Kudu connector#

The Kudu connector allows querying, inserting and deleting data in Apache Kudu.

Requirements#

To connect to Kudu, you need:

  • Kudu version 1.13.0 or higher.

  • Network access from the Trino coordinator and workers to Kudu. Port 7051 is the default port.

Configuration#

To configure the Kudu connector, create a catalog properties file etc/catalog/kudu.properties with the following contents, replacing the properties as appropriate:

connector.name=kudu

## Defaults to NONE
kudu.authentication.type = NONE

## List of Kudu master addresses, at least one is needed (comma separated)
## Supported formats: example.com, example.com:7051, 192.0.2.1, 192.0.2.1:7051,
##                    [2001:db8::1], [2001:db8::1]:7051, 2001:db8::1
kudu.client.master-addresses=localhost

## Kudu does not support schemas, but the connector can emulate them optionally.
## By default, this feature is disabled, and all tables belong to the default schema.
## For more details see connector documentation.
#kudu.schema-emulation.enabled=false

## Prefix to use for schema emulation (only relevant if `kudu.schema-emulation.enabled=true`)
## The standard prefix is `presto::`. Empty prefix is also supported.
## For more details see connector documentation.
#kudu.schema-emulation.prefix=

###########################################
### Advanced Kudu Java client configuration
###########################################

## Default timeout used for administrative operations (e.g. createTable, deleteTable, etc.)
#kudu.client.default-admin-operation-timeout = 30s

## Default timeout used for user operations
#kudu.client.default-operation-timeout = 30s

## Disable Kudu client's collection of statistics.
#kudu.client.disable-statistics = false

## Assign Kudu splits to replica host if worker and kudu share the same cluster
#kudu.allow-local-scheduling = false

Kerberos support#

In order to connect to a kudu cluster that uses kerberos authentication, you need to configure the following kudu properties:

kudu.authentication.type = KERBEROS

## The kerberos client principal name
kudu.authentication.client.principal = clientprincipalname

## The path to the kerberos keytab file
## The configured client principal must exist in this keytab file
kudu.authentication.client.keytab = /path/to/keytab/file.keytab

## The path to the krb5.conf kerberos config file
kudu.authentication.config = /path/to/kerberos/krb5.conf

## Optional and defaults to "kudu"
## If kudu is running with a custom SPN this needs to be configured
kudu.authentication.server.principal.primary = kudu

Querying data#

Apache Kudu does not support schemas, i.e. namespaces for tables. The connector can optionally emulate schemas by table naming conventions.

Default behaviour (without schema emulation)#

The emulation of schemas is disabled by default. In this case all Kudu tables are part of the default schema.

For example, a Kudu table named orders can be queried in Trino with SELECT * FROM example.default.orders or simple with SELECT * FROM orders if catalog and schema are set to kudu and default respectively.

Table names can contain any characters in Kudu. In this case, use double quotes. E.g. To query a Kudu table named special.table! use SELECT * FROM example.default."special.table!".

Example#

  • Create a users table in the default schema:

    CREATE TABLE example.default.users (
      user_id int WITH (primary_key = true),
      first_name VARCHAR,
      last_name VARCHAR
    ) WITH (
      partition_by_hash_columns = ARRAY['user_id'],
      partition_by_hash_buckets = 2
    );
    

    On creating a Kudu table you must/can specify additional information about the primary key, encoding, and compression of columns and hash or range partitioning. For details see the Creating a table section.

  • Describe the table:

    DESCRIBE example.default.users;
    
       Column   |  Type   |                      Extra                      | Comment
    ------------+---------+-------------------------------------------------+---------
     user_id    | integer | primary_key, encoding=auto, compression=default |
     first_name | varchar | nullable, encoding=auto, compression=default    |
     last_name  | varchar | nullable, encoding=auto, compression=default    |
    (3 rows)
    
  • Insert some data:

    INSERT INTO example.default.users VALUES (1, 'Donald', 'Duck'), (2, 'Mickey', 'Mouse');
    
  • Select the inserted data:

    SELECT * FROM example.default.users;
    

Behavior with schema emulation#

If schema emulation has been enabled in the connector properties, i.e. etc/catalog/example.properties, tables are mapped to schemas depending on some conventions.

