Skip to content

Serialization

The term serialization format refers to the manner in which a record's raw bytes are translated to and from information structures that ksqlDB can understand at runtime. ksqlDB offers several mechanisms for controlling serialization and deserialization.

The primary mechanism is by choosing the serialization format when you create a stream or table and specify FORMAT, KEY_FORMAT or VALUE_FORMAT in the WITH clause.

1
2
3
4
5
6
7
8
9
-- create table with JSON value format:
CREATE TABLE ORDERS (
    F0 INT PRIMARY KEY, 
    F1 STRING
  ) WITH (
    KEY_FORMAT='KAFKA',
    VALUE_FORMAT='JSON', 
    ...
  );

Serialization Formats

ksqlDB supports these serialization formats:

  • NONE used to indicate the data should not be deserialized.
  • DELIMITED supports comma separated values.
  • JSON and JSON_SR support JSON values, with and without schema registry integration
  • AVRO supports AVRO serialized values.
  • KAFKA supports primitives serialized using the standard Kafka serializers.
  • PROTOBUF and PROTOBUF_NOSR support Protocol Buffers, with and without schema registry integration

With the exception of the NONE format, all formats may be used as both key and value formats. See individual formats for details.

NONE

Feature Supported
As value format No
As key format Yes
Multi-Column Keys N/A
Schema Registry required No
Schema inference No
Single field wrapping No
Single field unwrapping No

The NONE format is a special marker format that is used to indicate ksqlDB should not attempt to deserialize that part of the Kafka record.

It's main use is as the KEY_FORMAT of key-less streams, especially where a default key format has been set, via ksql.persistence.default.format.key that supports Schema inference. If the key format was not overridden, the server would attempt to load the key schema from the Schema Registry. If the schema existed, the key columns would be inferred from the schema, which may not be the intent. If the schema did not exist, the statement would be rejected. In such situations, the key format can be set to NONE:

1
2
3
4
5
6
7
CREATE STREAM KEYLESS_STREAM (
    VAL STRING
  ) WITH (
    KEY_FORMAT='NONE',
    VALUE_FORMAT='JSON',
    KAFKA_TOPIC='foo'
  );

Any statement that sets the key format to NONE and has key columns defined, will result in an error.

If a CREATE TABLE AS or CREATE STREAM AS statement has a source with a key format of NONE, but the newly created table or stream has key columns, then you may either explicitly define the key format to use in the WITH clause, or the default key format, as set in ksql.persistence.default.format.key will be used.

Conversely, a CREATE STREAM AS statement that removes the key columns, i.e. via PARTITION BY null will automatically set the key format to NONE.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
-- keyless stream with NONE key format:
CREATE STREAM KEYLESS_STREAM (
    VAL STRING
  ) WITH (
    KEY_FORMAT='NONE',
    VALUE_FORMAT='JSON',
    KAFKA_TOPIC='foo'
  );

-- Table created from stream with explicit key format declared in WITH clause:
CREATE TABLE T WITH (KEY_FORMAT='KAFKA') AS 
  SELECT VAL, COUNT() FROM KEYLESS_STREAM
  GROUP BY VAL;

-- or, using the default key format set in the ksql.persistence.default.format.key config:
CREATE TABLE T AS 
  SELECT VAL, COUNT() FROM KEYLESS_STREAM
  GROUP BY VAL;

DELIMITED

Feature Supported
As value format Yes
As key format Yes
Multi-Column Keys Yes
Schema Registry required No
Schema inference No
Single field wrapping No
Single field unwrapping Yes

The DELIMITED format supports comma-separated values. You can use other delimiter characters by specifying the KEY_DELIMITER and/or VALUE_DELIMITER when you use FORMAT='DELIMITED' in a WITH clause. Only a single character is valid as a delimiter. The default is the comma character. For space- and tab-delimited values, use the special values SPACE or TAB, not an actual space or tab character.

