[SPARK-16449] unionAll raises "Task not serializable"

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Duplicate
• Affects Version/s: 1.6.1
• Component/s: PySpark

Environment

AWS EMR, Jupyter notebook

• Assignee: Unassigned
• Reporter: Jeff Levy
• Created: 08/Jul/16 17:35
• Updated: 12/Jul/16 12:25
• Resolved: 12/Jul/16 12:25

Description

Goal: Take the output from `describe` on a large DataFrame, then use a loop to calculate `skewness` and `kurtosis` from pyspark.sql.functions for each column, build them into a DataFrame of two rows, then use `unionAll` to merge them together.

Issue: Despite having the same column names, in the same order with the same dtypes, the `unionAll` fails with "Task not serializable". However, if I build two test rows using dummy data then `unionAll` works fine. Also, if I collect my results then turn them straight back into DataFrames, `unionAll` succeeds.

Step-by-step code and output with comments can be seen here: https://github.com/UrbanInstitute/pyspark-tutorials/blob/master/unionAll%20error.ipynb
The issue appears to be in the way the loop in code block 6 is building the rows before parallelizing, but the results look no different from the test rows that do work. I reproduced this on multiple datasets, so downloading the notebook and pointing it to any data of your own should replicate it.

Issue Links

• duplicates: SPARK-16173 Can't join describe() of DataFrame in Scala 2.10