  • With kudu.schema-emulation.enabled=true and kudu.schema-emulation.prefix=, the mapping works like:

    Kudu table name

    Trino qualified name

    orders

    kudu.default.orders

    part1.part2

    kudu.part1.part2

    x.y.z

    kudu.x."y.z"

    As schemas are not directly supported by Kudu, a special table named $schemas is created for managing the schemas.

  • With kudu.schema-emulation.enabled=true and kudu.schema-emulation.prefix=presto::, the mapping works like:

    Kudu table name

    Trino qualified name

    orders

    kudu.default.orders

    part1.part2

    kudu.default."part1.part2"

    x.y.z

    kudu.default."x.y.z"

    presto::part1.part2

    kudu.part1.part2

    presto:x.y.z

    kudu.x."y.z"

    As schemas are not directly supported by Kudu, a special table named presto::$schemas is created for managing the schemas.

Type mapping#

Because Trino and Kudu each support types that the other does not, this connector modifies some types when reading or writing data. Data types may not map the same way in both directions between Trino and the data source. Refer to the following sections for type mapping in each direction.

Kudu type to Trino type mapping#

The connector maps Kudu types to the corresponding Trino types following this table:

Kudu type to Trino type mapping#

Kudu type

Trino type

BOOL

BOOLEAN

INT8

TINYINT

INT16

SMALLINT

INT32

INTEGER

INT64

BIGINT

FLOAT

REAL

DOUBLE

DOUBLE

DECIMAL(p,s)

DECIMAL(p,s)

STRING

VARCHAR

BINARY

VARBINARY

UNIXTIME_MICROS

TIMESTAMP(3)

No other types are supported.

Trino type to Kudu type mapping#

The connector maps Trino types to the corresponding Kudu types following this table:

Trino type to Kudu type mapping#

Trino type

Kudu type

Notes

BOOLEAN

BOOL

TINYINT

INT8

SMALLINT

INT16

INTEGER

INT32

BIGINT

INT64

REAL

FLOAT

DOUBLE

DOUBLE

DECIMAL(p,s)

DECIMAL(p,s)

Only supported for Kudu server >= 1.7.0

VARCHAR

STRING

The optional maximum length is lost

VARBINARY

BINARY

DATE

STRING

TIMESTAMP(3)

UNIXTIME_MICROS

µs resolution in Kudu column is reduced to ms resolution

No other types are supported.

SQL support#

The connector provides read and write access to data and metadata in Kudu. In addition to the globally available and read operation statements, the connector supports the following features:

Inserting into tables#

INSERT INTO ... values and INSERT INTO ... select behave like UPSERT.

SQL DELETE#

If a WHERE clause is specified, the DELETE operation only works if the predicate in the clause can be fully pushed down to the data source.

Creating schemas#

CREATE SCHEMA is only allowed if schema emulation is enabled. See the Behavior with schema emulation section.

Dropping schemas#

DROP SCHEMA is only allowed if schema emulation is enabled. See the Behavior with schema emulation section.

Creating a table#

On creating a Kudu table, you need to provide the columns and their types, of course, but Kudu needs information about partitioning and optionally for column encoding and compression.

Simple Example:

CREATE TABLE user_events (
  user_id INTEGER WITH (primary_key = true),
  event_name VARCHAR WITH (primary_key = true),
  message VARCHAR,
  details VARCHAR WITH (nullable = true, encoding = 'plain')
) WITH (
  partition_by_hash_columns = ARRAY['user_id'],
  partition_by_hash_buckets = 5,
  number_of_replicas = 3
);

The primary key consists of user_id and event_name. The table is partitioned into five partitions by hash values of the column user_id, and the number_of_replicas is explicitly set to 3.

The primary key columns must always be the first columns of the column list. All columns used in partitions must be part of the primary key.

The table property number_of_replicas is optional. It defines the number of tablet replicas, and must be an odd number. If it is not specified, the default replication factor from the Kudu master configuration is used.

