Error in `read_pickle` when loading a DataFrame with MultiIndex columns from a pickle created in py27

[pedroreys]

Calling read_pickle to load a DataFrame with MultiIndex columns from a pickle file created in py27 throws an error saying:

UnicodeDecodeError: 'ascii' codec can't decode byte 0xd9 in position 0: ordinal not in range(128)

This issue was introduced by #28645, more specifically by this change:

diff --git a/pandas/io/pickle.py b/pandas/io/pickle.py
index adf0aa961..8f9bae0f7 100644
--- a/pandas/io/pickle.py
+++ b/pandas/io/pickle.py
@@ -142,18 +142,24 @@ def read_pickle(path, compression="infer"):

     # 1) try standard library Pickle
     # 2) try pickle_compat (older pandas version) to handle subclass changes
-    # 3) try pickle_compat with latin1 encoding
+
+    excs_to_catch = (AttributeError, ImportError)
+    if PY36:
+        excs_to_catch += (ModuleNotFoundError,)

     try:
         with warnings.catch_warnings(record=True):
             # We want to silence any warnings about, e.g. moved modules.
             warnings.simplefilter("ignore", Warning)
             return pickle.load(f)
-    except Exception:
-        try:
-            return pc.load(f, encoding=None)
-        except Exception:
-            return pc.load(f, encoding="latin1")
+    except excs_to_catch:
+        # e.g.
+        #  "No module named 'pandas.core.sparse.series'"
+        #  "Can't get attribute '__nat_unpickle' on <module 'pandas._libs.tslib"
+        return pc.load(f, encoding=None)
+    except UnicodeDecodeError:
+        # e.g. can occur for files written in py27; see GH#28645
+        return pc.load(f, encoding="latin-1")
     finally:
         f.close()
         for _f in fh:

Note how before when there was an Exception trying to load the file with the builtin pickle.load function, it would try to load them using pickle_compat with enconding=None and if that call also threw an Exception, it would then fallback to try to load with encoding="latin-1". With the change from #28645 the fallback to use encoding="latin-1" is only in the catch block of the initial pickle_load call, not the second one that tries to use pickle_compat with encoding=None.

This become an issue for py27 pickles with MultiIndex columns after FrozenNDArray was removed by #29840 as pickle.load(f) throws an AttributeError for FrozenNDArray and then pc.load(f, encoding=None) throws an UnicodeDecodeError.

Here is a full stack trace:

In [1]: import pandas as pd
   ...: df = pd.read_pickle('test_mi_py27.pkl')
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
~/anaconda3/envs/pandas-pickle/lib/python3.7/site-packages/pandas/io/pickle.py in read_pickle(filepath_or_buffer, compression)
    180             warnings.simplefilter("ignore", Warning)
--> 181             return pickle.load(f)
    182     except excs_to_catch:

AttributeError: Can't get attribute 'FrozenNDArray' on <module 'pandas.core.indexes.frozen' from '~/anaconda3/envs/pandas-pickle/lib/python3.7/site-packages/pandas/core/indexes/frozen.py'>

During handling of the above exception, another exception occurred:

UnicodeDecodeError                        Traceback (most recent call last)
<ipython-input-1-8c3800ea96cb> in <module>
      1 import pandas as pd
----> 2 df = pd.read_pickle('test_mi_py27.pkl')

~/anaconda3/envs/pandas-pickle/lib/python3.7/site-packages/pandas/io/pickle.py in read_pickle(filepath_or_buffer, compression)
    184         #  "No module named 'pandas.core.sparse.series'"
    185         #  "Can't get attribute '__nat_unpickle' on <module 'pandas._libs.tslib"
--> 186         return pc.load(f, encoding=None)
    187     except UnicodeDecodeError:
    188         # e.g. can occur for files written in py27; see GH#28645

~/anaconda3/envs/pandas-pickle/lib/python3.7/site-packages/pandas/compat/pickle_compat.py in load(fh, encoding, is_verbose)
    239         up.is_verbose = is_verbose
    240
--> 241         return up.load()
    242     except (ValueError, TypeError):
    243         raise

~/anaconda3/envs/pandas-pickle/lib/python3.7/pickle.py in load(self)
   1086                     raise EOFError
   1087                 assert isinstance(key, bytes_types)
-> 1088                 dispatch[key[0]](self)
   1089         except _Stop as stopinst:
   1090             return stopinst.value

~/anaconda3/envs/pandas-pickle/lib/python3.7/pickle.py in load_short_binstring(self)
   1262         len = self.read(1)[0]
   1263         data = self.read(len)
-> 1264         self.append(self._decode_string(data))
   1265     dispatch[SHORT_BINSTRING[0]] = load_short_binstring
   1266

~/anaconda3/envs/pandas-pickle/lib/python3.7/pickle.py in _decode_string(self, value)
   1202             return value
   1203         else:
-> 1204             return value.decode(self.encoding, self.errors)
   1205
   1206     def load_string(self):

UnicodeDecodeError: 'ascii' codec can't decode byte 0xd9 in position 0: ordinal not in range(128)

I believe the fix for this specific issue can be as simple as:

diff --git a/pandas/io/pickle.py b/pandas/io/pickle.py
index e51f24b55..5c4f2d8c4 100644
--- a/pandas/io/pickle.py
+++ b/pandas/io/pickle.py
@@ -183,7 +183,11 @@ def read_pickle(
         # e.g.
         #  "No module named 'pandas.core.sparse.series'"
         #  "Can't get attribute '__nat_unpickle' on <module 'pandas._libs.tslib"
-        return pc.load(f, encoding=None)
+        try:
+            return pc.load(f, encoding=None)
+        except UnicodeDecodeError:
+            # e.g. can occur for files written in py27;
+            return pc.load(f, encoding="latin-1")
     except UnicodeDecodeError:
         # e.g. can occur for files written in py27; see GH#28645
         return pc.load(f, encoding="latin-1")

I will open a PR with the fix above over the weekend.

Thanks!

[jorisvandenbossche]

Thanks for the report, and the analysis! PR very welcome