[ARROW-14630] [C++] DCHECK in GroupByNode when error encountered

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 6.0.0
• Fix Version/s: 6.0.1, 7.0.0
• Component/s: C++
• Labels: pull-request-available query-engine
• External issue URL: https://github.com/apache/arrow/issues/30174
• Assignee: David Li
• Reporter: Nicola Crane
• Created: 08/Nov/21 13:34
• Updated: 11/Jan/23 08:41
• Resolved: 08/Nov/21 21:01

Description

thisisnic found that this example crashes:

library(arrow)
library(dplyr)
write_dataset(group_by(iris, Species), "iris_data")
open_dataset("iris_data") %>%
  group_by(Species) %>%
  summarise(mean(Sepal.Length)) %>%
  collect()  

There are two bugs here:

• StopProducing is written in a way that causes a future to be finished twice, triggering a DCHECK.
• Consume() doesn't set the length of the key column batch, causing a spurious error because the group ID datum and the values datum will have different lengths.

Issue Links

• is related to: ARROW-14583 [R][C++] Crash when summarizing after filtering to no rows on partitioned data
• links to: GitHub Pull Request #11640