The delimiter is a Unicode character, as defined in java.lang.Character. For example, the smiley-face character works:

1
CREATE STREAM delim_stream (f1 STRING, f2 STRING) with (KAFKA_TOPIC='delim', FORMAT='DELIMITED', VALUE_DELIMITER='☺', ...);

The serialized object should be a Kafka-serialized string, which will be split into columns.

For example, given a SQL statement such as:

1
CREATE STREAM x (ORGID BIGINT KEY, ID BIGINT KEY, NAME STRING, AGE INT) WITH (FORMAT='DELIMITED', ...);

ksqlDB splits a key of 120,21 and a value of bob,49 into the four fields (two keys and two values) with ORGID KEY of 120, ID KEY of 21, NAME of bob and AGE of 49.

This data format supports all SQL data types except ARRAY, MAP and STRUCT.

  • TIMESTAMP typed data is serialized as a long value indicating the Unix epoch time in milliseconds.
  • TIME typed data is serialized as an int value indicating the number of milliseconds since the beginning of the day.
  • DATE typed data is serialized as an int value indicating the number of days since the Unix epoch.
  • BYTES typed data is serialized as a Base64-encoded string value.

JSON

Feature Supported
As value format Yes
As key format JSON: Yes, JSON_SR: Yes
Multi-Column Keys Yes
Schema Registry required JSON: No, JSON_SR: Yes
Schema inference JSON: No, JSON_SR: Yes
Single field unwrapping Yes

There are two JSON formats, JSON and JSON_SR. Both support serializing and deserializing JSON data. The latter offers integration with the Schema Registry, registering and retrieving JSON schemas while the former does not. These two formats are not byte compatible (you cannot read data produced by one by the other).

The JSON formats supports all SQL data types. By itself, JSON doesn't support a map type, so ksqlDB serializes MAP types as JSON objects. For this reason, the JSON format supports only MAP objects that have STRING keys.

The serialized object should be a Kafka-serialized string that contains a valid JSON value. The format supports JSON objects and top-level primitives, arrays, and maps.

Important

If you want the sources that you create to store their schemas in Schema Registry, specify the JSON_SR format.

JSON Objects

Values that are JSON objects are probably the most common.

For example, given a SQL statement such as:

1
CREATE STREAM x (ID BIGINT, NAME STRING, AGE INT) WITH (VALUE_FORMAT='JSON', ...);

And a JSON value of:

1
2
3
4
5
{
  "id": 120,
  "name": "bob",
  "age": "49"
}

ksqlDB deserializes the JSON object's fields into the corresponding fields of the stream.

Top-level primitives, arrays and maps

The JSON format supports reading and writing top-level primitives, arrays and maps.

For example, given a SQL statement with only a single field in the value schema and the WRAP_SINGLE_VALUE property set to false:

1
CREATE STREAM x (ID BIGINT) WITH (VALUE_FORMAT='JSON', WRAP_SINGLE_VALUE=false, ...);

And a JSON value of:

1
10

ksqlDB can deserialize the values into the ID field of the stream.

When serializing data with a single field, ksqlDB can serialize the field as an anonymous value if the WRAP_SINGLE_VALUE is set to false, for example:

1
CREATE STREAM y WITH (WRAP_SINGLE_VALUE=false) AS SELECT id FROM x EMIT CHANGES;

Tip

Explicit wrapping and unwrapping is only supported for value columns. For more information, see Single field (un)wrapping.

Decimal Serialization

ksqlDB accepts decimals that are serialized either as numbers or the text representation of the base 10 equivalent. For example, ksqlDB can read JSON data from both formats below:

1
2
3
4
{
  "numericDecimal": 1.12345678912345,
  "stringDecimal": "1.12345678912345"
}

Decimals with specified precision and scale are serialized as JSON numbers. For example:

1
2
3
{
  "value": 1.12345678912345
}

Timestamp Serialization

Timestamps are serialized as numbers indicating the Unix epoch time in milliseconds. For example, a timestamp at 1970-01-01T00:00:00.001 is serialized as

1
2
3
{
  "value": 1
}

ksqlDb deserializes a number as a TIMESTAMP if it corresponds to a TIMESTAMP typed field in the stream.

