diff --git a/xarray/backends/api.py b/xarray/backends/api.py
index 4929da2539c..babec1beb4b 100644
--- a/xarray/backends/api.py
+++ b/xarray/backends/api.py
@@ -288,7 +288,8 @@ def maybe_decode_store(store, lock=False):
                                          allow_remote=True)
         if engine == 'netcdf4':
             store = backends.NetCDF4DataStore(filename_or_obj, group=group,
-                                              autoclose=autoclose)
+                                              autoclose=autoclose,
+                                              allow_object=True)
         elif engine == 'scipy':
             store = backends.ScipyDataStore(filename_or_obj,
                                             autoclose=autoclose)
@@ -534,7 +535,8 @@ def open_mfdataset(paths, chunks=None, concat_dim=_CONCAT_DIM_DEFAULT,
 
 
 def to_netcdf(dataset, path_or_file=None, mode='w', format=None, group=None,
-              engine=None, writer=None, encoding=None, unlimited_dims=None):
+              engine=None, writer=None, encoding=None, unlimited_dims=None,
+              allow_object=False):
     """This function creates an appropriate datastore for writing a dataset to
     disk as a netCDF file
 
@@ -574,8 +576,8 @@ def to_netcdf(dataset, path_or_file=None, mode='w', format=None, group=None,
     sync = writer is None
 
     target = path_or_file if path_or_file is not None else BytesIO()
-    store = store_cls(target, mode, format, group, writer)
-
+    store = store_cls(target, mode, format, group, writer,
+                      allow_object=allow_object)
     if unlimited_dims is None:
         unlimited_dims = dataset.encoding.get('unlimited_dims', None)
     try:
diff --git a/xarray/backends/common.py b/xarray/backends/common.py
index cec55d22589..caf9b64add0 100644
--- a/xarray/backends/common.py
+++ b/xarray/backends/common.py
@@ -184,11 +184,25 @@ def sync(self):
 
 
 class AbstractWritableDataStore(AbstractDataStore):
-    def __init__(self, writer=None):
+    def __init__(self, writer=None, allow_object=False):
+        self.allow_object = allow_object
         if writer is None:
             writer = ArrayWriter()
         self.writer = writer
 
+    @property
+    def allow_object(self):
+        return self._allow_object
+
+    @allow_object.setter
+    def allow_object(self, value):
+        if value:
+            msg = "'{}' does not support native Python object " \
+                  "serialization'".format(self.__class__.__name__)
+            raise NotImplemented(msg)
+        else:
+            self._allow_object = value
+
     def set_dimension(self, d, l):  # pragma: no cover
         raise NotImplementedError
 
@@ -241,7 +255,8 @@ class WritableCFDataStore(AbstractWritableDataStore):
     def store(self, variables, attributes, *args, **kwargs):
         # All NetCDF files get CF encoded by default, without this attempting
         # to write times, for example, would fail.
-        cf_variables, cf_attrs = cf_encoder(variables, attributes)
+        cf_variables, cf_attrs = cf_encoder(variables, attributes,
+                                            allow_object=self.allow_object)
         AbstractWritableDataStore.store(self, cf_variables, cf_attrs,
                                         *args, **kwargs)
 
diff --git a/xarray/backends/h5netcdf_.py b/xarray/backends/h5netcdf_.py
index 0e7d866bc2a..d1f15134833 100644
--- a/xarray/backends/h5netcdf_.py
+++ b/xarray/backends/h5netcdf_.py
@@ -55,7 +55,7 @@ class H5NetCDFStore(WritableCFDataStore, DataStorePickleMixin):
     """Store for reading and writing data via h5netcdf
     """
     def __init__(self, filename, mode='r', format=None, group=None,
-                 writer=None, autoclose=False):
+                 writer=None, autoclose=False, allow_object=False):
         if format not in [None, 'NETCDF4']:
             raise ValueError('invalid format for h5netcdf backend')
         opener = functools.partial(_open_h5netcdf_group, filename, mode=mode,
@@ -71,7 +71,7 @@ def __init__(self, filename, mode='r', format=None, group=None,
         self._opener = opener
         self._filename = filename
         self._mode = mode
-        super(H5NetCDFStore, self).__init__(writer)
+        super(H5NetCDFStore, self).__init__(writer, allow_object=allow_object)
 
     def open_store_variable(self, name, var):
         with self.ensure_open(autoclose=False):
diff --git a/xarray/backends/memory.py b/xarray/backends/memory.py
index f79e92439fe..243ef743745 100644
--- a/xarray/backends/memory.py
+++ b/xarray/backends/memory.py
@@ -18,10 +18,12 @@ class InMemoryDataStore(AbstractWritableDataStore):
 
