Skip to content

Schema Inference With ID

For supported serialization formats, ksqlDB can use schema inference to retrieve (read) and register (write) schemas as needed. If you specify a KEY_SCHEMA_ID or VALUE_SCHEMA_ID explicitly in the CREATE statements, ksqlDB retrieves and registers the schema specified by the ID from Schema Registry, and it also serializes data using exactly the same schema referred to by the ID. This can spare you from defining columns and data types manually and also make sure the data are serialized by the specified physical schema, which can be consumed in downstream systems. Before using schema inference with explicit IDs in ksqlDB, make sure that the Schema Registry is up and running and ksqlDB is configured to use it.

Here's what you can do with schema inference with IDs in ksqlDB:

  • Declare streams and tables on Kafka topics with supported key and value formats by using CREATE STREAM and CREATE TABLE statements with KEY_SCHEMA_ID or VALUE_SCHEMA_ID properties, without the need to declare the key and value columns.
  • Declare derived views with CREATE STREAM AS SELECT and CREATE TABLE AS SELECT statements with KEY_SCHEMA_ID or VALUE_SCHEMA_ID properties. The schema of the view is registered in Schema Registry automatically.
  • Serialize output data using the schema referred to by KEY_SCHEMA_ID or VALUE_SCHEMA_ID, instead of the logical data source schema stored in ksqlDB.

If you're declaring a stream or table with a key format that's different from its value format, and only one of the two formats supports schema inference, you can explicitly provide the columns for the format that doesn't support schema inference while still having ksqlDB load columns for the format that does support schema inference from Schema Registry. This is known as partial schema inference. To infer value columns for a keyless stream, set the key format to NONE.

Tables require a PRIMARY KEY, so you must supply one explicitly in your CREATE TABLE statement. KEY columns are optional for streams, so if you don't supply one, the stream is created without a key column.

Note

The following example statements show how to create streams and tables that have Avro-formatted data. If you want to use Protobuf or JSON-formatted data, substitute PROTOBUF or JSON_SR for AVRO in each statement.

Create a new stream or table

When KEY_SCHEMA_ID or VALUE_SCHEMA_ID is used in statements to create a stream or table, the schema fetched from Schema Registry is used to infer data source's columns and serialize output data. See Schema inference and data serialization for details about how columns are inferred and data are serialized.

Important

  • The schemas referred to by KEY_SCHEMA_ID and VALUE_SCHEMA_ID must be registered in Schema Registry. They can be under any subject but must match the formats defined by KEY_FORMAT and VALUE_FORMAT, respectively.
  • You can't define key or value columns in a statement if a corresponding KEY_SCHEMA_ID or VALUE_SCHEMA_ID is supplied.

Without a key column

The following statement shows how to create a new pageviews stream by reading from a Kafka topic that has Avro-formatted message values.

1
2
3
4
5
6
CREATE STREAM pageviews
  WITH (
    KAFKA_TOPIC='pageviews-avro-topic',
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1
  );

In this example, you don't need to define any columns in the CREATE statement. ksqlDB infers this information automatically from Schema Registry using the provided VALUE_SCHEMA_ID.

With a key column

The following statement shows how to create a new pageviews stream by reading from a Kafka topic that has Avro-formatted key and message values.

1
2
3
4
5
6
7
CREATE STREAM pageviews WITH (
    KAFKA_TOPIC='pageviews-avro-topic',
    KEY_FORMAT='AVRO',
    VALUE_FORMAT='AVRO',
    KEY_SCHEMA_ID=1,
    VALUE_SCHEMA_ID=2
  );

In this example, ksqlDB infers the key and value columns automatically from Schema Registry using the provided KEY_SCHEMA_ID and VALUE_SCHEMA_ID.

With partial schema inference

The following statement shows how to create a new pageviews table by reading from a Kafka topic that has Avro-formatted message values and a KAFKA-formatted INT primary key.

1
2
3
4
5
6
7
8
CREATE TABLE pageviews (
    pageId INT PRIMARY KEY
  ) WITH (
    KAFKA_TOPIC='pageviews-avro-topic',
    KEY_FORMAT='KAFKA',
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1
  );

In this example, only the key column is supplied in the CREATE statement. ksqlDB infers the value columns automatically from Schema Registry using the provided VALUE_SCHEMA_ID.