Time Serialization

Times are serialized as numbers indicating the number of milliseconds since the beginning of the day. For example, 00:00:01 is serialized as

1
2
3
{
  "value": 1000
}

ksqlDb deserializes a number as a TIME if it corresponds to a TIME typed field in the stream.

Date Serialization

Dates are serialized as numbers indicating the number of days since the Unix epoch. For example, a timestamp at 1970-01-03 is serialized as

1
2
3
{
  "value": 2
}

ksqlDb deserializes a number as a DATE if it corresponds to a DATE typed field in the stream.

Bytes Serialization

Bytes are serialized as a Base64-encoded string value. For example, the byte array [61, 62, 63] is serialized as:

1
2
3
{
  "value": "YWJj"
}

ksqlDb deserializes a string as BYTES if it corresponds to a BYTES typed field in the stream.

Field Name Case Sensitivity

The format is case-insensitive when matching a SQL field name with a JSON document's property name. The first case-insensitive match is used.

Known Limitations

The following keywords aren't supported in the VALUE_SCHEMA_ID field for INSERT INTO and CREATE TABLE AS SELECT statements, and if present in your JSON schema will cause serialization errors. They are supported in CREATE statements.

  • additionalProperties: false
  • enum
  • required

Avro

Feature Supported
As value format Yes
As key format Yes
Multi-Column Keys Yes
Schema Registry required Yes
Schema inference Yes
Single field wrapping Yes
Single field unwrapping Yes

The AVRO format supports Avro binary serialization of all SQL data types, including records and top-level primitives, arrays, and maps.

Note

ksqlDB doesn't support creating streams or tables from a topic that has a recursive Avro schema.

The format requires ksqlDB to be configured to store and retrieve the Avro schemas from the Confluent Schema Registry. For more information, see Configure ksqlDB for Avro, Protobuf, and JSON schemas.

Avro Records

Avro records can be deserialized into matching ksqlDB schemas.

For example, given a SQL statement such as:

1
CREATE STREAM x (ID BIGINT, NAME STRING, AGE INT, TIME TIMESTAMP) WITH (VALUE_FORMAT='AVRO', ...);

And an Avro record serialized with the schema:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
{
  "type": "record",
  "namespace": "com.acme",
  "name": "UserDetails",
  "fields": [
    { "name": "id", "type": "long" },
    { "name": "name", "type": "string" },
    { "name": "age", "type": "int" },
    { "name": "time", "type": "long", "logicalType": "timestamp-millis"}
  ]
}

ksqlDB deserializes the Avro record's fields into the corresponding fields of the stream.

Important

By default, ksqlDB-registered schemas have the same name (KsqlDataSourceSchema) and the same namespace (io.confluent.ksql.avro_schemas). You can override this behavior by providing a VALUE_AVRO_SCHEMA_FULL_NAME property in the WITH clause, where you set the VALUE_FORMAT to 'AVRO'. As the name suggests, this property overrides the default name/namespace with the provided one. For example, com.mycompany.MySchema registers a schema with the MySchema name and the com.mycompany namespace.

Top-level primitives, arrays and maps

The Avro format supports reading and writing top-level primitives, arrays and maps.

For example, given a SQL statement with only a single field in the value schema and the WRAP_SINGLE_VALUE property set to false:

1
CREATE STREAM x (ID BIGINT) WITH (VALUE_FORMAT='AVRO', WRAP_SINGLE_VALUE=false, ...);

And an Avro value serialized with the schema:

1
2
3
{
  "type": "long"
}

ksqlDB can deserialize the values into the ID field of the stream.

