table " +
        "SET data = 'd' " +
        "WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))"
);

Deleting data

DELETE FROM removes the rows that match a filter, which can be a spatial filter:

Python
Scala
Java

sedona.sql("DELETE FROM org_catalog.test_db.test_table WHERE id = 1")

# Use ST_Intersects to delete rows that intersect with a polygon
sedona.sql("""
DELETE FROM org_catalog.test_db.test_table
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

sedona.sql("DELETE FROM org_catalog.test_db.test_table WHERE id = 1")

// Use ST_Intersects to delete rows that intersect with a polygon
sedona.sql("""
DELETE FROM org_catalog.test_db.test_table
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

sedona.sql("DELETE FROM org_catalog.test_db.test_table WHERE id = 1");

// Use ST_Intersects to delete rows that intersect with a polygon
sedona.sql(
        "DELETE FROM org_catalog.test_db.test_table " +
        "WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))"
);

Merging data using `MERGE INTO`

MERGE INTO has the same syntax as in Apache Iceberg; see Apache Iceberg - MERGE INTO. UPDATE, DELETE, and MERGE INTO work in both copy-on-write and merge-on-read mode, which the write.update.mode, write.delete.mode, and write.merge.mode table properties select.

Querying data

Load a table into a DataFrame with sedona.table(...):

Python
Scala
Java

df = sedona.table("org_catalog.test_db.test_table")

val df = sedona.table("org_catalog.test_db.test_table")

Dataset<Row> df = sedona.table("org_catalog.test_db.test_table");

You can set read options when you load the table. For example, with a smaller split size, the table is read into more partitions:

Python
Scala
Java

df = sedona.read.option('split-size', '1000').table("org_catalog.test_db.test_table")

val df = sedona.read.option("split-size", "1000").table("org_catalog.test_db.test_table")

Dataset<Row> df = sedona.read().option("split-size", "1000").table("org_catalog.test_db.test_table");

Run a spatial range query with Spatial SQL:

Python
Scala
Java

df = sedona.sql("""
SELECT * FROM org_catalog.test_db.test_table
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

val df = sedona.sql("""
SELECT * FROM org_catalog.test_db.test_table
WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))
""")

Dataset<Row> df = sedona.sql(
        "SELECT * FROM org_catalog.test_db.test_table " +
        "WHERE ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))"
);

Or with the DataFrame API:

Python
Scala
Java

df = sedona.table("org_catalog.test_db.test_table")\
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))")

val df = sedona.table("org_catalog.test_db.test_table")
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))")

Dataset<Row> df = sedona.table("org_catalog.test_db.test_table")
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))");

You can also load a table by naming the data source with .format("iceberg"). This loads an isolated table reference, which doesn’t refresh automatically when the table changes:

Python
Scala
Java

df = sedona.read.format("iceberg").load("org_catalog.test_db.test_table")\
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))")

val df = sedona.read.format("iceberg").load("org_catalog.test_db.test_table")
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))")

Dataset<Row> df = sedona.read().format("iceberg").load("org_catalog.test_db.test_table")
    .where("ST_Intersects(geom, ST_GeomFromText('POLYGON ((0 0, 0 2, 2 2, 2 0, 0 0))'))");

WherobotsDB skips the data files whose geometries can’t match a spatial filter, which makes selective spatial range queries fast. See Spatial Filter Push-down and Clustering for how it works and how to cluster a table so more files are skipped.

Working with geometry data

Every ST_ function in WherobotsDB works on a geometry column. For example, ST_Buffer creates a buffer around each geometry:

Python
Scala
Java

sedona.sql("SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table").show()

sedona.sql("SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table").show()

sedona.sql("SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table").show();

+----------------------+
|st_buffer(geom, 0.001)|
+----------------------+
|  POLYGON ((3.001 4...|
|  POLYGON ((1.001 2...|
|  POLYGON ((2.001 3...|
|  POLYGON ((1.001 2...|
|  POLYGON ((2.001 3...|
+----------------------+

To write the result back to a table, create the table with its geometry column first, then insert the result of the query. ST_Buffer keeps the SRID of its input, so the buffered values match a geometry(4326) column:

