[SPARK-22395][SQL][PYTHON] Fix the behavior of timestamp values for Pandas to respect session timezone

docs/sql-programming-guide.md

@@ -1716,6 +1716,8 @@ options.
     </table>
 
     Note that, for <b>DecimalType(38,0)*</b>, the table above intentionally does not cover all other combinations of scales and precisions because currently we only infer decimal type like `BigInteger`/`BigInt`. For example, 1.1 is inferred as double type.
+  - In PySpark, now we need Pandas 0.19.2 or upper if you want to use Pandas related functionalities, such as `toPandas`, `createDataFrame` from Pandas DataFrame, etc.
+  - In PySpark, the behavior of timestamp values for Pandas related functionalities was changed to respect session timezone. If you want to use the old behavior, you need to set a configuration `spark.sql.execution.pandas.respectSessionTimeZone` to `False`. See [SPARK-22395](https://issues.apache.org/jira/browse/SPARK-22395) for details.
 
 ## Upgrading From Spark SQL 2.1 to 2.2
 

python/pyspark/serializers.py

@@ -206,11 +206,12 @@ def __repr__(self):
         return "ArrowSerializer"
 
 
-def _create_batch(series):
+def _create_batch(series, timezone):
     """
     Create an Arrow record batch from the given pandas.Series or list of Series, with optional type.
 
     :param series: A single pandas.Series, list of Series, or list of (series, arrow_type)
+    :param timezone: A timezone to respect when handling timestamp values
     :return: Arrow RecordBatch
     """
 
@@ -227,7 +228,7 @@ def _create_batch(series):
     def cast_series(s, t):
         if type(t) == pa.TimestampType:
             # NOTE: convert to 'us' with astype here, unit ignored in `from_pandas` see ARROW-1680
-            return _check_series_convert_timestamps_internal(s.fillna(0))\
+            return _check_series_convert_timestamps_internal(s.fillna(0), timezone)\
                 .values.astype('datetime64[us]', copy=False)
         # NOTE: can not compare None with pyarrow.DataType(), fixed with Arrow >= 0.7.1
         elif t is not None and t == pa.date32():
@@ -253,6 +254,10 @@ class ArrowStreamPandasSerializer(Serializer):
     Serializes Pandas.Series as Arrow data with Arrow streaming format.
     """
 
+    def __init__(self, timezone):
+        super(ArrowStreamPandasSerializer, self).__init__()
+        self._timezone = timezone
+
     def dump_stream(self, iterator, stream):
         """
         Make ArrowRecordBatches from Pandas Series and serialize. Input is a single series or
@@ -262,7 +267,7 @@ def dump_stream(self, iterator, stream):
         writer = None
         try:
             for series in iterator:
-                batch = _create_batch(series)
+                batch = _create_batch(series, self._timezone)
                 if writer is None:
                     write_int(SpecialLengths.START_ARROW_STREAM, stream)
                     writer = pa.RecordBatchStreamWriter(stream, batch.schema)
@@ -280,7 +285,7 @@ def load_stream(self, stream):
         reader = pa.open_stream(stream)
         for batch in reader:
             # NOTE: changed from pa.Columns.to_pandas, timezone issue in conversion fixed in 0.7.1
-            pdf = _check_dataframe_localize_timestamps(batch.to_pandas())
+            pdf = _check_dataframe_localize_timestamps(batch.to_pandas(), self._timezone)
             yield [c for _, c in pdf.iteritems()]
 
     def __repr__(self):

python/pyspark/sql/dataframe.py

@@ -39,6 +39,7 @@
 from pyspark.sql.streaming import DataStreamWriter
 from pyspark.sql.types import IntegralType
 from pyspark.sql.types import *
+from pyspark.util import _exception_message
 
 __all__ = ["DataFrame", "DataFrameNaFunctions", "DataFrameStatFunctions"]
 
@@ -1881,6 +1882,13 @@ def toPandas(self):
         1    5    Bob
         """
         import pandas as pd
+
+        if self.sql_ctx.getConf("spark.sql.execution.pandas.respectSessionTimeZone").lower() \
+           == "true":
+            timezone = self.sql_ctx.getConf("spark.sql.session.timeZone")
+        else:
+            timezone = None
+
         if self.sql_ctx.getConf("spark.sql.execution.arrow.enabled", "false").lower() == "true":
             try:
                 from pyspark.sql.types import _check_dataframe_localize_timestamps
@@ -1889,13 +1897,13 @@ def toPandas(self):
                 if tables:
                     table = pyarrow.concat_tables(tables)
                     pdf = table.to_pandas()
-                    return _check_dataframe_localize_timestamps(pdf)
+                    return _check_dataframe_localize_timestamps(pdf, timezone)
                 else:
                     return pd.DataFrame.from_records([], columns=self.columns)
             except ImportError as e:
                 msg = "note: pyarrow must be installed and available on calling Python process " \
                       "if using spark.sql.execution.arrow.enabled=true"
-                raise ImportError("%s\n%s" % (e.message, msg))
+                raise ImportError("%s\n%s" % (_exception_message(e), msg))
         else:
             pdf = pd.DataFrame.from_records(self.collect(), columns=self.columns)
 