When serializing data with a single field, ksqlDB can serialize the field as an anonymous value if the WRAP_SINGLE_VALUE is set to false, for example:

1
CREATE STREAM y WITH (WRAP_SINGLE_VALUE=false) AS SELECT id FROM x EMIT CHANGES;

Tip

Explicit wrapping and unwrapping is only supported for value columns. For more information, see Single field (un)wrapping.

Field Name Case Sensitivity

The format is case-insensitive when matching a SQL field name with an Avro record's field name. The first case-insensitive match is used.

KAFKA

Feature Supported
As value format Yes
As key format Yes
Multi-Column Keys No
Schema Registry required No
Schema inference No
Single field wrapping No
Single field unwrapping Yes

The KAFKA format supports INT, BIGINT, DOUBLE and STRING primitives that have been serialized using Kafka's standard set of serializers.

The format is designed primarily to support primitive message keys. It can be used as a value format, though certain operations aren't supported when this is the case.

Unlike some other formats, the KAFKA format does not perform any type coercion, so it's important to correctly match the field type to the underlying serialized form to avoid deserialization errors.

The table below details the SQL types the format supports, including details of the associated Kafka Java Serializer, Deserializer and Connect Converter classes you would need to use to write the key to Kafka, read the key from Kafka, or use to configure Apache Connect to work with the KAFKA format, respectively.

SQL Field Type Kafka Type Kafka Serializer Kafka Deserializer Connect Converter
INT / INTEGER A 32-bit signed integer org.apache.kafka.common.serialization.IntegerSerializer org.apache.kafka.common.serialization.IntegerDeserializer org.apache.kafka.connect.converters.IntegerConverter
BIGINT A 64-bit signed integer org.apache.kafka.common.serialization.LongSerializer org.apache.kafka.common.serialization.LongDeserializer org.apache.kafka.connect.converters.LongConverter
DOUBLE A 64-bit floating point number org.apache.kafka.common.serialization.DoubleSerializer org.apache.kafka.common.serialization.DoubleDeserializer org.apache.kafka.connect.converters.DoubleConverter
STRING / VARCHAR A UTF-8 encoded text string org.apache.kafka.common.serialization.StringSerializer org.apache.kafka.common.serialization.StringDeserializer org.apache.kafka.connect.storage.StringConverter

Because the format supports only primitive types, you can only use it when the schema contains a single field.

For example, if your Kafka messages have a long key, you can make them available to ksqlDB by using a statement like:

1
CREATE STREAM USERS (ID BIGINT KEY, NAME STRING) WITH (VALUE_FORMAT='JSON', ...);

If you integrate ksqlDB with Confluent Schema Registry, and your ksqlDB application uses a compatible value format (Avro, JSON_SR, or Protobuf), you can just supply the key column, and ksqlDB loads the value columns from Schema Registry:

1
CREATE STREAM USERS (ID BIGINT KEY) WITH (VALUE_FORMAT='JSON_SR', ...);

The key column must be supplied, because ksqlDB supports only keys in KAFKA format.

Protobuf

Feature Supported
As value format Yes
As key format Yes
Multi-Column Keys Yes
Schema Registry required PROTOBUF: Yes, PROTOBUF_NOSR: No
Schema inference PROTOBUF: Yes, PROTOBUF_NOSR: No
Single field wrapping Yes
Single field unwrapping No

Protobuf handles null values differently than AVRO and JSON. Protobuf doesn't have the concept of a null value, so the conversion between PROTOBUF and Java (Kafka Connect) objects is undefined. Usually, Protobuf resolves a "missing field" to the default value of its type.

  • String: the default value is the empty string.
  • Byte: the default value is empty bytes.
  • Bool: the default value is false.
  • Numeric type: the default value is zero.
  • Enum: the default value is the first defined enum value, which must be zero.
  • Message field: the field is not set. Its exact value is language-dependent. See the generated code guide for details.