     This store exists purely for internal testing purposes.
     """
-    def __init__(self, variables=None, attributes=None, writer=None):
+    def __init__(self, variables=None, attributes=None, writer=None,
+                 allow_object=False):
         self._variables = OrderedDict() if variables is None else variables
         self._attributes = OrderedDict() if attributes is None else attributes
-        super(InMemoryDataStore, self).__init__(writer)
+        super(InMemoryDataStore, self).__init__(writer,
+                                                allow_object=allow_object)
 
     def get_attrs(self):
         return self._attributes
diff --git a/xarray/backends/netCDF4_.py b/xarray/backends/netCDF4_.py
index 93af50f4ae5..4174a872585 100644
--- a/xarray/backends/netCDF4_.py
+++ b/xarray/backends/netCDF4_.py
@@ -11,11 +11,11 @@
 from ..core import indexing
 from ..core.utils import (FrozenOrderedDict, NdimSizeLenMixin,
                           DunderArrayMixin, close_on_error,
-                          is_remote_uri)
+                          is_remote_uri, encode_pickle, decode_pickle)
 from ..core.pycompat import iteritems, basestring, OrderedDict, PY3
 
 from .common import (WritableCFDataStore, robust_getitem,
-                     DataStorePickleMixin, find_root)
+                     DataStorePickleMixin, find_root, ArrayWriter)
 from .netcdf3 import (encode_nc3_attr_value, encode_nc3_variable,
                       maybe_convert_to_char_array)
 
@@ -41,6 +41,11 @@ def __init__(self, variable_name, datastore):
             # represent variable length strings; it also prevents automatic
             # string concatenation via conventions.decode_cf_variable
             dtype = np.dtype('O')
+        if dtype == np.uint8 and array.datatype.name == 'object':
+            self.is_object = True
+            dtype = np.dtype('O')
+        else:
+            self.is_object = False
         self.dtype = dtype
 
     def get_array(self):
@@ -75,6 +80,8 @@ def __getitem__(self, key):
             # arrays (slicing them always returns a 1-dimensional array):
             # https://github.com/Unidata/netcdf4-python/pull/220
             data = np.asscalar(data)
+        if self.is_object:
+            data = decode_pickle(data)
         return data
 
 
@@ -200,9 +207,14 @@ class NetCDF4DataStore(WritableCFDataStore, DataStorePickleMixin):
     """
     def __init__(self, filename, mode='r', format='NETCDF4', group=None,
                  writer=None, clobber=True, diskless=False, persist=False,
-                 autoclose=False):
+                 autoclose=False, allow_object=False):
         if format is None:
             format = 'NETCDF4'
+
+        assert not allow_object or format.startswith('NETCDF4'), \
+            """serializing native Python objects is only possible with
+            'NETCDF4' format.  Current format is '{}'""".format(format)
+
         opener = functools.partial(_open_netcdf4_group, filename, mode=mode,
                                    group=group, clobber=clobber,
                                    diskless=diskless, persist=persist,
@@ -215,7 +227,22 @@ def __init__(self, filename, mode='r', format='NETCDF4', group=None,
         self._filename = filename
         self._mode = 'a' if mode == 'w' else mode
         self._opener = functools.partial(opener, mode=self._mode)
-        super(NetCDF4DataStore, self).__init__(writer)
+        super(NetCDF4DataStore, self).__init__(writer,
+                                               allow_object=allow_object)
+
+    @WritableCFDataStore.allow_object.setter
+    def allow_object(self, value):
+        self._allow_object = value
+
+    @property
+    def _object_datatype(self):
+        msg = "Object datatype only supported with keyword 'allow_object'"
+        assert self.allow_object, msg
+        dtype = self.ds.vltypes.get('object',
+                                    self.ds.createVLType(np.uint8, 'object'))
+        msg = "Object datatype is '{}'. Should be 'uint8'.".format(dtype.dtype)
+        assert dtype.dtype == np.uint8, msg
+        return dtype
 
     def open_store_variable(self, name, var):
         with self.ensure_open(autoclose=False):
@@ -287,17 +314,25 @@ def set_variables(self, *args, **kwargs):
         with self.ensure_open(autoclose=False):
             super(NetCDF4DataStore, self).set_variables(*args, **kwargs)
 
