Skip to content

Native shuffle inaccurate estimate of builder memory allocation #1449

Closed
@andygrove

Description

@andygrove

Describe the bug

As demonstrated in unit tests added in #1440, we are allocating ~100kb for a batch when the actual memory used in less than 10kb.

assert_eq!(8376, batch.get_array_memory_size());
assert_eq!(106496, buffer.reservation.size());

Rather than use the current slot_size logic that produces an estimate based on schema, it would be more accurate to estimate memory size based on the memory used by the incoming batches.

Steps to reproduce

No response

Expected behavior

No response

Additional context

No response

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions