[Backport pandas-dev#14596] BUG: Fix parse empty df

closes pandas-dev#14515

This commit fixes a bug where `read_csv` failed when given a file with
a multiindex header and empty content. Because pandas reads index
names as a separate line following the header lines, the reader looks
for the line with index names in it. If the content of the dataframe
is empty, the reader will choke. This bug surfaced after
pandas-dev#6618 stopped writing an
extra line after multiindex columns, which led to a situation where
pandas could write CSV's that it couldn't then read.     This commit
changes that behavior by explicitly checking if the index name row
exists, and processing it correctly if it doesn't.

Author: Ben Kandel <[email protected]>

Closes pandas-dev#14596 from bkandel/fix-parse-empty-df and squashes the following commits:

32e3b0a [Ben Kandel] lint
e6b1237 [Ben Kandel] lint
fedfff8 [Ben Kandel] fix multiindex column parsing
518982d [Ben Kandel] move to 0.19.2
fc23e5c [Ben Kandel] fix errant this_columns
3d9bbdd [Ben Kandel] whatsnew
68eadf3 [Ben Kandel] Modify test.
17e44dd [Ben Kandel] fix python parser too
72adaf2 [Ben Kandel] remove unnecessary test
bfe0423 [Ben Kandel] typo
2f64d57 [Ben Kandel] pep8
b8200e4 [Ben Kandel] BUG: read_csv with empty df

(cherry picked from commit f862b52)

Author: Ben Kandel

doc/source/whatsnew/v0.19.2.txt

@@ -29,7 +29,7 @@ Bug Fixes
 
 - Compat with ``dateutil==2.6.0``; segfault reported in the testing suite (:issue:`14621`)
 - Allow ``nanoseconds`` in ``Timestamp.replace`` as a kwarg (:issue:`14621`)
-
+- Bug in ``pd.read_csv`` where reading files fails, if the number of headers is equal to the number of lines in the file (:issue:`14515`)
 
 
 

pandas/io/parsers.py

@@ -1509,10 +1509,11 @@ def read(self, nrows=None):
             if self._first_chunk:
                 self._first_chunk = False
                 names = self._maybe_dedup_names(self.orig_names)
-
                 index, columns, col_dict = _get_empty_meta(
                     names, self.index_col, self.index_names,
                     dtype=self.kwds.get('dtype'))
+                columns = self._maybe_make_multi_index_columns(
+                    columns, self.col_names)
 
                 if self.usecols is not None:
                     columns = self._filter_usecols(columns)
@@ -1979,8 +1980,11 @@ def read(self, rows=None):
         if not len(content):  # pragma: no cover
             # DataFrame with the right metadata, even though it's length 0
             names = self._maybe_dedup_names(self.orig_names)
-            return _get_empty_meta(names, self.index_col,
-                                   self.index_names)
+            index, columns, col_dict = _get_empty_meta(
+                names, self.index_col, self.index_names)
+            columns = self._maybe_make_multi_index_columns(
+                columns, self.col_names)
+            return index, columns, col_dict
 
         # handle new style for names in index
         count_empty_content_vals = count_empty_vals(content[0])
@@ -2083,6 +2087,12 @@ def _infer_columns(self):
                     # We have an empty file, so check
                     # if columns are provided. That will
                     # serve as the 'line' for parsing
+                    if have_mi_columns and hr > 0:
+                        if clear_buffer:
+                            self._clear_buffer()
+                        columns.append([None] * len(columns[-1]))
+                        return columns, num_original_columns
+
                     if not self.names:
                         raise EmptyDataError(
                             "No columns to parse from file")

pandas/io/tests/parser/common.py

@@ -606,6 +606,28 @@ def test_multi_index_no_level_names(self):
         expected = self.read_csv(StringIO(data), index_col=[1, 0])
         tm.assert_frame_equal(df, expected, check_names=False)
 
+    def test_multi_index_blank_df(self):
+        # GH 14545
+        data = """a,b
+"""
+        df = self.read_csv(StringIO(data), header=[0])
+        expected = DataFrame(columns=['a', 'b'])
+        tm.assert_frame_equal(df, expected)
+        round_trip = self.read_csv(StringIO(
+            expected.to_csv(index=False)), header=[0])
+        tm.assert_frame_equal(round_trip, expected)
+
+        data_multiline = """a,b
+c,d
+"""
+        df2 = self.read_csv(StringIO(data_multiline), header=[0, 1])
+        cols = MultiIndex.from_tuples([('a', 'c'), ('b', 'd')])
+        expected2 = DataFrame(columns=cols)
+        tm.assert_frame_equal(df2, expected2)
+        round_trip = self.read_csv(StringIO(
+            expected2.to_csv(index=False)), header=[0, 1])
+        tm.assert_frame_equal(round_trip, expected2)
+
     def test_no_unnamed_index(self):
         data = """ id c0 c1 c2
 0 1 0 a b

pandas/parser.pyx

@@ -714,7 +714,9 @@ cdef class TextReader:
                     start = self.parser.line_start[0]
 
                 # e.g., if header=3 and file only has 2 lines
-                elif self.parser.lines < hr + 1:
+                elif (self.parser.lines < hr + 1
+                      and not isinstance(self.orig_header, list)) or (
+                          self.parser.lines < hr):
                     msg = self.orig_header
                     if isinstance(msg, list):
                         msg = "[%s], len of %d," % (
@@ -937,7 +939,7 @@ cdef class TextReader:
                 raise_parser_error('Error tokenizing data', self.parser)
             footer = self.skipfooter
 
-        if self.parser_start == self.parser.lines:
+        if self.parser_start >= self.parser.lines:
             raise StopIteration
         self._end_clock('Tokenization')
 

pandas/tests/frame/test_to_csv.py

@@ -587,7 +587,7 @@ def _make_frame(names=None):
             df = _make_frame(True)
             df.to_csv(path, tupleize_cols=False)
 
-            for i in [5, 6, 7]:
+            for i in [6, 7]:
                 msg = 'len of {i}, but only 5 lines in file'.format(i=i)
                 with assertRaisesRegexp(CParserError, msg):
                     read_csv(path, tupleize_cols=False,