[
https://issues.apache.org/jira/browse/SPARK-12467?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15994833#comment-15994833
]
John Berryman commented on SPARK-12467:
---------------------------------------
I believe there is still something here that needs to be fixed.
Consider the following code:
{code}
from pyspark.sql import Row
from datetime import datetime
rows = [
dict(number=1, letters='real1', some_date=datetime(2017,12,1,3,15)),
dict(number=2, letters='real2', some_date=datetime(2017,12,2,3,15)),
dict(number=3, letters='real3', some_date=datetime(2017,12,3,3,15)),
]
rows_rdd = spark.sparkContext.parallelize(rows).map(lambda r: Row(**r))
df = spark.createDataFrame(rows_rdd)
spark.sql('CREATE DATABASE test_trash')
df.write.mode(saveMode='overwrite').saveAsTable('test_trash.thingy')
schema = spark.sql('SELECT letters, number, some_date FROM
test_trash.thingy').schema # A-works
# schema = spark.sql('SELECT some_date, number, letters FROM
test_trash.thingy').schema # B-fails
schema = spark.sql('SELECT number, letters, some_date FROM
test_trash.thingy').schema # C-works
rows_rdd = spark.sparkContext.parallelize(rows).map(lambda r: Row(**r))
df = spark.createDataFrame(rows_rdd, schema)
df.count()
{code}
If I uncomment line #A it works, line #B fails, and line #C works. The only
difference is the ordering of the named fields. The behavior is inconsistent.
Also, ``Row`` objects have named fields so why should there be any dependence
upon ordering at all? Also, the errors don't really convey the problem
``AttributeError: 'int' object has no attribute 'tzinfo'`` - the error should
be about some explicit schema mismatch (though I contend that this isn't really
a mismatch at all; the above lines should all work).
> Get rid of sorting in Row's constructor in pyspark
> --------------------------------------------------
>
> Key: SPARK-12467
> URL: https://issues.apache.org/jira/browse/SPARK-12467
> Project: Spark
> Issue Type: Bug
> Components: PySpark, SQL
> Affects Versions: 1.5.2
> Reporter: Irakli Machabeli
> Priority: Minor
>
> Current implementation of Row's __new__ sorts columns by name
> First of all there is no obvious reason to sort, second, if one converts
> dataframe to rdd and than back to dataframe, order of column changes. While
> this is not a bug, nevetheless it makes looking at the data really
> inconvenient.
> def __new__(self, *args, **kwargs):
> if args and kwargs:
> raise ValueError("Can not use both args "
> "and kwargs to create Row")
> if args:
> # create row class or objects
> return tuple.__new__(self, args)
> elif kwargs:
> # create row objects
> names = sorted(kwargs.keys()) # just get rid of sorting here!!!
> row = tuple.__new__(self, [kwargs[n] for n in names])
> row.__fields__ = names
> return row
> else:
> raise ValueError("No args or kwargs")
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