Python
Scala
Java

sedona.sql("""
CREATE TABLE org_catalog.test_db.test_table_buf (
    geom geometry(4326)
) USING iceberg""")

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table_buf
SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table
""")

sedona.sql("""
CREATE TABLE org_catalog.test_db.test_table_buf (
    geom geometry(4326)
) USING iceberg""")

sedona.sql("""
INSERT INTO org_catalog.test_db.test_table_buf
SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table
""")

sedona.sql(
        "CREATE TABLE org_catalog.test_db.test_table_buf (" +
        "geom geometry(4326)" +
        ") USING iceberg"
);

sedona.sql(
        "INSERT INTO org_catalog.test_db.test_table_buf " +
        "SELECT ST_Buffer(geom, 0.001) FROM org_catalog.test_db.test_table"
);

Or call writeTo(...).append() on the DataFrame the query returns:

Python
Scala
Java

sedona.sql("SELECT ST_Buffer(geom, 0.001) AS geom FROM org_catalog.test_db.test_table")\
    .writeTo("org_catalog.test_db.test_table_buf").append()

sedona.sql("SELECT ST_Buffer(geom, 0.001) AS geom FROM org_catalog.test_db.test_table")
    .writeTo("org_catalog.test_db.test_table_buf").append()

sedona.sql("SELECT ST_Buffer(geom, 0.001) AS geom FROM org_catalog.test_db.test_table")
    .writeTo("org_catalog.test_db.test_table_buf").append();

Geography columns

GEOGRAPHY(<srid>) interprets coordinates on a sphere rather than a plane, so ST_Distance, ST_Area, and ST_Length return spherical results in meters without projecting first.

sedona.sql("""
CREATE TABLE org_catalog.test_db.stations (
    id bigint,
    geog geography(4326)
) USING iceberg
""")

sedona.sql("""
INSERT INTO org_catalog.test_db.stations
SELECT 1L, ST_GeogFromWKT('POINT (-122.3937 37.7955)', 4326)
""")

A geography column rejects a value whose SRID is a different CRS, the same as a geometry column. Unlike a geometry column, it accepts a value with no SRID, such as the result of ST_GeogFromWKT without the SRID argument, and stores it under the column CRS. There is no cast from GEOMETRY to GEOGRAPHY; use ST_GeomToGeography to convert, and ST_GeogToGeometry in the other direction. See the geography function reference for the full list.

Iceberg Tables Overview

Format versions, spatial file skipping, and format-version 3 features.

Upgrade Existing Geometry Tables

Copy a table with no declared CRS, or geometry stored as WKB, WKT, or GeoJSON, into a table with a GEOMETRY(<srid>) column.

Spatial Filter Push-down and Clustering

Cluster or partition a table so WherobotsDB skips more data files for spatial filters.

Apache Iceberg documentation for Spark

Table maintenance, procedures, and the rest of the Iceberg Spark integration.

Start here

Havasu data catalogs

Vector data

Raster data

Iceberg tables

Maps and visualization

Geometry and Geography Tables

Readable as geometry by other engines

One CRS per column, enforced

Creating a table with a geometry column

Inspecting the schema

Columns with no declared CRS

Data in another CRS

Writing data

INSERT INTO

Matching the column CRS

Writing a DataFrame to a table

Creating a table from a query or a DataFrame

Updating data

Deleting data

Merging data using `MERGE INTO`

Querying data

Working with geometry data

Geography columns

Iceberg Tables Overview

Upgrade Existing Geometry Tables

Spatial Filter Push-down and Clustering

Apache Iceberg documentation for Spark

Readable as geometry by other engines

One CRS per column, enforced

Creating a table with a geometry column

Inspecting the schema

Columns with no declared CRS

Data in another CRS

Writing data

INSERT INTO

Matching the column CRS

Writing a DataFrame to a table

Creating a table from a query or a DataFrame

Updating data

Deleting data

Merging data using MERGE INTO

Querying data

Working with geometry data

Geography columns

Related pages

Iceberg Tables Overview

Upgrade Existing Geometry Tables

Spatial Filter Push-down and Clustering

Apache Iceberg documentation for Spark

Merging data using `MERGE INTO`