+    def _extract_variable_and_datatype(self, variable):
+        if self.format == 'NETCDF4':
+            if variable.dtype.kind == 'O' and self.allow_object:
+                datatype = self._object_datatype
+                variable.data = encode_pickle(variable)
+            else:
+                variable, datatype = _nc4_values_and_dtype(variable)
+        else:
+            variable = encode_nc3_variable(variable)
+            datatype = variable.dtype
+        return variable, datatype
+
     def prepare_variable(self, name, variable, check_encoding=False,
                          unlimited_dims=None):
         attrs = variable.attrs.copy()
 
         variable = _force_native_endianness(variable)
 
-        if self.format == 'NETCDF4':
-            variable, datatype = _nc4_values_and_dtype(variable)
-        else:
-            variable = encode_nc3_variable(variable)
-            datatype = variable.dtype
+        variable, datatype = self._extract_variable_and_datatype(variable)
 
         self.set_necessary_dimensions(variable, unlimited_dims=unlimited_dims)
 
diff --git a/xarray/backends/scipy_.py b/xarray/backends/scipy_.py
index ab896043db7..18d003d261e 100644
--- a/xarray/backends/scipy_.py
+++ b/xarray/backends/scipy_.py
@@ -90,7 +90,7 @@ class ScipyDataStore(WritableCFDataStore, DataStorePickleMixin):
     It only supports the NetCDF3 file-format.
     """
     def __init__(self, filename_or_obj, mode='r', format=None, group=None,
-                 writer=None, mmap=None, autoclose=False):
+                 writer=None, mmap=None, autoclose=False, allow_object=False):
         import scipy
         import scipy.io
 
@@ -122,7 +122,7 @@ def __init__(self, filename_or_obj, mode='r', format=None, group=None,
         self._opener = opener
         self._mode = mode
 
-        super(ScipyDataStore, self).__init__(writer)
+        super(ScipyDataStore, self).__init__(writer, allow_object=allow_object)
 
     def open_store_variable(self, name, var):
         with self.ensure_open(autoclose=False):
diff --git a/xarray/conventions.py b/xarray/conventions.py
index d39ae20925a..42965f8a629 100644
--- a/xarray/conventions.py
+++ b/xarray/conventions.py
@@ -663,7 +663,7 @@ def maybe_encode_bools(var):
     return var
 
 
-def _infer_dtype(array):
+def _infer_dtype(array, allow_object=False):
     """Given an object array with no missing values, infer its dtype from its
     first element
     """
@@ -676,12 +676,15 @@ def _infer_dtype(array):
             # the length of their first element
             dtype = np.dtype(dtype.kind)
         elif dtype.kind == 'O':
-            raise ValueError('unable to infer dtype; xarray cannot '
-                             'serialize arbitrary Python objects')
+            if allow_object:
+                dtype = np.dtype(object)
+            else:
+                raise ValueError('unable to infer dtype; xarray cannot '
+                                 'serialize arbitrary Python objects')
     return dtype
 
 
-def ensure_dtype_not_object(var):
+def maybe_infer_dtype(var, allow_object=False):
     # TODO: move this from conventions to backends? (it's not CF related)
     if var.dtype.kind == 'O':
         dims, data, attrs, encoding = _var_as_tuple(var)
@@ -689,7 +692,8 @@ def ensure_dtype_not_object(var):
         if missing.any():
             # nb. this will fail for dask.array data
             non_missing_values = data[~missing]
-            inferred_dtype = _infer_dtype(non_missing_values)
+            inferred_dtype = _infer_dtype(non_missing_values,
+                                          allow_object=allow_object)
 
             if inferred_dtype.kind in ['S', 'U']:
                 # There is no safe bit-pattern for NA in typical binary string
@@ -706,12 +710,13 @@ def ensure_dtype_not_object(var):
             data = np.array(data, dtype=inferred_dtype, copy=True)
             data[missing] = fill_value
         else:
-            data = data.astype(dtype=_infer_dtype(data))
+            data = data.astype(dtype=_infer_dtype(data,
+                                                  allow_object=allow_object))
         var = Variable(dims, data, attrs, encoding)
     return var
 
 
-def encode_cf_variable(var, needs_copy=True, name=None):
+def encode_cf_variable(var, needs_copy=True, name=None, allow_object=False):
     """
     Converts an Variable into an Variable which follows some
     of the CF conventions:
@@ -725,6 +730,9 @@ def encode_cf_variable(var, needs_copy=True, name=None):
     ----------
     var : xarray.Variable
         A variable holding un-encoded data.
+    allow_object : bool
+        Whether to allow objects to pass the encoder or to throw an error if
+        this is attempted.
 