To enable alternative representations for null values in protobuf, protobuf-specific properties can be passed to CREATE statements. For example, the following CREATE statement will create a protobuf schema that wraps all primitive types into the corresponding standard wrappers (e.g. google.protobuf.StringValue for string).

1
CREATE STREAM USERS (ID STRING KEY, i INTEGER, s STRING) WITH (VALUE_FORMAT='PROTOBUF', VALUE_PROTOBUF_NULLABLE_REPRESENTATION='WRAPPER');

This way, null can be distinguished from default values. Similarly, when VALUE_PROTOBUF_NULLABLE_REPRESENTATION is set to OPTIONAL, all fields in protobuf will be declared optional, also allowing null primitive fields to be distinguished from default values.

The same property values can be used with the KEY_PROTOBUF_NULLABLE_REPRESENTATION property to customize the protobuf serialization of the key.

Single field (un)wrapping

(de)serialization of single keys

ksqlDB assumes that any single key is unwrapped, which means that it's not contained in an outer record or object. Conversely, ksqlDB assumes that any key with multiple columns (for example, CREATE STREAM x (K1 INT KEY, K2 INT KEY, C1 INT)) is wrapped, which means that it is a record with each column as a field within the key.

To declare a single-column key that's wrapped, specify a STRUCT type with a single column. For example, K STRUCT<F1 INT> KEY. See the next two sections on single values for more information about wrapped and unwrapped data.

Controlling deserializing of single values

When ksqlDB deserializes a Kafka message into a row, the key is deserialized into the key field, and the message's value is deserialized into the value fields.

By default, ksqlDB expects any value with a single-field schema to have been serialized as a named field within a record. However, this is not always the case. ksqlDB also supports reading data that has been serialized as an anonymous value.

For example, a value with multiple fields might look like the following in JSON:

1
2
3
4
{
   "id": 134,
   "name": "John"
}

If the value only had the id field, ksqlDB would still expect the value to be serialized as a named field, for example:

1
2
3
{
   "id": 134
}

If your data contains only a single field, and that field is not wrapped within a JSON object, or an Avro record is using the AVRO format, then you can use the WRAP_SINGLE_VALUE property in the WITH clause of your CREATE TABLE or CREATE STREAM statements. Setting the property to false tells ksqlDB that the value isn't wrapped, so the example above would be a JSON number:

1
134

For example, the following creates a table where the values in the underlying topic have been serialized as an anonymous JSON number:

1
2
3
4
5
6
7
CREATE TABLE TRADES (
    ID INT PRIMARY KEY, 
    VALUE INT
  ) WITH (
    WRAP_SINGLE_VALUE=false, 
    ...
  );

If a statement doesn't set the value wrapping explicitly, ksqlDB uses the system default, which is defined by ksql.persistence.wrap.single.values. You can change the system default, if the format supports it. For more information, see ksql.persistence.wrap.single.values.

Important

ksqlDB treats null keys and values as a special case. We recommend avoiding unwrapped single-field schemas if the field can have a null value.

A null value in a table's topic is treated as a tombstone, which indicates that a row has been removed. If a table's source topic has an unwrapped single-field key schema and the value is null, it's treated as a tombstone, resulting in any previous value for the key being removed from the table.

A null key or value in a stream's topic is ignored when the stream is part of a join. A null value in a table's topic is treated as a tombstone, and a null key is ignored when the table is part of a join.

When you have an unwrapped single-field schema, ensure that any null key or value has the desired result.

Controlling serialization of single values

When ksqlDB serializes a row into a Kafka message, the key field is serialized into the message's key, and any value fields are serialized into the message's value.

By default, if the value has only a single field, ksqlDB serializes the single field as a named field within a record. However, this doesn't always match the requirements of downstream consumers, so ksqlDB allows the value to be serialized as an anonymous value.

