[SPARK-44717] "pyspark.pandas.resample" is incorrect when DST is overlapped and setting "spark.sql.timestampType" to TIMESTAMP_NTZ does not help

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.4.0, 3.4.1, 4.0.0
• Fix Version/s: 3.5.0, 4.0.0
• Component/s: Pandas API on Spark
• Assignee: Hyukjin Kwon
• Reporter: Attila Zsolt Piros
• Created: 08/Aug/23 04:01
• Updated: 09/Aug/23 02:04
• Resolved: 09/Aug/23 02:04

Description

Use one of the existing test:

• "11H" case of test_dataframe_resample (pyspark.pandas.tests.test_resample.ResampleTests)
• "1001H" case of test_series_resample (pyspark.pandas.tests.test_resample.ResampleTests)

After setting the TZ for example to New York (like by using the following python code in a "setUpClass":

os.environ["TZ"] = 'America/New_York'

)

You will get the error for the latter mentioned test:

======================================================================
FAIL [4.219s]: test_series_resample (pyspark.pandas.tests.test_resample.ResampleTests)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/__w/spark/spark/python/pyspark/pandas/tests/test_resample.py", line 276, in test_series_resample
    self._test_resample(self.pdf3.A, self.psdf3.A, ["1001H"], "right", "right", "sum")
  File "/__w/spark/spark/python/pyspark/pandas/tests/test_resample.py", line 259, in _test_resample
    self.assert_eq(
  File "/__w/spark/spark/python/pyspark/testing/pandasutils.py", line 457, in assert_eq
    _assert_pandas_almost_equal(lobj, robj)
  File "/__w/spark/spark/python/pyspark/testing/pandasutils.py", line 228, in _assert_pandas_almost_equal
    raise PySparkAssertionError(
pyspark.errors.exceptions.base.PySparkAssertionError: [DIFFERENT_PANDAS_SERIES] Series are not almost equal:
Left:
Freq: 1001H
float64
Right:
float64

The problem is the in the pyspark resample there will be more resampled rows in the result. The DST change will cause those extra lines as the computed _tmp_resample_bin_col_ be something like:

| __index_level_0__  | __tmp_resample_bin_col__ | A
.....
|2011-03-08 00:00:00|2011-03-26 11:00:00     |0.3980551570183919  |
|2011-03-09 00:00:00|2011-03-26 11:00:00     |0.6511376673995046  |
|2011-03-10 00:00:00|2011-03-26 11:00:00     |0.6141085426890365  |
|2011-03-11 00:00:00|2011-03-26 11:00:00     |0.11557638066163867 |
|2011-03-12 00:00:00|2011-03-26 11:00:00     |0.4517788243490799  |
|2011-03-13 00:00:00|2011-03-26 11:00:00     |0.8637060550157284  |
|2011-03-14 00:00:00|2011-03-26 10:00:00     |0.8169499149450166  |
|2011-03-15 00:00:00|2011-03-26 10:00:00     |0.4585916249356583  |
|2011-03-16 00:00:00|2011-03-26 10:00:00     |0.8362472880832088  |
|2011-03-17 00:00:00|2011-03-26 10:00:00     |0.026716901748386812|
|2011-03-18 00:00:00|2011-03-26 10:00:00     |0.9086816462089563  |

You can see the extra lines around when the DST kicked in on 2011-03-13 in New York.

Even setting the conf "spark.sql.timestampType" to"TIMESTAMP_NTZ" does not help.

You can see my tests here:
https://github.com/attilapiros/spark/pull/5

Pandas timestamps are TZ less:
`

import pandas as pd
a = pd.Timestamp(year=2011, month=3, day=13, hour=1)
b = pd.Timedelta(hours=1)

>> a 
Timestamp('2011-03-13 01:00:00')
>>> a+b
Timestamp('2011-03-13 02:00:00')
>>> a+b+b
Timestamp('2011-03-13 03:00:00')

But pyspark TimestampType uses TZ and DST:

>>> sql("select  TIMESTAMP '2011-03-13 01:00:00'").show()
+-------------------------------+
|TIMESTAMP '2011-03-13 01:00:00'|
+-------------------------------+
|            2011-03-13 01:00:00|
+-------------------------------+

>>> sql("select  TIMESTAMP '2011-03-13 01:00:00' + make_interval(0,0,0,0,1,0,0)").show()
+--------------------------------------------------------------------+
|TIMESTAMP '2011-03-13 01:00:00' + make_interval(0, 0, 0, 0, 1, 0, 0)|
+--------------------------------------------------------------------+
|                                                 2011-03-13 03:00:00|
+--------------------------------------------------------------------+

The current resample code uses the above interval based calculation.