diff --git a/.travis.yml b/.travis.yml
index 99dada51..a95faa2b 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -23,7 +23,7 @@ jobs:
     env:
 
 install:
-- pip install -U setuptools
+- pip install -U setuptools cython
 - python setup.py install
 - pip install -Ur requirements.txt
 - pip install codecov
diff --git a/aiopg/sa/result.py b/aiopg/sa/result.pyx
similarity index 67%
rename from aiopg/sa/result.py
rename to aiopg/sa/result.pyx
index d58bf7b6..a6d8846d 100644
--- a/aiopg/sa/result.py
+++ b/aiopg/sa/result.pyx
@@ -1,69 +1,84 @@
-import weakref
-from collections.abc import Mapping, Sequence
-
+#cython: language_level=3, boundscheck=False, wraparound=False, nonecheck=False
+from cpython cimport Py_XINCREF, PyObject
+from cpython.tuple cimport PyTuple_SET_ITEM, PyTuple_GET_ITEM, PyTuple_GetItem
+from cpython.dict cimport PyDict_GetItem
+from cpython.list cimport PyList_New, PyList_SET_ITEM
 from sqlalchemy.sql import expression, sqltypes
 
 from . import exc
 
 
-class RowProxy(Mapping):
-    __slots__ = ('_result_proxy', '_row', '_processors', '_keymap')
+cdef inline object _raise_on_ambigous_column(key):
+    raise exc.InvalidRequestError(
+        "Ambiguous column name %s in result set! "
+        "try 'use_labels' option on select statement." % repr(key)
+    )
+
 
-    def __init__(self, result_proxy, row, processors, keymap):
+cdef class RowProxy:
+    cdef ResultMetaData _result_metadata
+    cdef tuple _row
+    cdef dict _key_to_index
+
+    def __cinit__(self, result_metadata, row, key_to_index):
         """RowProxy objects are constructed by ResultProxy objects."""
-        self._result_proxy = result_proxy
+        self._result_metadata = result_metadata
         self._row = row
-        self._processors = processors
-        self._keymap = keymap
+        self._key_to_index = key_to_index
 
     def __iter__(self):
-        return iter(self._result_proxy.keys)
+        return iter(self._result_metadata.keys)
 
     def __len__(self):
         return len(self._row)
 
-    def __getitem__(self, key):
-        try:
-            processor, obj, index = self._keymap[key]
-        except KeyError:
-            processor, obj, index = self._result_proxy._key_fallback(key)
-        # Do we need slicing at all? RowProxy now is Mapping not Sequence
-        # except TypeError:
-        #     if isinstance(key, slice):
-        #         l = []
-        #         for processor, value in zip(self._processors[key],
-        #                                     self._row[key]):
-        #             if processor is None:
-        #                 l.append(value)
-        #             else:
-        #                 l.append(processor(value))
-        #         return tuple(l)
-        #     else:
-        #         raise
-        if index is None:
-            raise exc.InvalidRequestError(
-                "Ambiguous column name '%s' in result set! "
-                "try 'use_labels' option on select statement." % key)
-        if processor is not None:
-            return processor(self._row[index])
-        else:
-            return self._row[index]
+    cdef object _getitem(self, key):
+        cdef PyObject* item
+        cdef int index
+        cdef tuple data
+
+        item = PyDict_GetItem(self._key_to_index, key)
+
+        if item is NULL:
+            fallback_key = self._result_metadata._key_fallback(key)
+            if fallback_key is None:
+                _raise_on_ambigous_column(key)
+            item = PyDict_GetItem(self._key_to_index, fallback_key)
+            if item is NULL:
+                _raise_on_ambigous_column(key)
+
+        index = <object>item
 
-    def __getattr__(self, name):
+        return <object>PyTuple_GET_ITEM(self._row, index)
+
+    def __getitem__(self, item):
+        cdef PyObject* result
+
+        if isinstance(item, int):
+            result = PyTuple_GetItem(self._row, item)
+            if result is NULL:
+                raise KeyError(item)
+            return <object>result
+        return self._getitem(item)
+
+    cdef object _getattr(self, str name):
         try:
-            return self[name]
+            return self._getitem(name)
         except KeyError as e:
             raise AttributeError(e.args[0])
 
+    def __getattr__(self, str item):
+        return self._getattr(item)
+
     def __contains__(self, key):
-        return self._result_proxy._has_key(self._row, key)
+        return self._result_metadata._has_key(self._row, key)
 
