Skip to content

Shuffle flushes can resend buffered tuples #8241

Description

@carloea2

Feature Summary

Hash-based and range-based shuffle partitioners emit pending tuples during control and state flushes but leave those tuples in their buffers. A later flush sends the same tuples again. Round-robin, broadcast, and one-to-one partitioners already drain their flushed buffers.

Before: first flush sends data and control; second flush sends the same data again

After: first flush sends data and control; second flush sends only control

Expected behavior:

A successful flush should drain only the emitted receiver buffer so tuples cannot be duplicated across control or state boundaries.

Reproduction evidence:

On 50321e403c82df299a13deb50a7f9849dd93bdba, add one tuple to a hash shuffle partitioner and flush the same receiver twice.

Observed result:

first=['list', 'EmbeddedControlMessage']
second=['list', 'EmbeddedControlMessage']

The same retained-buffer behavior occurs for range shuffle and for state flushes.

Version and commit evidence:

1.3.0-incubating-SNAPSHOT (main)

Commit Hash (Optional)

50321e403c82df299a13deb50a7f9849dd93bdba

Proposed Solution or Design

After: first flush sends data and control; second flush sends only control

Expected behavior:

Affected Area

Workflow Engine (Amber)

Activity

  1. added theissue type on Aug 31, 2026
  2. changed the issue type fromtoon Aug 31, 2026
  3. carloea2 commented on Sep 5, 2026

    @carloea2
    ContributorAuthor

    /take

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions