[IMPALA-11059] Speed up zipping unnest by reading collection elements in batches

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Won't Fix
• Component/s: Backend
• Labels: performance
• Epic Link: Complex types performance improvements
• Epic Color: ghx-label-1

Language

C++

• Assignee: Daniel Becker
• Reporter: Daniel Becker
• Created: 13/Dec/21 16:55
• Updated: 29/Mar/23 08:52
• Resolved: 29/Mar/23 08:52

Description

Try to speed up zipping unnest by reading collection elements in batches.

Now we read from the collections row-wise, that is we read one element from each collection and we store them in the corresponding columns in the current row, then proceed to the next element in each collection for the next row etc. The proposal here is to fill the row batch column-wise, i.e. filling the column corresponding to the first collection, then the second collection etc.

Issue Links

• is related to: IMPALA-10920 UNNEST function for arrays in the select list