@@ -1913,7 +1921,17 @@ def toPandas(self):
 
             for f, t in dtype.items():
                 pdf[f] = pdf[f].astype(t, copy=False)
-            return pdf
+
+            if timezone is None:
+                return pdf
+            else:
+                from pyspark.sql.types import _check_series_convert_timestamps_local_tz
+                for field in self.schema:
+                    # TODO: handle nested timestamps, such as ArrayType(TimestampType())?
+                    if isinstance(field.dataType, TimestampType):
+                        pdf[field.name] = \
+                            _check_series_convert_timestamps_local_tz(pdf[field.name], timezone)
+                return pdf
 
     def _collectAsArrow(self):
         """

python/pyspark/sql/session.py

@@ -34,8 +34,9 @@
 from pyspark.sql.dataframe import DataFrame
 from pyspark.sql.readwriter import DataFrameReader
 from pyspark.sql.streaming import DataStreamReader
-from pyspark.sql.types import Row, DataType, StringType, StructType, _make_type_verifier, \
-    _infer_schema, _has_nulltype, _merge_type, _create_converter, _parse_datatype_string
+from pyspark.sql.types import Row, DataType, StringType, StructType, TimestampType, \
+    _make_type_verifier, _infer_schema, _has_nulltype, _merge_type, _create_converter, \
+    _parse_datatype_string
 from pyspark.sql.utils import install_exception_handler
 
 __all__ = ["SparkSession"]
@@ -444,11 +445,34 @@ def _get_numpy_record_dtype(self, rec):
             record_type_list.append((str(col_names[i]), curr_type))
         return np.dtype(record_type_list) if has_rec_fix else None
 
-    def _convert_from_pandas(self, pdf):
+    def _convert_from_pandas(self, pdf, schema, timezone):
         """
          Convert a pandas.DataFrame to list of records that can be used to make a DataFrame
          :return list of records
         """
+        if timezone is not None:
+            from pyspark.sql.types import _check_series_convert_timestamps_tz_local
+            copied = False
+            if isinstance(schema, StructType):
+                for field in schema:
+                    # TODO: handle nested timestamps, such as ArrayType(TimestampType())?
+                    if isinstance(field.dataType, TimestampType):
+                        s = _check_series_convert_timestamps_tz_local(pdf[field.name], timezone)
+                        if not copied and s is not pdf[field.name]:
+                            # Copy once if the series is modified to prevent the original Pandas
+                            # DataFrame from being updated
+                            pdf = pdf.copy()
+                            copied = True
+                        pdf[field.name] = s
+            else:
+                for column, series in pdf.iteritems():
+                    s = _check_series_convert_timestamps_tz_local(pdf[column], timezone)
+                    if not copied and s is not pdf[column]:
+                        # Copy once if the series is modified to prevent the original Pandas
+                        # DataFrame from being updated
+                        pdf = pdf.copy()
+                        copied = True
+                    pdf[column] = s
 
         # Convert pandas.DataFrame to list of numpy records
         np_records = pdf.to_records(index=False)
@@ -462,15 +486,19 @@ def _convert_from_pandas(self, pdf):
         # Convert list of numpy records to python lists
         return [r.tolist() for r in np_records]
 
-    def _create_from_pandas_with_arrow(self, pdf, schema):
+    def _create_from_pandas_with_arrow(self, pdf, schema, timezone):
         """
         Create a DataFrame from a given pandas.DataFrame by slicing it into partitions, converting
         to Arrow data, then sending to the JVM to parallelize. If a schema is passed in, the
         data types will be used to coerce the data in Pandas to Arrow conversion.
         """
         from pyspark.serializers import ArrowSerializer, _create_batch
-        from pyspark.sql.types import from_arrow_schema, to_arrow_type, TimestampType
-        from pandas.api.types import is_datetime64_dtype, is_datetime64tz_dtype
+        from pyspark.sql.types import from_arrow_schema, to_arrow_type, \
+            _old_pandas_exception_message, TimestampType
+        try:
+            from pandas.api.types import is_datetime64_dtype, is_datetime64tz_dtype
+        except ImportError as e:
+            raise ImportError(_old_pandas_exception_message(e))
 
         # Determine arrow types to coerce data when creating batches
         if isinstance(schema, StructType):
@@ -488,7 +516,8 @@ def _create_from_pandas_with_arrow(self, pdf, schema):
         pdf_slices = (pdf[start:start + step] for start in xrange(0, len(pdf), step))
 
         # Create Arrow record batches
-        batches = [_create_batch([(c, t) for (_, c), t in zip(pdf_slice.iteritems(), arrow_types)])
+        batches = [_create_batch([(c, t) for (_, c), t in zip(pdf_slice.iteritems(), arrow_types)],
+                                 timezone)
                    for pdf_slice in pdf_slices]
 
         # Create the Spark schema from the first Arrow batch (always at least 1 batch after slicing)
@@ -606,6 +635,11 @@ def createDataFrame(self, data, schema=None, samplingRatio=None, verifySchema=Tr
         except Exception:
             has_pandas = False
         if has_pandas and isinstance(data, pandas.DataFrame):
+            if self.conf.get("spark.sql.execution.pandas.respectSessionTimeZone").lower() \
+               == "true":
+                timezone = self.conf.get("spark.sql.session.timeZone")
+            else:
+                timezone = None
 
             # If no schema supplied by user then get the names of columns only
             if schema is None:
@@ -614,11 +648,11 @@ def createDataFrame(self, data, schema=None, samplingRatio=None, verifySchema=Tr
             if self.conf.get("spark.sql.execution.arrow.enabled", "false").lower() == "true" \
                     and len(data) > 0:
                 try:
-                    return self._create_from_pandas_with_arrow(data, schema)
+                    return self._create_from_pandas_with_arrow(data, schema, timezone)
                 except Exception as e:
                     warnings.warn("Arrow will not be used in createDataFrame: %s" % str(e))
                     # Fallback to create DataFrame without arrow if raise some exception
-            data = self._convert_from_pandas(data)
+            data = self._convert_from_pandas(data, schema, timezone)
 
         if isinstance(schema, StructType):
             verify_func = _make_type_verifier(schema) if verifySchema else lambda _: True