diff --git a/lib/iris/cube.py b/lib/iris/cube.py
index 3b6a8420e3..db1147eef1 100644
--- a/lib/iris/cube.py
+++ b/lib/iris/cube.py
@@ -1658,9 +1658,14 @@ def fill_value(self):
     @fill_value.setter
     def fill_value(self, fill_value):
         if fill_value is not None:
+            # Convert the given value to the dtype of the cube.
+            fill_value = np.asarray([fill_value])[0]
+            target_dtype = self.dtype
+            if fill_value.dtype.kind == 'f' and target_dtype.kind == 'i':
+                # Perform rounding when converting floats to ints.
+                fill_value = np.rint(fill_value)
             try:
-                fill_value = np.asarray([fill_value],
-                                        dtype=self.dtype)[0]
+                fill_value = np.asarray([fill_value], dtype=target_dtype)[0]
             except OverflowError:
                 emsg = 'Fill value of {!r} invalid for cube {!r}.'
                 raise ValueError(emsg.format(fill_value, self.dtype))
@@ -1763,12 +1768,17 @@ def data(self, value):
                 raise ValueError('Require cube data with shape %r, got '
                                  '%r.' % (self.shape, value.shape))
 
+        # Set lazy or real data, and reset the other.
         if is_lazy_data(value):
             self._dask_array = value
             self._numpy_array = None
-
         else:
             self._numpy_array = value
+            self._dask_array = None
+
+        # Cancel any 'realisation' datatype conversion, and fill value.
+        self._dtype = None
+        self.fill_value = None
 
     def has_lazy_data(self):
         return self._numpy_array is None
diff --git a/lib/iris/tests/unit/cube/test_Cube.py b/lib/iris/tests/unit/cube/test_Cube.py
index 74686d33f4..9260d8dc86 100644
--- a/lib/iris/tests/unit/cube/test_Cube.py
+++ b/lib/iris/tests/unit/cube/test_Cube.py
@@ -1359,7 +1359,6 @@ def test_lazydata_dtype_change(self):
         self.assertEqual(cube.core_data.dtype, np.dtype('f4'))
         self.assertIsNone(cube.fill_value)
 
-
     def test_lazydata_maskedints_dtype_change(self):
         # Check that re-assigning dtype resets fill_value.
         cube = self._sample_cube(dtype=np.int16, masked=True, lazy=True,
@@ -1384,6 +1383,31 @@ def test_fill_value__float_dtype_to_int(self):
         self.assertEqual(cube.fill_value.dtype, np.dtype('i2'))
         self.assertArrayAllClose(cube.fill_value, 1735)
 
+    def test_set_fill_value(self):
+        cube = self._sample_cube()
+        cube.fill_value = -74.6
+        self.assertEqual(cube.fill_value.dtype, np.dtype('f4'))
+        self.assertArrayAllClose(cube.fill_value, -74.6)
+
+    def test_set_fill_value__typecast(self):
+        cube = self._sample_cube(dtype=np.int16)
+        cube.fill_value = -74.6
+        self.assertEqual(cube.fill_value.dtype, np.dtype('i2'))
+        self.assertArrayAllClose(cube.fill_value, -75)
+
+    def test_set_fill_value__casterror(self):
+        cube = self._sample_cube(dtype=np.int16)
+        msg = "invalid for cube dtype\('int16'\)"
+        with self.assertRaisesRegexp(ValueError, msg):
+            # NOTE: this doesn't actually work properly in most cases.
+            # E.G. it will happily assign 1e12 to an int16 and gets 4096.
+            cube.fill_value = -1.0e23
+
+    def test_clear_fill_value(self):
+        cube = self._sample_cube(cube_fill_value=123.768)
+        cube.fill_value = None
+        self.assertIsNone(cube.fill_value)
+
 
 class TestSubset(tests.IrisTest):
     def test_scalar_coordinate(self):