[SPARK-20012] spark.read.csv schemas effectively ignore headers

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Not A Problem
• Affects Version/s: 2.1.0
• Component/s: Input/Output

Environment

pyspark

• Assignee: Unassigned
• Reporter: david cottrell
• Created: 18/Mar/17 11:47
• Updated: 24/Mar/17 07:45
• Resolved: 24/Mar/17 07:45

Description

New to Spark, so please direct me elsewhere if there is another place for this kind of discussion.

To my understanding, schema are ordered named structures however it seems the names are not being used when reading files with headers.

I had a quick look at the DataFrameReader code and it seems like it might not be too hard to
a) let the schema set the "global" order of the columns
b) per file, map the columns by name to the schema ordering and apply the types on load.

A simple way of saying this is that the schema is an ordered dictionary and the files with headers only define dictionaries.

A typical example showing what I think are the implications of this problem:

In [248]: a = spark.read.csv('./data/test.csv.gz', header=True, inferSchema=True).toPandas()

In [249]: b = spark.read.csv('./data/0.csv.gz', header=True, inferSchema=True).toPandas()

In [250]: d = pd.concat([a, b])

In [251]: df = spark.read.csv('./data/{test,0}.csv.gz', header=True, inferSchema=True).toPandas()

In [252]: df[['b', 'c', 'd', 'e']] = df[['b', 'c', 'd', 'e']].astype(float)

In [253]: a
Out[253]:
      a         b         e         d         c
0  test -0.874197  0.168660 -0.948726  0.479723
1  test  1.124383  0.620870  0.159186  0.993676
2  test -1.429108 -0.048814 -0.057273 -1.331702

In [254]: b
Out[254]:
   a         b         c         d         e
0  0 -1.671828 -1.259530  0.905029  0.487244
1  0 -0.024553 -1.750904  0.004466  1.978049
2  0  1.686806  0.175431  0.677609 -0.851670

In [255]: d
Out[255]:
      a         b         c         d         e
0  test -0.874197  0.479723 -0.948726  0.168660
1  test  1.124383  0.993676  0.159186  0.620870
2  test -1.429108 -1.331702 -0.057273 -0.048814
0     0 -1.671828 -1.259530  0.905029  0.487244
1     0 -0.024553 -1.750904  0.004466  1.978049
2     0  1.686806  0.175431  0.677609 -0.851670

In [256]: df
Out[256]:
      a         b         c         d         e
0  test -0.874197  0.168660 -0.948726  0.479723
1  test  1.124383  0.620870  0.159186  0.993676
2  test -1.429108 -0.048814 -0.057273 -1.331702
3     0 -1.671828 -1.259530  0.905029  0.487244
4     0 -0.024553 -1.750904  0.004466  1.978049
5     0  1.686806  0.175431  0.677609 -0.851670

Example also posted here: http://stackoverflow.com/questions/42637497/pyspark-2-1-0-spark-read-csv-scrambles-columns