Declaring a derived view with schema ID.

The following statement shows how to create a materialized view derived from an existing source with the VALUE_SCHEMA_ID property. The schema referred to by VALUE_SCHEMA_ID is used to check column compatibility with output columns and serialize output data. For more information, see Schema inference and data serialization.

1
2
3
4
5
6
7
8
9
CREATE STREAM pageviews_new
  WITH (
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1
  ) AS 
  SELECT
    pageId,
    ts
  FROM pageviews

Important

The schema referred to by VALUE_SCHEMA_ID must be compatible with the logical schema defined by the SELECT clause. For more information, see Schema inference and data serialization.

Schema inference and data serialization

The schema in Schema Registry is a "physical schema", and the schema in ksqlDB is a "logical schema". The physical schema, not the logical schema, is registered under the subject <topic-name>-key or <topic-name>-value if corresponding KEY_SCHEMA_ID or VALUE_SCHEMA_ID values are provided.

Schema inference schema requirements

If WRAP_SINGLE_VALUE is set to true in the SQL statement, the physical schema is expected to be a struct type, and the field names are used as data source column names. Field types are inferred from corresponding column data types.

  • In AVRO, the struct type corresponds with the record type.
  • In PROTOBUF the struct type corresponds with the message type.
  • In JSON_SR, the struct type corresponds with the object type.

Note

In the following examples, the AVRO schema string in Schema Registry is a single-line raw string without newline characters (\n). The strings are shown as human-readable text for convenience.

For example, the following a physical schema is in AVRO format and is registered with Schema Registry with ID 1:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
{
  "schema": {
    "type": "record",
    "name": "PageViewValueSchema",
    "namespace": "io.confluent.ksql.avro_schemas",
    "fields": [
      {
        "name": "page_name",
        "type": "string",
        "default": "abc"
      },
      {
        "name": "ts",
        "type": "int",
        "default": 123
      }
    ]
  }
}

The following CREATE statement defines a stream on the pageviews topic and specifies the physical schema that has an ID of 1.

1
2
3
4
5
6
7
8
9
CREATE STREAM pageviews (
    pageId INT KEY
  ) WITH (
    KAFKA_TOPIC='pageviews-avro-topic',
    KEY_FORMAT='KAFKA',
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1,
    PARTITIONS=1
  );

The following output from the describe pageviews command shows the inferred logical schema for the pageviews stream:

1
2
3
4
5
6
7
8
9
ksql> describe pageviews;

Name                 : PAGEVIEWS
 Field     | Type
------------------------------------
 PAGEID    | INTEGER          (key)
 page_name | VARCHAR(STRING)
 ts        | INTEGER
------------------------------------

If WRAP_SINGLE_VALUE is false in the statement, and if KEY_SCHEMA_ID is set, ROWKEY is used as the key's column name.

If VALUE_SCHEMA_ID is set, ROWVAL is used as the value's column name. The physical schema is used as the column data type.

For example, the following physical schema is AVRO and is defined in Schema Registry with ID 2:

1
{"schema": "int"}

The following CREATE statement defines a table on the pageview-count topic and specifies the physical schema that has ID 2.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
CREATE TABLE pageview_count (
    pageId INT PRIMARY KEY
  ) WITH (
    KAFKA_TOPIC='pageview-count',
    KEY_FORMAT='KAFKA',
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=2,
    WRAP_SINGLE_VALUE=false,
    PARTITIONS=1
  );

The inferred logical schema for the pageview_count table is:

1
2
3
4
5
6
Name                 : PAGEVIEW_COUNT
 Field  | Type
-----------------------------------------
 PAGEID | INTEGER          (primary key)
 ROWVAL | INTEGER
-----------------------------------------

For more information about WRAP_SINGLE_VALUE, see Single Field (un)wrapping.

Schema Inference Type Handling

ksqlDB supports the null value for KEY and VALUE columns. If a field in the physical schema is a required type, it's translated to an optional type in the logical schema. This has subtle implications for data serialization which are explained in the following section.

Important

  • ksqlDB ignores unsupported types in the physical schema and continues translating supported types to the logical schema. You should verify that the logical schema is translated as expected.
  • During schema translation from a physical schema to a logical schema, struct type field names are used as column names in the logical schema. Field names are not translated to uppercase, in contrast with schema inference without a schema id, which does translate field names to uppercase.

