Kafka Connector Data Schema

GridGain Kafka Connectors support data schemas. This enables numerous existing non-Ignite Sink Connectors to understand data injected with Ignite Source Connector and the Ignite Sink Connector to understand data injected by non-Ignite Source Connectors.

Ignite Type Support

Source and Sink Connectors work with Ignite data in Ignite Binary format.

The table below provides mappings between Java Ignite types and Kafka Schema types and Logical types

Java Type / Ignite Type Kafka Type Examples and Notes

Java Type / Ignite Type	Kafka Type	Examples and Notes
byte (Byte)	INT8
short (Short)	INT16
int (INT)	INT32
char (CHAR)	INT32
long (LONG)	INT64
float (FLOAT)	FLOAT32
double (DOUBLE)	FLOAT64
boolean (BOOL)	BOOLEAN
String (STR)	STRING
Class<?> (CLASS)	STRING	Same as `Class.getName()`.
UUID	STRING
enum (ENUM)	INT32	Same as `enumPolicy.ORDINAL`.
enum (ENUM)	STRING	Same as `enumPolicy.SHORT_NAME`,`enumPolicy.LONG_NAME` and `enumPolicy.FULL_NAME`.
java.util.Date (DATE)	TIMESTAMP
java.sql.Date (OBJ)	DATE	Kafka expects that time part (HH:mm:ss:ms) of Date instance is set to 0.
java.sql.Time (TIME)	TIME	Kafka expects that date part (DD:MM:YYYY) of Time instance is set to 0.
java.sql.Timestamp (TIMESTAMP)	TIMESTAMP
BigDecimal (DECIMAL)	DECIMAL
byte[] (BYTE_ARR)	BYTES
Map<K,V> (MAP)	MAP (KV_schema)	The schemas of maps are inferred from first `map_entry` or `array_element`. Heterogeneous collections/maps (e.g.`[1, «foo», 42.0]`) can trigger `DataException`.
Collection<V> (COL)	ARRAY (element_schema)	The schemas of collection(array) are inferred from first `map_entry` or `array_element`. Heterogeneous collections/maps (e.g.`[1, «foo», 42.0]`) can trigger `DataException`.
V[] (SHORT_ARR INT_ARR LONG_ARR FLOAT_ARR DOUBLE_ARR CHAR_ARR BOOLEAN_ARR DECIMAL_ARR STRING_ARR UUID_ARR DATE_ARR OBJ_ARR ENUM_ARR TIME_ARR DATE_ARR TIMESTAMP_ARR DECIMAL_ARR)	ARRAY (element_schema)	The schemas of collection(array) are inferred from first `map_entry` or `array_element`. Heterogeneous collections/maps (e.g.`[1, «foo», 42.0]`) can trigger `DataException`.
Pojo (OBJ)	STRUCT

byte (Byte)

INT8

short (Short)

INT16

int (INT)

INT32

char (CHAR)

INT32

long (LONG)

INT64

float (FLOAT)

FLOAT32

double (DOUBLE)

FLOAT64

boolean (BOOL)

BOOLEAN

String (STR)

STRING

Class<?> (CLASS)

STRING

Same as Class.getName().

UUID

STRING

enum (ENUM)

INT32

Same as enumPolicy.ORDINAL.

enum (ENUM)

STRING

Same as enumPolicy.SHORT_NAME,enumPolicy.LONG_NAME and enumPolicy.FULL_NAME.

java.util.Date (DATE)

TIMESTAMP

java.sql.Date (OBJ)

DATE

Kafka expects that time part (HH:mm:ss:ms) of Date instance is set to 0.

java.sql.Time (TIME)

TIME

Kafka expects that date part (DD:MM:YYYY) of Time instance is set to 0.

java.sql.Timestamp (TIMESTAMP)

TIMESTAMP

BigDecimal (DECIMAL)

DECIMAL

byte[] (BYTE_ARR)

BYTES

Map<K,V> (MAP)

MAP (KV_schema)

The schemas of maps are inferred from first map_entry or array_element. Heterogeneous collections/maps (e.g.[1, «foo», 42.0]) can trigger DataException.

Collection<V> (COL)

ARRAY (element_schema)

The schemas of collection(array) are inferred from first map_entry or array_element. Heterogeneous collections/maps (e.g.[1, «foo», 42.0]) can trigger DataException.

V[] (SHORT_ARR INT_ARR LONG_ARR FLOAT_ARR DOUBLE_ARR CHAR_ARR BOOLEAN_ARR DECIMAL_ARR STRING_ARR UUID_ARR DATE_ARR OBJ_ARR ENUM_ARR TIME_ARR DATE_ARR TIMESTAMP_ARR DECIMAL_ARR)

ARRAY (element_schema)

The schemas of collection(array) are inferred from first map_entry or array_element. Heterogeneous collections/maps (e.g.[1, «foo», 42.0]) can trigger DataException.

