New Record State Extraction
Change event structure
Debezium generates data change events that have a complex structure. Each event consists of three parts:
Metadata, which includes but is not limited to:
The operation that made the change
Source information such as the names of the database and table where the change was made
Time stamp for when the change was made
Optional transaction information
Row data before the change
Row data after the change
For example, part of the structure of an UPDATE
change event looks like this:
{
"op": "u",
"source": {
...
},
"ts_ms" : "...",
"before" : {
"field1" : "oldvalue1",
"field2" : "oldvalue2"
},
"after" : {
"field1" : "newvalue1",
"field2" : "newvalue2"
}
}
More details about change event structure are provided in the documentation for each connector.
This complex format provides the most information about changes happening in the system. However, other connectors or other parts of the Kafka ecosystem usually expect the data in a simple format like this:
{
"field1" : "newvalue1",
"field2" : "newvalue2"
}
To provide the needed Kafka record format for consumers, configure the event flattening SMT.
Behavior
The event flattening SMT extracts the after
field from a Debezium change event in a Kafka record. The SMT replaces the original change event with only its after
field to create a simple Kafka record.
You can configure the event flattening SMT for a Debezium connector or for a sink connector that consumes messages emitted by a Debezium connector. The advantage of configuring event flattening for a sink connector is that records stored in Apache Kafka contain whole Debezium change events. The decision to apply the SMT to a source or sink connector depends on your particular use case.
You can configure the transformation to do any of the following:
Add metadata from the change event to the simplified Kafka record. The default behavior is that the SMT does not add metadata.
Keep Kafka records that contain change events for
DELETE
operations in the stream. The default behavior is that the SMT drops Kafka records forDELETE
operation change events because most consumers cannot yet handle them.
A database DELETE
operation causes Debezium to generate two Kafka records:
A record that contains
"op": "d",
thebefore
row data, and some other fields.A tombstone record that has the same key as the deleted row and a value of
null
. This record is a marker for Apache Kafka. It indicates that log compaction can remove all records that have this key.
Instead of dropping the record that contains the before
row data, you can configure the event flattening SMT to do one of the following:
Keep the record in the stream and edit it to have only the
"value": "null"
field.Keep the record in the stream and edit it to have a
value
field that contains the key/value pairs that were in thebefore
field with an added"__deleted": "true"
entry.
Similary, instead of dropping the tombstone record, you can configure the event flattening SMT to keep the tombstone record in the stream.
Configuration
Configure the Debezium event flattening SMT in a Kafka Connect source or sink connector by adding the SMT configuration details to your connector’s configuration. To obtain the default behavior, in a .properties
file, you would specify something like the following:
transforms=unwrap,...
transforms.unwrap.type=io.debezium.transforms.ExtractNewRecordState
As for any Kafka Connect connector configuration, you can set transforms=
to multiple, comma-separated, SMT aliases in the order in which you want Kafka Connect to apply the SMTs.
The following .properties
example sets several event flattening SMT options:
transforms=unwrap,...
transforms.unwrap.type=io.debezium.transforms.ExtractNewRecordState
transforms.unwrap.drop.tombstones=false
transforms.unwrap.delete.handling.mode=rewrite
transforms.unwrap.add.fields=table,lsn
drop.tombstones=false
Keeps tombstone records for DELETE
operations in the event stream.
delete.handling.mode=rewrite
For DELETE
operations, edits the Kafka record by flattening the value
field that was in the change event. The value
field directly contains the key/value pairs that were in the before
field. The SMT adds __deleted
and sets it to true
, for example:
"value": {
"pk": 2,
"cola": null,
"__deleted": "true"
}
add.fields=table,lsn
Adds change event metadata for the table
and lsn
fields to the simplified Kafka record.
Adding metadata
The event flattening SMT can add original, change event metadata to the simplified Kafka record. For example, you might want the simplified record’s header or value to contain any of the following:
The type of operation that made the change
The name of the database or table that was changed
Connector-specific fields such as the Postgres LSN field
For more information on what is available see the documentation for each connector.
To add metadata to the simplified Kafka record’s header, specify the add.header
option. To add metadata to the simplified Kafka record’s value, specify the add.fields
option. Each of these options takes a comma separated list of change event field names. Do not specify spaces. When there are duplicate field names, to add metadata for one of those fields, specify the struct as well as the field. For example:
transforms=unwrap,...
transforms.unwrap.type=io.debezium.transforms.ExtractNewRecordState
transforms.unwrap.add.fields=op,table,lsn,source.ts_ms
transforms.unwrap.add.headers=db
transforms.unwrap.delete.handling.mode=rewrite
With that configuration, a simplified Kafka record would contain something like the following:
{
...
"__op" : "c",
"__table": "MY_TABLE",
"__lsn": "123456789",
"__source_ts_ms" : "123456789",
...
}
Also, simplified Kafka records would have a __db
header.
In the simplified Kafka record, the SMT prefixes the metadata field names with a double underscore. When you specify a struct, the SMT also inserts an underscore between the struct name and the field name.
To add metadata to a simplified Kafka record that is for a DELETE
operation, you must also configure delete.handling.mode=rewrite
.
Configuration options
The following table describes the options that you can specify to configure the event flattening SMT.
Option | Default | Description |
---|---|---|
| Debezium generates a tombstone record for each | |
| Debezium generates a change event record for each | |
To use row data to determine the topic to route the record to, set this option to an | ||
(double-underscore) | Set this optional string to prefix a field. | |
Set this option to a comma-separated list, with no spaces, of metadata fields to add to the simplified Kafka record’s value. When there are duplicate field names, to add metadata for one of those fields, specify the struct as well as the field, for example | ||
(double-underscore) | Set this optional string to prefix a header. | |
Set this option to a comma-separated list, with no spaces, of metadata fields to add to the header of the simplified Kafka record. When there are duplicate field names, to add metadata for one of those fields, specify the struct as well as the field, for example |