[
https://issues.apache.org/jira/browse/SPARK-16394?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15906266#comment-15906266
]
Keith Bourgoin commented on SPARK-16394:
----------------------------------------
We've been having the same issue. To illustrate, this is what happens when you
put datetimes first into an RDD and then a DataFrame:
{code}
In [10]: rdd = sc.parallelize([(1, 2, dt.datetime.now(),
pytz.utc.localize(dt.datetime.now()))])
In [11]: rdd.collect()
Out[11]:
[(1,
2,
datetime.datetime(2017, 3, 11, 12, 40, 25, 834984),
datetime.datetime(2017, 3, 11, 12, 40, 25, 834996, tzinfo=<UTC>))]
In [12]: df = sqlContext.createDataFrame(rdd)
In [13]: df.collect()
Out[13]: [Row(_1=1, _2=2, _3=datetime.datetime(2017, 3, 11, 12, 40, 25,
834984), _4=datetime.datetime(2017, 3, 11, 7, 40, 25, 834996))]
{code}
The datetime in the df has lost its timezone information and is now a different
value. This is incredibly confusing, and I just lost about a day of time trying
to figure out what exactly was going on.
As the previous commenter said, the only way around this is to use a string or
int representation of the value. If you do that, you lose all of Spark's
date-related functionality.
IMO, this is more of a Major bug than a Minor.
> Timestamp conversion error in pyspark.sql.Row because of timezones
> ------------------------------------------------------------------
>
> Key: SPARK-16394
> URL: https://issues.apache.org/jira/browse/SPARK-16394
> Project: Spark
> Issue Type: Bug
> Components: PySpark
> Affects Versions: 1.6.1
> Reporter: Martin Tapp
> Priority: Minor
>
> We use DataFrame.map to convert each row to a dictionary using Row.asDict().
> The problem occurs when a Timestamp column is converted. It seems the
> Timestamp gets converted to a naive Python datetime. This causes processing
> errors since all naive datetimes get adjusted to the process' timezone. For
> instance, a Timestamp with a time of midnight see's it's time bounce based on
> the local timezone (+/- x hours).
> Current fix is to apply the pytz.utc timezone to each datetime instance.
> Proposed solution is to make all datetime instances aware and use the
> pytz.utc timezone.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
