[ARROW-16808] [C++] count_distinct aggregates incorrectly across row groups

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Duplicate
• Fix Version/s: 9.0.0

Environment

> arrow::arrow_info()
Arrow package version: 8.0.0.9000

Capabilities:
               
dataset TRUE
substrait FALSE
parquet TRUE
json TRUE
s3 TRUE
utf8proc TRUE
re2 TRUE
snappy TRUE
gzip TRUE
brotli TRUE
zstd TRUE
lz4 TRUE
lz4_frame TRUE
lzo FALSE
bz2 TRUE
jemalloc TRUE
mimalloc FALSE

Memory:
                   
Allocator jemalloc
Current 37.25 Kb
Max 925.42 Kb

Runtime:
                        
SIMD Level none
Detected SIMD Level none

Build:
                                                             
C++ Library Version 9.0.0-SNAPSHOT
C++ Compiler AppleClang
C++ Compiler Version 13.1.6.13160021
Git ID d9d78946607f36e25e9d812a5cc956bd00ab2bc9

• External issue URL: https://github.com/apache/arrow/issues/32139

Language

C++

• Assignee: Unassigned
• Reporter: Edward Visel
• Created: 10/Jun/22 17:00
• Updated: 11/Jan/23 11:46
• Resolved: 10/Jun/22 17:04

Description

When reading from parquet files with multiple row groups, count_distinct (wrapped by `n_distinct` in R) returns inaccurate and inconsistent results:

library(dplyr, warn.conflicts = FALSE)

path <- tempfile(fileext = '.parquet')
arrow::write_parquet(dplyr::starwars, path, chunk_size = 20L)

ds <- arrow::open_dataset(path)

ds %>% count(sex) %>% collect()
#> # A tibble: 5 × 2
#>   sex                n
#>   <chr>          <int>
#> 1 male              60
#> 2 none               6
#> 3 female            16
#> 4 hermaphroditic     1
#> 5 <NA>               4

ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    19
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    17
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    17
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    16
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    16
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    17
ds %>% summarise(n = n_distinct(sex)) %>% collect()
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1    17

# correct
ds %>% collect() %>% summarise(n = n_distinct(sex))
#> # A tibble: 1 × 1
#>       n
#>   <int>
#> 1     5

If the file is stored as a single row group, results are correct. When grouped, results are correct.

I can reproduce this in Python as well using the same file and pyarrow.compute.count_distinct:

import pyarrow as pa
import pyarrow.parquet as pq

pa.__version__
#> 8.0.0

starwars = pq.read_table('/var/folders/0j/zz6p_mjx2_b727p6xdpm5chc0000gn/T//Rtmp2wnWl5/file1744f6cc6cea8.parquet')

print(pa.compute.count_distinct(starwars.column('sex')).as_py())
#> 15
print(pa.compute.unique(starwars.column('sex')))
#> [
#>   "male",
#>   "none",
#>   "female",
#>   "hermaphroditic",
#>    null
#> ]

This seems likely to be the same problem in this StackOverflow question: https://stackoverflow.com/questions/72561901/how-do-i-compute-the-number-of-unique-values-in-a-pyarrow-array which is working from orc files.