[ 
https://issues.apache.org/jira/browse/SPARK-61423?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

L. C. Hsieh updated SPARK-61423:
--------------------------------
    Fix Version/s: connect-rust-4.3.0

> Reject an out-of-range decimal precision or scale when converting to Arrow 
> instead of wrapping it, in Spark Connect Rust
> ------------------------------------------------------------------------------------------------------------------------
>
>                 Key: SPARK-61423
>                 URL: https://issues.apache.org/jira/browse/SPARK-61423
>             Project: Spark
>          Issue Type: Bug
>          Components: Connect
>    Affects Versions: connect-rust-4.3.0
>            Reporter: L. C. Hsieh
>            Priority: Major
>             Fix For: connect-rust-4.3.0
>
>
> {{DataType::Decimal}} holds {{i32}} precision and scale and, as PySpark's 
> {{DecimalType}}, is not validated when it is built or parsed. The two places 
> in {{crates/spark-connect/src/session.rs}} that turn it into an Arrow 
> {{Decimal128}} ({{datatype_to_arrow}} and the decimal branch of 
> {{values_to_arrow}}) narrow the fields with {{as u8}} / {{as i8}}.
> So for a schema, or for a column with no non-null value, an out-of-range type 
> wraps silently instead of raising: {{decimal(257,0)}} becomes the valid 
> {{Decimal128(1, 0)}}, {{decimal(38,256)}} gets a scale of 0, and 
> {{decimal(300,0)}} becomes a precision of 44 and fails later with a message 
> about 44.
> PySpark does not validate {{DecimalType}} in the constructor, in a JSON 
> schema or in the proto conversion (the server rejects the type on analysis), 
> but its Arrow conversion ({{pa.decimal128(dt.precision, dt.scale)}}) does 
> raise for a precision outside 1 to 38.
> Use checked conversions and Arrow's own Decimal128 limits at the Arrow 
> boundary and return an error naming the type. Parsing stays unvalidated, as 
> in PySpark.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

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

Reply via email to