Pojo (OBJ)

STRUCT

Updates and Removals

By default, Source Connector does not process removed Ignite cache entries. Set the shallProcessRemovals configuration setting to true to make the Source Connector process removals. In this case Source Connector injects a record with null value into Kafka to indicate that the key was removed. Sink Connector removes keys with null values from the cache. Using null as a value to indicate a removed entry works because Ignite does not support null cache values.

For performance reasons, Sink Connector does not support existing cache entry update by default. Set shallProcessUpdates configuration setting to true to make the Sink Connector update existing entries.

Schema Migration

Schema migration is implicit for GridGain Connectors. Both the Source and Sink Connectors pull and push cache entries in cross-platform Ignite Binary format, which intrinsically supports changing schemas. Ignite cache keys and values are dynamic objects that could have a different set of fields.

For performance reasons, Source connector reuses key and values schemas that are created when the first cache entry is pulled. If it is expected that key or value schema could be changed over time, consider setting isSchemaDynamic to true.

Schemaless Operation

Source Connector does not generate schemas if the isSchemaless configuration setting is true.

Disabling schemas improves performance because the Connectors would not build schemas and would not convert keys and values into Kafka format. This comes at a cost of non-Ignite Sink converters unable to understand the data injected into Kafka in the Ignite Binary format.

Some examples when disabling Source schema makes sense:

You are ready to do some coding to extend a non-Ignite converter to process the Ignite Binary objects to achieve higher performance.
The Ignite Data Replication example does not need schemas since both the Source and Sink are GridGain connectors.

Schema Caching

GridGain connector has a feature for caching Kafka schemas of already resolved binary objects. This can be extremely useful in case if same cache stores objects with different binary types.

For example, let us assume that caches stores 2 types of objects: A and B. A stream of objects that were pulled by connector from the cache might look like this: A₁ B₁ A₂ A₃ A₄ B₂ B₃ …. With the default settings, the connector will infer initial Kafka schema from A₁ and try to apply it for all subsequent B_1..3 objects. This will lead to errors. As described above, you must set the isSchemaDynamic=true property to get correct results, this option forces the connector to perform schema resolution for each incoming object. Performing schema resolution each time is not a cheap process from performance perspective, so we can improve it by enabling schema caching with the enableSchemaCache=true property. After that, the connector will resolve kafka schemas only during first encounter of object with unknown type, for example:

Kafka schema for type A will be calculated during A₁ processing and reused for A₂ A₃ A₄ … A_x
Kafka schema for type B will be calculated during B₁ processing and reused for B₂ B₃ … B_x

When caching is enabled, the connector is still able to detect simple changes in binary types, like appearance of new field, but it cannot detect changes in inner composite types. For example let us assume that type A has field of type C.

public class A {
    C foo;
}

public class C {
    String f1;
}

If Kafka schema for type A was already resolved, but instance of class C, that is stored in field foo of new incoming object A_x, contains new field f2, connector will not detect it because binary schema of type A was not changed.

Nullability

Kafka Connect always requires that you provide a type schema for provided data. GridGain connector can infer proper kafka schema for fields of scalar types or other known logical types (date, time, decimal), but it is not always possible for complex objects or collections. For example, if you have an object obj of composite type T1 with field foo = null:

public class T1 {
    T2 foo = null;
}

While processing binary representation of obj, GridGain Connector will not be able to infer obj.foo type, because it does not have a binary instance of T2 that it can use for schema resolution. Similar case happens when some field is supposed to store a generic collection like List<Person>. If collection value is null, the collection is empty or collection consists of nulls only, we will not be able to infer the type of collection elements as well.

To work around this obstacle, GridGain Connector introduces a special fieldNullabilityPolicy option with the followingmodes:

LAZY mode, connector will simply skip nullable fields, so they will not be a part of either kafka schema or kafka connect record. This is same behavior that the connector had prior to the appearance of this policy (ver < 8.9.8).
EAGER mode will force the GridGain connector to do its best to infer Kafka schema of nullable field. In case if it is not possible to successfully complete schema resolution STRUCT(name=o.g.k.c.s.undef) placeholder will be used.

© 2024 GridGain Systems, Inc. All Rights Reserved. Privacy Policy | Legal Notices. GridGain® is a registered trademark of GridGain Systems, Inc.
Apache, Apache Ignite, the Apache feather and the Apache Ignite logo are either registered trademarks or trademarks of The Apache Software Foundation.

Last updated on Nov 26, 2024