     __hash__ = None
 
     def __eq__(self, other):
-        if isinstance(other, RowProxy):
+        if hasattr(other, 'as_tuple'):
             return self.as_tuple() == other.as_tuple()
-        elif isinstance(other, Sequence):
+        elif hasattr(other, 'index') and hasattr(other, 'count'):
             return self.as_tuple() == other
         else:
             return NotImplemented
@@ -71,26 +86,30 @@ def __eq__(self, other):
     def __ne__(self, other):
         return not self == other
 
-    def as_tuple(self):
+    cdef tuple as_tuple(self):
         return tuple(self[k] for k in self)
 
     def __repr__(self):
         return repr(self.as_tuple())
 
 
-class ResultMetaData(object):
+cdef class ResultMetaData:
     """Handle cursor.description, applying additional info from an execution
     context."""
+    cdef public list keys
+    cdef public dict _keymap
+    cdef public list _processors
+    cdef public dict _key_to_index
 
-    def __init__(self, result_proxy, cursor_description):
-        self._processors = processors = []
-
+    def __init__(self, ResultProxy result_proxy, cursor_description):
         map_type, map_column_name = self.result_map(result_proxy._result_map)
 
         # We do not strictly need to store the processor in the key mapping,
         # though it is faster in the Python version (probably because of the
         # saved attribute lookup self._processors)
         self._keymap = keymap = {}
+        self._processors = processors = []
+        self._key_to_index = key_to_index = {}
         self.keys = []
         dialect = result_proxy.dialect
 
@@ -102,12 +121,11 @@ def __init__(self, result_proxy, cursor_description):
         assert dialect.case_sensitive, \
             "Doesn't support case insensitive database connection"
 
-        # high precedence key values.
-        primary_keymap = {}
-
         assert not dialect.description_encoding, \
             "psycopg in py3k should not use this"
 
+        ambiguous = []
+
         for i, rec in enumerate(cursor_description):
             colname = rec[0]
             coltype = rec[1]
@@ -116,43 +134,27 @@ def __init__(self, result_proxy, cursor_description):
             # if dialect.requires_name_normalize:
             #     colname = dialect.normalize_name(colname)
 
-            name, obj, type_ = (
-                map_column_name.get(colname, colname),
-                None,
-                map_type.get(colname, typemap.get(coltype, sqltypes.NULLTYPE))
+            name = str(map_column_name.get(colname, colname))
+            type_ = map_type.get(
+                colname,
+                typemap.get(coltype, sqltypes.NULLTYPE)
             )
-
             processor = type_._cached_result_processor(dialect, coltype)
+            if processor:
+                processors.append((processor, i))
 
-            processors.append(processor)
-            rec = (processor, obj, i)
-
-            # indexes as keys. This is only needed for the Python version of
-            # RowProxy (the C version uses a faster path for integer indexes).
-            primary_keymap[i] = rec
+            rec = (processor, i)
 
-            # populate primary keymap, looking for conflicts.
-            if primary_keymap.setdefault(name, rec) is not rec:
-                # place a record that doesn't have the "index" - this
-                # is interpreted later as an AmbiguousColumnError,
-                # but only when actually accessed.   Columns
-                # colliding by name is not a problem if those names
-                # aren't used; integer access is always
-                # unambiguous.
-                primary_keymap[name] = rec = (None, obj, None)
+            if name in keymap:
+                ambiguous.append(name)
+            keymap[name] = rec
+            key_to_index[name] = i
 
             self.keys.append(name)
-            if obj:
-                for o in obj:
-                    keymap[o] = rec
-                    # technically we should be doing this but we
-                    # are saving on callcounts by not doing so.
-                    # if keymap.setdefault(o, rec) is not rec:
-                    #    keymap[o] = (None, obj, None)
-
-        # overwrite keymap values with those of the
-        # high precedence keymap.
-        keymap.update(primary_keymap)
+
+        for name in ambiguous:
+            keymap.pop(name)
+            key_to_index.pop(name)
 
     def result_map(self, data_map):
         data_map = data_map or {}
@@ -166,50 +168,32 @@ def result_map(self, data_map):
 
         return map_type, map_column_name
 
-    def _key_fallback(self, key, raiseerr=True):
+    def _key_fallback(self, key):
         map = self._keymap
         result = None
         if isinstance(key, str):
-            result = map.get(key)
+            result = key
         # fallback for targeting a ColumnElement to a textual expression
         # this is a rare use case which only occurs when matching text()
         # or colummn('name') constructs to ColumnElements, or after a
         # pickle/unpickle roundtrip
         elif isinstance(key, expression.ColumnElement):
             if (key._label and key._label in map):
-                result = map[key._label]
+                result = key._label
             elif (hasattr(key, 'key') and key.key in map):
                 # match is only on name.
-                result = map[key.key]
-            # search extra hard to make sure this
-            # isn't a column/label name overlap.
-            # this check isn't currently available if the row
-            # was unpickled.
-            if result is not None and result[1] is not None:
-                for obj in result[1]:
-                    if key._compare_name_for_result(obj):
-                        break
-                else:
-                    result = None
-        if result is None:
-            if raiseerr:
-                raise exc.NoSuchColumnError(
-                    "Could not locate column in row for column '%s'" %
-                    expression._string_or_unprintable(key))
-            else:
-                return None
-        else:
-            map[key] = result
+                result = key.key
+
         return result
 