For example, consider the statements:

1
2
CREATE STREAM x (f0 INT, f1 STRING) WITH (VALUE_FORMAT='JSON', ...);
CREATE STREAM y AS SELECT f0 FROM x EMIT CHANGES;

The second statement defines a stream with only a single field in the value, named f0.

By default, when ksqlDB writes out the result to Kafka, it persists the single field as a named field within a JSON object, or an Avro record if using the AVRO format:

1
2
3
{
   "F0": 10
}

If you require the value to be serialized as an anonymous value, for example:

1
10

Then you can use the WRAP_SINGLE_VALUE property in your statement. For example,

1
CREATE STREAM y WITH(WRAP_SINGLE_VALUE=false) AS SELECT f0 FROM x EMIT CHANGES;

If a statement doesn't set the value wrapping explicitly, ksqlDB uses the system default, defined by ksql.persistence.wrap.single.values, if the format supports it. You can change the system default. For more information, see ksql.persistence.wrap.single.values.

Important

ksqlDB treats null keys and values as a special case. We recommended avoiding unwrapped single-field schemas if the field can have a null value.

A null value in a table's topic is treated as a tombstone, which indicates that a row has been removed. If a table's source topic has an unwrapped single-field key schema and the value is null, it's treated as a tombstone, resulting in any previous value for the key being removed from the table.

A null key or value in a stream's topic is ignored when the stream is part of a join. A null value in a table's topic is treated as a tombstone, and a null key is ignored when the table is part of a join.

When you have an unwrapped single-field schema, ensure that any null key or value has the desired result.

Single-field serialization examples

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
-- Assuming system configuration is at the default:
--  ksql.persistence.wrap.single.values=true

-- creates a stream, picking up the system default of wrapping values.
-- the serialized value is expected to be wrapped.
-- if the serialized forms do not match the expected wrapping it will result in a deserialization error.
CREATE STREAM IMPLICIT_SOURCE (NAME STRING) WITH (...);

-- override 'ksql.persistence.wrap.single.values' to false
-- the serialized value is expected to not be unwrapped.
CREATE STREAM EXPLICIT_SOURCE (ID INT) WITH (WRAP_SINGLE_VALUE=false, ...);

-- results in an error as the value schema is multi-field
CREATE STREAM BAD_SOURCE (ID INT, NAME STRING) WITH (WRAP_SINGLE_VALUE=false, ...);

-- creates a stream, picking up the system default of wrapping values.
-- the serialized values in the sink topic will be wrapped.
CREATE STREAM IMPLICIT_SINK AS SELECT ID FROM S EMIT CHANGES;

-- override 'ksql.persistence.wrap.single.values' to false
-- the serialized values will not be wrapped.
CREATE STREAM EXPLICIT_SINK WITH(WRAP_SINGLE_VALUE=false) AS SELECT ID FROM S EMIT CHANGES;

-- results in an error as the value schema is multi-field
CREATE STREAM BAD_SINK WITH(WRAP_SINGLE_VALUE=true) AS SELECT ID, COST FROM S EMIT CHANGES;

Defining schemas for Schema Registry

Schema Registry enables both schema inference and defining schemas explicitly. You can use the identifiers for specific schemas in your CREATE STREAM and CREATE TABLE statements with KEY_SCHEMA_ID and VALUE_SCHEMA_ID properties. For more information, see Schema Inference With ID.

Some serialization formats require special syntax. For example, in JSON, the following message maps to a STRUCT<> with any field that contains an INT type.

1
2
3
4
5
6
"metadata": {
  "type": "object",
  "additionalProperties": {
     "type": "int"
  }
}

But JSON_SR doesn't support this case. In JSON_SR, the schema must be adapted to make it a MAP<STRING, INT> by using the connect.type config:

1
2
3
4
5
6
7
"metadata": {
  "connect.type": "map",
  "type": "object",
  "additionalProperties": {
     "type": "int"
  }
}

Suggested Reading


Last update: 2023-08-22