[ 
https://issues.apache.org/jira/browse/SPARK-58069?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18095082#comment-18095082
 ] 

Eric Yang commented on SPARK-58069:
-----------------------------------

I'm working on a PR to fix it.

> approx_top_k silently returns incorrect results (internal collation sort keys 
> / garbage bytes) for collated string columns
> --------------------------------------------------------------------------------------------------------------------------
>
>                 Key: SPARK-58069
>                 URL: https://issues.apache.org/jira/browse/SPARK-58069
>             Project: Spark
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions: 5.0.0
>            Reporter: Eric Yang
>            Priority: Major
>
> *Description:*
> {{`approx_top_k`}} over a STRING column with a non-{{{}UTF8_BINARY{}}} 
> collation returns wrong item labels. The frequency counts are correct, but 
> the returned items are not actual input values: for ICU collations they are 
> raw sort-key bytes (invalid UTF-8, shown as U+FFFD replacement chars), and 
> for {{UTF8_LCASE}} they are the lower-cased form, which may never have 
> appeared in the data. Silent wrong-result, no error thrown.
> {code:sql}
> SELECT approx_top_k(c, 5, 100)
> FROM (
>   SELECT CAST(col AS STRING COLLATE UNICODE_CI) AS c
>   FROM VALUES ('HELLO'), ('HELLO'), ('HELLO'), ('hello'), ('world') AS t(col)
> );
> {code}
> Top item (count 4) by collation:
> ||collation||returned item||expected||
> |UTF8_BINARY|'HELLO' (correct)|'HELLO'|
> |UTF8_LCASE|'hello' (never appeared verbatim)|an actual value, e.g. 'HELLO'|
> |UNICODE_CI|bytes 39 33 41 41 47 01 09 00 (garbage)|an actual value, e.g. 
> 'HELLO'|
> |UNICODE|bytes 39 33 41 41 47 01 09 01 ef bf bd ... (garbage)|an actual 
> value, e.g. 'HELLO'|
> The equivalent exact query ({{{}... GROUP BY c ORDER BY count( *) DESC{}}}) 
> returns the correct actual value in every case.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to