Schema Compatibility Check for Derived View

You can use schema IDs when creating a materialized view, but instead of inferring the logical schema for the view, the schema is used to check compatibility against the query's projection and serialized output data. For compatibility checks, the inferred logical schema must be a superset of the query's projection schema, which means corresponding column names, types, and order must match. The inferred logical schema may have extra columns.

The following example creates the pageviews_new topic as the result of a SELECT query:

1
2
3
4
5
6
7
8
9
CREATE STREAM pageviews_new
  WITH (
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1
  ) AS 
  SELECT
    pageId,
    ts
  FROM pageviews

If the pageviews value column has the type ts INT, the logical schema of pageviews_new is decided by the projection in the query SELECT pageId, ts FROM pageviews. When VALUE_SCHEMA_ID is used, the inferred logical schema is checked against ts INT for compatibility.

The following example shows a compatible physical schema:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
{
  "schema": {
    "type": "record",
    "name": "PageViewNewSchema",
    "namespace": "io.confluent.ksql.avro_schemas",
    "fields": [
      {
        "name": "ts",
        "type": "int",
        "default": 123
      },
      {
        "name": "title",
        "type": "string",
        "default": "title"
      }
    ]
  }
}

In this AVRO schema, title is an extra field. Because the physical schema is used for data serialization, the title field with a default value appears in serialized data, even though the inserted value can never set the title field, because it's not in the logical schema defined by the SELECT clause of the query.

The following example shows an incompatible physical schema, which is incompatible because of the type mismatch for pageId.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
{
  "schema": {
    "type": "record",
    "name": "PageViewNewSchema",
    "namespace": "io.confluent.ksql.avro_schemas",
    "fields": [
      {
        "name": "pageId",
        "type": "string",
        "default": "id"
      },
      {
        "name": "ts",
        "type": "int",
        "default": 123
      }
    ]
  }
}

Data Serialization

When a schema ID is provided, and schema inference is successful, ksqlDB can create the data source. When writing to the data source, the physical schema inferred by the schema ID is used to serialize data, instead of the logical schema that's used in other cases. Because ksqlDB's logical schema accepts null values but the physical schema may not, serialization can fail even if the inserted value is valid for the logical schema.

The following example shows a physical schema that's defined in Schema Registry with ID 1. No default values are specified for the page_name and ts fields. 

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
{
  "schema": {
    "type": "record",
    "name": "PageViewValueSchema",
    "namespace": "io.confluent.ksql.avro_schemas",
    "fields": [
      {
        "name": "page_name",
        "type": "string"
      },
      {
        "name": "ts",
        "type": "int"
      }
    ]
  }
}

The following example creates a stream with schema ID 1:

1
2
3
4
5
6
7
8
9
CREATE STREAM pageviews (
    pageId INT KEY
  ) WITH (
    KAFKA_TOPIC='pageviews-avro-topic',
    KEY_FORMAT='KAFKA',
    VALUE_FORMAT='AVRO',
    VALUE_SCHEMA_ID=1,
    PARTITIONS=1
  );

ksqlDB infers the following schema for pageviews:

1
2
3
4
5
6
7
Name                 : PAGEVIEWS
 Field     | Type
------------------------------------
 PAGEID    | INTEGER          (key)
 page_name | VARCHAR(STRING)
 ts        | INTEGER
------------------------------------

If you insert values to pageviews with null values, ksqlDB returns an error:

1
insert into pageviews values (1, null, null);

1
Failed to insert values into 'PAGEVIEWS'. Could not serialize value: [ null | null ]. Error serializing message to topic: pageviews-avro-topic1. Invalid value: null used for required field: "page_name", schema type: STRING

This error occurs because page_name and ts are required fields without default values in the specified physical schema.

Important

ksqlDB doesn't check that null can be serialized in a physical schema that contains required fields. You must ensure that null can be handled properly, either by making physical schema fields optional or by using the IFNULL function to ensure that null is never inserted.

Important

If key_schema_id is used in table creation with windowed aggregation, the serialized key value also contain window information in addition to original key value.

Last update: 2022-08-08