[FLINK-36072] Optimizing the Overhead of the Network Layer in Adaptive Execution Scenarios

• Type: Sub-task
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Component/s: Runtime / Network
• Labels: pull-request-available
• Assignee: Junrui Lee
• Reporter: Junrui Li
• Created: 15/Aug/24 08:22
• Updated: 13/Dec/24 02:14
• Resolved: 13/Dec/24 02:14

Description

In adaptive execution scenarios, hash edges may transition to broadcast edges; however, at that point, the upstream may have already produced data based on hashes.

For example, in the adaptive broadcast join case, this results in each downstream task needing to connect to all upstream tasks and create a partition reader for each subpartition, leading to significant overhead (O(N²)). To optimize the overhead of the network layer, we need to enable downstream tasks to consume all subpartitions from a specific Task Manager using a single channel and a single partition reader in such situations.

Issue Links

• links to: GitHub Pull Request #25738