[BEAM-12438] Sharing input to both a GBK and a CoGBK causes LP error w/schemas

• Type: Improvement
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Fix Version/s: Not applicable
• Component/s: sdk-go
• Assignee: Robert Burke
• Reporter: Robert Burke
• Created: 01/Jun/21 17:39
• Updated: 07/Jun/21 16:45
• Resolved: 07/Jun/21 16:45

Description

Constructing a pipeline with a PCol<KV> (call it A) used as an input to a GBK, using that output to generate an PCol<KV> (call it B) and using both A and B as inputs to a CoGBK causes the DoFn after the CoGBK to fail with an encoding error.

Essentially the Length Prefix coder applied to A isn't applied to the output of the CoGBK for the expand, causing the pipeline to fail. Only happens with Row encoded structs.

This is likely due to CoGBK being a synthetic construct on top of a GBK with a extra Inject and Expand operations, with the Inject missing the LP coding that the earlier GBK received.

Issue Links

• links to: GitHub Pull Request #14924