Leave bytestring exception messages as bytestrings

Fixed a regression introduced in version 1.2 where a refactor
of the :class:`.SQLAlchemyError` base exception class introduced an
inappropriate coercion of a plain string message into Unicode under
python 2k, which is not handled by the Python interpreter for characters
outside of the platform's encoding (typically ascii).  The
:class:`.SQLAlchemyError` class now passes a bytestring through under
Py2K for ``__str__()`` as is the behavior of exception objects in general
under Py2K, does a safe coercion to unicode utf-8 with
backslash fallback for ``__unicode__()``.  For Py3K the message is
typically unicode already, but if not is again safe-coerced with utf-8
with backslash fallback for the ``__str__()`` method.

Fixes: #4429
Change-Id: I2289da3f2c45c7d0041fa43d838958f7614defc3

1 files changed, 12 insertions, 0 deletions

diff --git a/lib/sqlalchemy/util/compat.py b/lib/sqlalchemy/util/compat.py
index 7963eebb6..6c24f75e1 100644
--- a/lib/sqlalchemy/util/compat.py
+++ b/lib/sqlalchemy/util/compat.py
@@ -92,6 +92,9 @@ if py3k:
     def b64encode(x):
         return base64.b64encode(x).decode("ascii")
 
+    def decode_backslashreplace(text, encoding):
+        return text.decode(encoding, errors="backslashreplace")
+
     def cmp(a, b):
         return (a > b) - (a < b)
 
@@ -195,6 +198,15 @@ else:
     def ue(s):
         return unicode(s, "unicode_escape")  # noqa
 
+    def decode_backslashreplace(text, encoding):
+        try:
+            return text.decode(encoding)
+        except UnicodeDecodeError:
+            # regular "backslashreplace" for an incompatible encoding raises:
+            # "TypeError: don't know how to handle UnicodeDecodeError in
+            # error callback"
+            return repr(text)[1:-1].decode()
+
     # not as nice as that of Py3K, but at least preserves
     # the code line where the issue occurred
     exec(