[
https://issues.apache.org/jira/browse/FLINK-20060?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17282128#comment-17282128
]
Blake Wilson edited comment on FLINK-20060 at 2/10/21, 2:29 AM:
----------------------------------------------------------------
Voicing support for addressing this. Deserialization with a Collector would be
a great usability enhancement for applications which encode a series of values
in a single Kinesis record. It would also save a great deal of serialization
since the current API requires adding a separate task to flatten kinesis
records:
KinesisConsumer ------ DataStream<List<T>> ----- FlattenItems ------
DataStream<T> ---->
The serialization cost of introducing this extra task is quite dramatic when
using a durable backend like the RocksDB one used in Kinesis Data Analytics
was (Author: blake wilson):
Voicing support for addressing this. Deserialization with a Collector would be
a great usability enhancement for applications which encode a series of values
in a single Kinesis record. It would also save a great deal of serialization
since the current API requires adding a separate task to flatten kinesis
records:
KinesisConsumer — DataStream<List<T>> FlattenItems ------ DataStream<T> ---->
The serialization cost of introducing this extra task is quite dramatic when
using a durable backend like the RocksDB one used in Kinesis Data Analytics
> Add a Collector to KinsesisDeserializationSchema
> ------------------------------------------------
>
> Key: FLINK-20060
> URL: https://issues.apache.org/jira/browse/FLINK-20060
> Project: Flink
> Issue Type: New Feature
> Components: Connectors / Kinesis
> Reporter: Timo Walther
> Priority: Major
>
> We did not add support for a collector in the KinesisDeserializationSchema.
> The problem with that connector lays in the threading model, where there is a
> pool of threads that read and deserialize records and then they handover the
> deserialized messages through a queue to the main thread. The problem is that
> we would need to create many temporary collections to put the deserialized
> records into the handover queue, which potentially would significantly affect
> performance, especially in the usual case of deserializing a single record
> from a single message.
> This means that we can currently not support the Debezium format for in the
> SQL connector if the Debezium record needs to emit 2 rows (UPDATE_BEFORE and
> UPDATE_AFTER).
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
