[FLINK-11186] Event-time balancing for multiple Kafka partitions

• Type: New Feature
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Component/s: API / DataStream, Connectors / Kafka
• Assignee: Unassigned
• Reporter: Tom Schamberger
• Created: 17/Dec/18 15:44
• Updated: 04/May/20 07:51
• Resolved: 04/May/20 07:51

Description

Currently, it is not possible with Flink to back-pressure individual Kafka partitions, which are faster in terms of event-time. This leads to unnecessary memory consumption and can lead to deadlocks in the case of back-pressure.

When multiple Kafka topics are consumed, succeeding event-time window operators have to wait until the last Kafka partition has produced a sufficient watermark to be triggered. If individual Kafka partitions differ in read performance or the event-time of messages within partitions is not monotonically distributed, this can lead to a situation, where 'fast' partitions (event-time makes fast progress) outperform slower partitions until back-pressuring prevents all partitions from being further consumed. This leads to a deadlock of the application.

I suggest, that windows should be able to back-pressure individual partitions, which progress faster in terms of event-time, so that slow partitions can keep up.

Issue Links

• duplicates: FLINK-12675 Event time synchronization in Kafka consumer