[BEAM-12328] Conversion from Avro GenericRecords to Beam Rows takes too much time

• Type: Improvement
• Status: Open
• Priority: P3
• Resolution: Unresolved
• Component/s: dsl-sql, sdk-java-core
• Assignee: Unassigned
• Reporter: Ismaël Mejía
• Created: 12/May/21 15:39
• Updated: 04/Jun/22 20:06

Description

While running TPC-DS SQL query 3 on Dataflow with a input dataset of 1TB I noticed the performancew was not good compared with a native version implemented with Raw GenericRecords.

I enabled profiling and noticed that the performance of the Convert transform that converts Avro GenericRecords into Beam Rows is taking too much time, specially for those records that are relatively simple (not nested recordes / logical types or anything complex).

There is definitely something odd that can be improved in this transformation.