asddfl created SPARK-60562:
------------------------------
Summary: pandas-on-Spark assign broadcasts a callable-returned
ndarray into every row
Key: SPARK-60562
URL: https://issues.apache.org/jira/browse/SPARK-60562
Project: Spark
Issue Type: Bug
Components: Pandas API on Spark, PySpark
Affects Versions: 4.2.0
Environment: Platform: Ubuntu 24.04 Linux-6.17.0-29-generic x86_64
Python: 3.11.17
pyspark 4.2.0
findspark 2.0.1
pandas 3.0.6
numpy 2.4.6
Reporter: asddfl
A callable passed to DataFrame.assign returns an ndarray from an existing
column. Pandas assigns its elements to the corresponding rows, producing scalar
c1 values 1 and 2. pandas-on-Spark accepts the callable but instead stores the
entire [1, 2] array in each c1 cell.
The operation succeeds without an exception while changing the shape and
meaning of each assigned value.
{code:python}
import pandas as pd
from pyspark.sql import SparkSession
import pyspark.pandas as ps
t0 = pd.DataFrame({'c0': [1, 2]})
print("Pandas:", flush=True)
result = t0.assign(c1=lambda t1: t1['c0'].to_numpy())
print(result, flush=True)
spark = (
SparkSession.builder
.master("local[2]")
.config("spark.sql.ansi.enabled", "false")
.config("spark.sql.shuffle.partitions", "2")
.getOrCreate()
)
ps.set_option("compute.ops_on_diff_frames", True)
t0 = ps.DataFrame({'c0': [1, 2]})
print("PySpark Pandas:", flush=True)
result = t0.assign(c1=lambda t1: t1['c0'].to_numpy())
print(result, flush=True)
spark.stop()
{code}
{code:bash}
Pandas:
c0 c1
0 1 1
1 2 2
PySpark Pandas:
c0 c1
0 1 [1, 2]
1 2 [1, 2]
{code}
*Expected behavior*
If this callable output is supported, return the Pandas-equivalent scalar c1
values 1 and 2. If position-based ndarray assignment is unsupported, reject the
evaluated callable output with a clear error. It should not silently turn the
whole array into a value repeated in every row.
*Actual behavior*
Pandas returns rows (c0=1, c1=1) and (c0=2, c1=2). pandas-on-Spark returns rows
(c0=1, c1=[1, 2]) and (c0=2, c1=[1, 2]).
*Reproduction notes*
Inspection of the installed DataFrame._assign implementation shows that it
validates the assignee before evaluating a callable. The resulting ndarray is
not checked again and reaches the literal-value construction path. This report
concerns inconsistent result-type validation and silent array broadcasting; it
does not assume that positional ndarray assignment is an API support promise.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