     def _has_key(self, row, key):
         if key in self._keymap:
             return True
         else:
-            return self._key_fallback(key, False) is not None
+            return self._key_fallback(key) is not None
 
 
-class ResultProxy:
+cdef class ResultProxy:
     """Wraps a DB-API cursor object to provide easier access to row columns.
 
     Individual columns may be accessed by their integer position,
@@ -228,15 +212,20 @@ class ResultProxy:
     data using sqlalchemy TypeEngine objects, which are referenced from
     the originating SQL statement that produced this result set.
     """
-
-    def __init__(self, connection, cursor, dialect, result_map=None):
+    cdef object _dialect
+    cdef public object _result_map
+    cdef object _cursor
+    cdef object _connection
+    cdef object _rowcount
+    cdef object _metadata
+
+    def __cinit__(self, connection, cursor, dialect, result_map=None):
         self._dialect = dialect
         self._result_map = result_map
         self._cursor = cursor
         self._connection = connection
         self._rowcount = cursor.rowcount
         self._metadata = None
-        self._weak = None
         self._init_metadata()
 
     @property
@@ -290,10 +279,8 @@ def _init_metadata(self):
         cursor_description = self.cursor.description
         if cursor_description is not None:
             self._metadata = ResultMetaData(self, cursor_description)
-            self._weak = weakref.ref(self, lambda wr: self.cursor.close())
         else:
             self.close()
-            self._weak = None
 
     @property
     def returns_rows(self):
@@ -333,7 +320,9 @@ def close(self):
             self.cursor.close()
             # allow consistent errors
             self._cursor = None
-            self._weak = None
+
+    def __del__(self):
+        self.close()
 
     def __aiter__(self):
         return self
@@ -353,13 +342,29 @@ def _non_result(self):
         else:
             raise exc.ResourceClosedError("This result object is closed.")
 
-    def _process_rows(self, rows):
-        process_row = RowProxy
+    cdef list _process_rows(self, list rows):
+        cdef int i
+        cdef list results
+
+        results = PyList_New(len(rows))
         metadata = self._metadata
-        keymap = metadata._keymap
-        processors = metadata._processors
-        return [process_row(metadata, row, processors, keymap)
-                for row in rows]
+        key_to_index = metadata._key_to_index
+        processors = self._metadata._processors
+        i = 0
+        for row in rows:
+            for processor, index in processors:
+                value = processor(<object>PyTuple_GET_ITEM(row, index))
+                PyTuple_SET_ITEM(row, index, value)
+            row_proxy = RowProxy(
+                metadata,
+                row,
+                key_to_index
+            )
+            Py_XINCREF(<PyObject*>row_proxy)
+            check = PyList_SET_ITEM(results, i, row_proxy)
+            i += 1
+
+        return results
 
     async def fetchall(self):
         """Fetch all rows, just like DB-API cursor.fetchall()."""
diff --git a/setup.py b/setup.py
index c4e41fe4..a981e690 100644
--- a/setup.py
+++ b/setup.py
@@ -1,7 +1,10 @@
 import os
 import re
 
-from setuptools import setup, find_packages
+from setuptools import setup, find_packages, Extension
+from Cython.Build import cythonize
+
+CFLAGS = ['-O3']
 
 install_requires = ['psycopg2-binary>=2.7.0']
 extras_require = {'sa': ['sqlalchemy[postgresql_psycopg2binary]>=1.1']}
@@ -50,6 +53,14 @@ def read_changelog(path='CHANGES.txt'):
     'Framework :: AsyncIO',
 ]
 
+EXTENSIONS = [
+    Extension(
+        "aiopg.sa.result",
+        sources=["aiopg/sa/result.pyx"],
+        extra_compile_args=CFLAGS
+    )
+]
+
 setup(
     name='aiopg',
     version=read_version(),
@@ -76,5 +87,6 @@ def read_changelog(path='CHANGES.txt'):
     packages=find_packages(),
     install_requires=install_requires,
     extras_require=extras_require,
-    include_package_data=True
+    include_package_data=True,
+    ext_modules=cythonize(EXTENSIONS)
 )