Kudu supports two different kinds of partitioning: hash and range partitioning. Hash partitioning distributes rows by hash value into one of many buckets. Range partitions distributes rows using a totally-ordered range partition key. The concrete range partitions must be created explicitly. Kudu also supports multi-level partitioning. A table must have at least one partitioning, either hash or range. It can have at most one range partitioning, but multiple hash partitioning ‘levels’.

For more details see Partitioning design.

Column properties#

Besides column name and type, you can specify some more properties of a column.

Column property name

Type

Description

primary_key

BOOLEAN

If true, the column belongs to primary key columns. The Kudu primary key enforces a uniqueness constraint. Inserting a second row with the same primary key results in updating the existing row (‘UPSERT’). See also Primary Key Design in the Kudu documentation.

nullable

BOOLEAN

If true, the value can be null. Primary key columns must not be nullable.

encoding

VARCHAR

The column encoding can help to save storage space and to improve query performance. Kudu uses an auto encoding depending on the column type if not specified. Valid values are: 'auto', 'plain', 'bitshuffle', 'runlength', 'prefix', 'dictionary', 'group_varint'. See also Column encoding in the Kudu documentation.

compression

VARCHAR

The encoded column values can be compressed. Kudu uses a default compression if not specified. Valid values are: 'default', 'no', 'lz4', 'snappy', 'zlib'. See also Column compression in the Kudu documentation.

Example:

CREATE TABLE example_table (
  name VARCHAR WITH (primary_key = true, encoding = 'dictionary', compression = 'snappy'),
  index BIGINT WITH (nullable = true, encoding = 'runlength', compression = 'lz4'),
  comment VARCHAR WITH (nullable = true, encoding = 'plain', compression = 'default'),
   ...
) WITH (...);

Changing tables#

Adding a column to an existing table uses the SQL statement ALTER TABLE ... ADD COLUMN .... You can specify the same column properties as on creating a table.

Example:

ALTER TABLE example_table ADD COLUMN extraInfo VARCHAR WITH (nullable = true, encoding = 'plain')

See also Column properties.

ALTER TABLE ... RENAME COLUMN is only allowed if not part of a primary key.

ALTER TABLE ... DROP COLUMN is only allowed if not part of a primary key.

Procedures#

Partitioning design#

A table must have at least one partitioning (either hash or range). It can have at most one range partitioning, but multiple hash partitioning ‘levels’. For more details see Apache Kudu documentation: Partitioning.

If you create a Kudu table in Trino, the partitioning design is given by several table properties.

Hash partitioning#

You can provide the first hash partition group with two table properties:

The partition_by_hash_columns defines the column(s) belonging to the partition group and partition_by_hash_buckets the number of partitions to split the hash values range into. All partition columns must be part of the primary key.

Example:

CREATE TABLE example_table (
  col1 VARCHAR WITH (primary_key=true),
  col2 VARCHAR WITH (primary_key=true),
  ...
) WITH (
  partition_by_hash_columns = ARRAY['col1', 'col2'],
  partition_by_hash_buckets = 4
)

This defines a hash partitioning with the columns col1 and col2 distributed over 4 partitions.

To define two separate hash partition groups, also use the second pair of table properties named partition_by_second_hash_columns and partition_by_second_hash_buckets.

Example:

CREATE TABLE example_table (
  col1 VARCHAR WITH (primary_key=true),
  col2 VARCHAR WITH (primary_key=true),
  ...
) WITH (
  partition_by_hash_columns = ARRAY['col1'],
  partition_by_hash_buckets = 2,
  partition_by_second_hash_columns = ARRAY['col2'],
  partition_by_second_hash_buckets = 3
)

This defines a two-level hash partitioning, with the first hash partition group over the column col1 distributed over 2 buckets, and the second hash partition group over the column col2 distributed over 3 buckets. As a result you have table with 2 x 3 = 6 partitions.

