[SPARK-29101] CSV datasource returns incorrect .count() from file with malformed records

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 2.4.0, 2.4.1, 2.4.2, 2.4.3, 2.4.4
• Fix Version/s: 2.4.5, 3.0.0
• Component/s: SQL
• Labels: correctness
• Assignee: Sandeep Katta
• Reporter: Stuart White
• Created: 16/Sep/19 19:38
• Updated: 12/Dec/22 18:10
• Resolved: 18/Sep/19 14:34

Description

Spark 2.4 introduced a change to the way csv files are read.  See Upgrading From Spark SQL 2.3 to 2.4 for more details.

In that document, it states: To restore the previous behavior, set spark.sql.csv.parser.columnPruning.enabled to false.

I am configuring Spark 2.4.4 as such, yet I'm still getting results inconsistent with pre-2.4.  For example:

Consider this file (fruit.csv).  Notice it contains a header record, 3 valid records, and one malformed record.

fruit,color,price,quantity
apple,red,1,3
banana,yellow,2,4
orange,orange,3,5
xxx

 
With Spark 2.1.1, if I call .count() on a DataFrame created from this file (using option DROPMALFORMED), "3" is returned.

(using Spark 2.1.1)
scala> spark.read.option("header", "true").option("mode", "DROPMALFORMED").csv("fruit.csv").count
19/09/16 14:28:01 WARN CSVRelation: Dropping malformed line: xxx
res1: Long = 3

With Spark 2.4.4, I set the "spark.sql.csv.parser.columnPruning.enabled" option to false to restore the pre-2.4 behavior for handling malformed records, then call .count() and "4" is returned.

(using spark 2.4.4)
scala> spark.conf.set("spark.sql.csv.parser.columnPruning.enabled", false)
scala> spark.read.option("header", "true").option("mode", "DROPMALFORMED").csv("fruit.csv").count
res1: Long = 4

So, using the spark.sql.csv.parser.columnPruning.enabled option did not actually restore previous behavior.

How can I, using Spark 2.4+, get a count of the records in a .csv which excludes malformed records?

Issue Links

• links to: GitHub Pull Request #25820
• links to: GitHub Pull Request #25843