Arthur created SPARK-60312:
------------------------------
Summary: VariantVal decoding rounds decimals with more than 28
digits under the default Python decimal context
Key: SPARK-60312
URL: https://issues.apache.org/jira/browse/SPARK-60312
Project: Spark
Issue Type: Bug
Components: PySpark
Affects Versions: 4.0.0
Reporter: Arthur
VariantUtils._get_decimal in python/pyspark/sql/variant_utils.py rebuilds a
variant decimal by multiplying the unscaled value with a power of ten using
decimal context arithmetic. Under the default Python context (precision 28) a
stored DECIMAL16 value with 38 significant digits loses digits in
VariantVal.toPython() and toJson(). With the Inexact trap enabled, decoding
raises instead.
Reproduction (PySpark 4.0.0 and master):
{code:python}
from pyspark.sql.types import VariantVal
value = VariantVal.parseJson("0.12345678901234567890123456789012345678")
print(value.toPython())
{code}
Actual: 0.1234567890123456789012345679
Expected: 0.12345678901234567890123456789012345678
Fix: https://github.com/apache/spark/pull/59201 rebuilds the value from the
coefficient digits and the stored scale, without context arithmetic, and adds a
regression test for DECIMAL4, DECIMAL8 and DECIMAL16 under precisions 1, 28 and
38.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