Range partitioning#

You can provide at most one range partitioning in Apache Kudu. The columns are defined with the table property partition_by_range_columns. The ranges themselves are given either in the table property range_partitions on creating the table. Or alternatively, the procedures kudu.system.add_range_partition and kudu.system.drop_range_partition can be used to manage range partitions for existing tables. For both ways see below for more details.

Example:

CREATE TABLE events (
  rack VARCHAR WITH (primary_key=true),
  machine VARCHAR WITH (primary_key=true),
  event_time TIMESTAMP WITH (primary_key=true),
  ...
) WITH (
  partition_by_hash_columns = ARRAY['rack'],
  partition_by_hash_buckets = 2,
  partition_by_second_hash_columns = ARRAY['machine'],
  partition_by_second_hash_buckets = 3,
  partition_by_range_columns = ARRAY['event_time'],
  range_partitions = '[{"lower": null, "upper": "2018-01-01T00:00:00"},
                       {"lower": "2018-01-01T00:00:00", "upper": null}]'
)

This defines a tree-level partitioning with two hash partition groups and one range partitioning on the event_time column. Two range partitions are created with a split at “2018-01-01T00:00:00”.

Table property range_partitions#

With the range_partitions table property you specify the concrete range partitions to be created. The range partition definition itself must be given in the table property partition_design separately.

Example:

CREATE TABLE events (
  serialno VARCHAR WITH (primary_key = true),
  event_time TIMESTAMP WITH (primary_key = true),
  message VARCHAR
) WITH (
  partition_by_hash_columns = ARRAY['serialno'],
  partition_by_hash_buckets = 4,
  partition_by_range_columns = ARRAY['event_time'],
  range_partitions = '[{"lower": null, "upper": "2017-01-01T00:00:00"},
                       {"lower": "2017-01-01T00:00:00", "upper": "2017-07-01T00:00:00"},
                       {"lower": "2017-07-01T00:00:00", "upper": "2018-01-01T00:00:00"}]'
);

This creates a table with a hash partition on column serialno with 4 buckets and range partitioning on column event_time. Additionally, three range partitions are created:

  1. for all event_times before the year 2017, lower bound = null means it is unbound

  2. for the first half of the year 2017

  3. for the second half the year 2017

This means any attempt to add rows with event_time of year 2018 or greater fails, as no partition is defined. The next section shows how to define a new range partition for an existing table.

Managing range partitions#

For existing tables, there are procedures to add and drop a range partition.

  • adding a range partition

    CALL example.system.add_range_partition(<schema>, <table>, <range_partition_as_json_string>)
    
  • dropping a range partition

    CALL example.system.drop_range_partition(<schema>, <table>, <range_partition_as_json_string>)
    
    • <schema>: schema of the table

    • <table>: table names

    • <range_partition_as_json_string>: lower and upper bound of the range partition as JSON string in the form '{"lower": <value>, "upper": <value>}', or if the range partition has multiple columns: '{"lower": [<value_col1>,...], "upper": [<value_col1>,...]}'. The concrete literal for lower and upper bound values are depending on the column types.

      Examples:

      Trino data Type

      JSON string example

      BIGINT

      ‘{“lower”: 0, “upper”: 1000000}’

      SMALLINT

      ‘{“lower”: 10, “upper”: null}’

      VARCHAR

      ‘{“lower”: “A”, “upper”: “M”}’

      TIMESTAMP

      ‘{“lower”: “2018-02-01T00:00:00.000”, “upper”: “2018-02-01T12:00:00.000”}’

      BOOLEAN

      ‘{“lower”: false, “upper”: true}’

      VARBINARY

      values encoded as base64 strings

      To specified an unbounded bound, use the value null.

Example:

CALL example.system.add_range_partition('example_schema', 'events', '{"lower": "2018-01-01", "upper": "2018-06-01"}')

This adds a range partition for a table events in the schema example_schema with the lower bound 2018-01-01, more exactly 2018-01-01T00:00:00.000, and the upper bound 2018-07-01.

Use the SQL statement SHOW CREATE TABLE to query the existing range partitions (they are shown in the table property range_partitions).

Limitations#

  • Only lower case table and column names in Kudu are supported.