     Returns
     -------
@@ -738,7 +746,7 @@ def encode_cf_variable(var, needs_copy=True, name=None):
     var = maybe_encode_dtype(var, name)
     var = maybe_default_fill_value(var)
     var = maybe_encode_bools(var)
-    var = ensure_dtype_not_object(var)
+    var = maybe_infer_dtype(var, allow_object=allow_object)
     return var
 
 
@@ -1058,7 +1066,7 @@ def encode_dataset_coordinates(dataset):
                                non_dim_coord_names=non_dim_coord_names)
 
 
-def cf_encoder(variables, attributes):
+def cf_encoder(variables, attributes, allow_object=False):
     """
     A function which takes a dicts of variables and attributes
     and encodes them to conform to CF conventions as much
@@ -1075,6 +1083,9 @@ def cf_encoder(variables, attributes):
         A dictionary mapping from variable name to xarray.Variable
     attributes : dict
         A dictionary mapping from attribute name to value
+    allow_object : bool
+        Whether to allow objects to pass the encoder or to throw an error if
+        this is attempted.
 
     Returns
     -------
@@ -1085,6 +1096,7 @@ def cf_encoder(variables, attributes):
 
     See also: encode_cf_variable
     """
-    new_vars = OrderedDict((k, encode_cf_variable(v, name=k))
+    new_vars = OrderedDict((k, encode_cf_variable(v, name=k,
+                                                  allow_object=allow_object))
                            for k, v in iteritems(variables))
     return new_vars, attributes
diff --git a/xarray/core/dataarray.py b/xarray/core/dataarray.py
index c12983c20a2..c8a9fe0feaf 100644
--- a/xarray/core/dataarray.py
+++ b/xarray/core/dataarray.py
@@ -1314,6 +1314,8 @@ def to_netcdf(self, *args, **kwargs):
             Nested dictionary with variable names as keys and dictionaries of
             variable specific encodings as values, e.g.,
             ``{'my_variable': {'dtype': 'int16', 'scale_factor': 0.1, 'zlib': True}, ...}``
+        allow_object : bool, optional
+            If True, allow native Python objects to be serialized.
 
         Notes
         -----
diff --git a/xarray/core/dataset.py b/xarray/core/dataset.py
index ae5499a46a7..43fd1838bbb 100644
--- a/xarray/core/dataset.py
+++ b/xarray/core/dataset.py
@@ -898,6 +898,7 @@ def reset_coords(self, names=None, drop=False, inplace=False):
     def dump_to_store(self, store, encoder=None, sync=True, encoding=None,
                       unlimited_dims=None):
         """Store dataset contents to a backends.*DataStore object."""
+
         if encoding is None:
             encoding = {}
         variables, attrs = conventions.encode_dataset_coordinates(self)
@@ -918,7 +919,8 @@ def dump_to_store(self, store, encoder=None, sync=True, encoding=None,
             store.sync()
 
     def to_netcdf(self, path=None, mode='w', format=None, group=None,
-                  engine=None, encoding=None, unlimited_dims=None):
+                  engine=None, encoding=None, unlimited_dims=None,
+                  allow_object=False):
         """Write dataset contents to a netCDF file.
 
         Parameters
@@ -968,13 +970,16 @@ def to_netcdf(self, path=None, mode='w', format=None, group=None,
             By default, no dimensions are treated as unlimited dimensions.
             Note that unlimited_dims may also be set via
             ``dataset.encoding['unlimited_dims']``.
+        allow_object : bool, optional
+            If True, allow native Python objects to be serialized.
         """
         if encoding is None:
             encoding = {}
         from ..backends.api import to_netcdf
         return to_netcdf(self, path, mode, format=format, group=group,
                          engine=engine, encoding=encoding,
-                         unlimited_dims=unlimited_dims)
+                         unlimited_dims=unlimited_dims,
+                         allow_object=allow_object)
 
     def __unicode__(self):
         return formatting.dataset_repr(self)
diff --git a/xarray/core/utils.py b/xarray/core/utils.py
index 89d1462328c..702f2c56bd1 100644
--- a/xarray/core/utils.py
+++ b/xarray/core/utils.py
@@ -9,6 +9,7 @@
 import re
 import warnings
 from collections import Mapping, MutableMapping, Iterable
+from six.moves import cPickle as pickle
 
 import numpy as np
 import pandas as pd
@@ -489,3 +490,13 @@ def ensure_us_time_resolution(val):
     elif np.issubdtype(val.dtype, np.timedelta64):
         val = val.astype('timedelta64[us]')
     return val
+
+
+@functools.partial(np.vectorize, otypes='O')
+def encode_pickle(obj):
+    return np.frombuffer(pickle.dumps(obj), dtype=np.uint8)
+
+
+@functools.partial(np.vectorize, otypes='O')
+def decode_pickle(obj):
+    return pickle.loads(obj.tostring())