[SPARK-45745] Extremely slow execution of sum of columns in Spark 3.4.1

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 3.4.1
• Component/s: PySpark
• Assignee: Unassigned
• Reporter: Javier
• Created: 31/Oct/23 21:59
• Updated: 28/Aug/24 14:53
• Resolved: 28/Aug/24 14:53

Description

We are in the process of upgrading some pySpark jobs from Spark 3.1.2 to Spark 3.4.1 and some code that was running fine is now basically never ending even for small dataframes.

We have simplified the problematic piece of code and the minimum pySpark example below shows the issue:

n_cols = 50
data = [{f"col{i}": i for i in range(n_cols)} for _ in range(5)]
df_data = sql_context.createDataFrame(data)

df_data = df_data.withColumn(
    "col_sum", sum([F.col(f"col{i}") for i in range(n_cols)])
)
df_data.show(10, False) 

Basically, this code with Spark 3.1.2 runs fine but with 3.4.1 the computation time seems to explode when the value of `n_cols` is bigger than about 25 columns. A colleague suggested that it could be related to the limit of 22 elements in a tuple in Scala 2.13 (https://www.scala-lang.org/api/current/scala/Tuple22.html), since the 25 columns are suspiciously close to this. Is there any known defect in the logical plan optimization in 3.4.1? Or is this kind of operations (sum of multiple columns) supposed to be implemented differently?

Issue Links

• duplicates: SPARK-45071 Optimize the processing speed of `BinaryArithmetic#dataType` when processing multi-column data