diff --git a/common/column_store.py b/common/column_store.py
index 17ae16d34e..6f162db1b1 100644
--- a/common/column_store.py
+++ b/common/column_store.py
@@ -6,7 +6,7 @@ from contextlib import closing
 from common.file_helpers import mkdirs_exists_ok
 
 class ColumnStoreReader():
-  def __init__(self, path, mmap=False, allow_pickle=False, direct_io=False):
+  def __init__(self, path, mmap=False, allow_pickle=False, direct_io=False, prefix="", np_data=None):
     if not (path and os.path.isdir(path)):
       raise ValueError("Not a column store: {}".format(path))
 
@@ -15,6 +15,9 @@ class ColumnStoreReader():
     self._mmap = mmap
     self._allow_pickle = allow_pickle
     self._direct_io = direct_io
+    self._is_dict = 'columnstore' in self._keys
+    self._prefix = prefix
+    self._np_data = np_data
 
   @property
   def path(self):
@@ -30,6 +33,11 @@ class ColumnStoreReader():
       return None
 
   def keys(self):
+    if self._is_dict:
+      if not self._np_data:
+        self._np_data = self._load(os.path.join(self._path, 'columnstore'))
+      return self._np_data.keys()
+
     return list(self._keys)
 
   def iteritems(self):
@@ -49,25 +57,42 @@ class ColumnStoreReader():
     else:
       return None
 
+  def _load(self, path):
+    if self._mmap:
+      # note that direct i/o does nothing for mmap since file read/write interface is not used
+      return np.load(path, mmap_mode='r', allow_pickle=self._allow_pickle, fix_imports=False)
+
+    if self._direct_io:
+      opener = lambda path, flags: os.open(path, os.O_RDONLY | os.O_DIRECT)
+      with open(path, 'rb', buffering=0, opener=opener) as f:
+        return np.load(f, allow_pickle=self._allow_pickle, fix_imports=False)
+
+    return np.load(path, allow_pickle=self._allow_pickle, fix_imports=False)
+
   def __getitem__(self, key):
     try:
-      path = os.path.join(self._path, key)
+      if self._is_dict:
+        path = os.path.join(self._path, 'columnstore')
+      else:
+        path = os.path.join(self._path, key)
 
       # TODO(mgraczyk): This implementation will need to change for zip.
-      if os.path.isdir(path):
+      if not self._is_dict and os.path.isdir(path):
         return ColumnStoreReader(path)
       else:
-        if self._mmap:
-          # note that direct i/o does nothing for mmap since file read/write interface is not used
-          ret = np.load(path, mmap_mode='r', allow_pickle=self._allow_pickle, fix_imports=False)
+        if self._is_dict and self._np_data:
+          ret = self._np_data
         else:
-          if self._direct_io:
-            opener = lambda path, flags: os.open(path, os.O_RDONLY | os.O_DIRECT)
-            with open(path, 'rb', buffering=0, opener=opener) as f:
-              ret = np.load(f, allow_pickle=self._allow_pickle, fix_imports=False)
-          else:
-            ret = np.load(path, allow_pickle=self._allow_pickle, fix_imports=False)
+          ret = self._load(path)
+
         if type(ret) == np.lib.npyio.NpzFile:
+          if self._is_dict:
+            if self._prefix+key in ret.keys():
+              return ret[self._prefix+key]
+            if any(k.startswith(self._prefix+key+"/") for k in ret.keys()):
+              return ColumnStoreReader(self._path, mmap=self._mmap, allow_pickle=self._allow_pickle, direct_io=self._direct_io, prefix=self._prefix+key+"/", np_data=ret)
+            raise KeyError(self._prefix+key)
+
           # if it's saved as compressed, it has arr_0 only in the file. deref this
           return ret['arr_0']
         else:
@@ -128,6 +153,27 @@ class ColumnStoreWriter():
     # TODO(mgraczyk): This implementation will need to change if we add zip or compression.
     return ColumnStoreWriter(os.path.join(self._path, group_name))
 
+  def add_dict(self, data, dtype=None, compression=False, overwrite=False):
+    # default name exists to have backward compatibility with equivalent directory structure
+    npy_path = os.path.join(self._path, "columnstore")
+    mkdirs_exists_ok(os.path.dirname(npy_path))
+
+    flat_dict = dict()
+    _flatten_dict(flat_dict, "", data)
+    for k, v in flat_dict.items():
+      flat_dict[k] = np.array(v, copy=False, dtype=dtype)
+
+    if overwrite:
+      f = open(npy_path, "wb")
+    else:
+      f = os.fdopen(os.open(npy_path, os.O_WRONLY | os.O_CREAT | os.O_EXCL), "wb")
+
+    with closing(f) as f:
+      if compression:
+        np.savez_compressed(f, **flat_dict)
+      else:
+        np.savez(f, **flat_dict)
+
   def close(self):
     pass
 
@@ -135,6 +181,15 @@ class ColumnStoreWriter():
   def __exit__(self, type, value, traceback): self.close()
 
 
+def _flatten_dict(flat, ns, d):
+  for k, v in d.items():
+    p = (ns + "/" if len(ns) else "") + k
+    if isinstance(v, collections.Mapping):
+      _flatten_dict(flat, p, v)
+    else:
+      flat[p] = v
+
+
 def _save_dict_as_column_store(values, writer, compression):
   for k, v in values.items():
     if isinstance(v, collections.Mapping):
@@ -147,4 +202,3 @@ def save_dict_as_column_store(values, output_path, compression=False):
   with ColumnStoreWriter(output_path) as writer:
     _save_dict_as_column_store(values, writer, compression)
 
-