diff --git a/redisdl.py b/redisdl.py
index 4ffd360..726f2f1 100755
--- a/redisdl.py
+++ b/redisdl.py
@@ -141,18 +141,18 @@ def dumps(host='localhost', port=6379, password=None, db=0, pretty=False,
 class BytesWriteWrapper(object):
     def __init__(self, stream):
         self.stream = stream
-        
+
     def write(self, str):
         return self.stream.write(str.encode())
 
 def dump(fp, host='localhost', port=6379, password=None, db=0, pretty=False,
          unix_socket_path=None, encoding='utf-8', keys='*'):
-    
+
     try:
         fp.write('')
     except TypeError:
         fp = BytesWriteWrapper(fp)
-    
+
     if pretty:
         # hack to avoid implementing pretty printing
         fp.write(dumps(host=host, port=port, password=password, db=db,
@@ -276,28 +276,76 @@ def _read_key(key, r, pretty, encoding):
     return (type, ttl, value)
 
 def _reader(r, pretty, encoding, keys='*'):
-    for encoded_key in r.keys(keys):
-        key = encoded_key.decode(encoding)
-        handled = False
-        for i in range(10):
-            try:
-                type, ttl, value = _read_key(encoded_key, r, pretty, encoding)
-                yield key, type, ttl, value
-                handled = True
-                break
-            except KeyDeletedError:
-                # do not dump the key
-                handled = True
-                break
-            except redis.WatchError:
-                # same logic as key type changed
-                pass
-            except KeyTypeChangedError:
-                # retry reading type again
-                pass
-        if not handled:
-            # ran out of retries
-            raise ConcurrentModificationError('Key %s is being concurrently modified' % key)
+    encoded_keys = r.keys(keys)
+    i = 0
+    while i < len(encoded_keys):
+        for key, type, ttl, value in _read_keys(r, encoded_keys[i:i+10000],
+                pretty=pretty, encoding=encoding):
+            yield key, type, ttl, value
+        i += 10000
+
+def _read_keys(r, encoded_keys, pretty, encoding):
+    decoded_keys = [encoded_key.decode() for encoded_key in encoded_keys]
+    do_keys = decoded_keys
+    retries = 5
+    type_results = None
+    while len(do_keys) > 0 and retries > 0:
+        next_do_keys = []
+        next_type_results = []
+
+        if type_results is None:
+            # first pass, need to get the types.
+            # on subsequent passes we know the types
+            # because the previous pass retrieved them and
+            # found a type mismatch
+            p = r.pipeline()
+            for key in do_keys:
+                p.type(key)
+            type_results = p.execute()
+
+        p = r.pipeline()
+        for i in range(len(do_keys)):
+            key = decoded_keys[i]
+            type = type_results[i].decode('ascii')
+            if type == 'none':
+                # key was deleted by a concurrent operation on the data store.
+                # issue noops so that the number of results does not change
+                p.type(key)
+                p.type(key)
+                p.type(key)
+                continue
+            reader = readers.get(type)
+            if reader is None:
+                raise UnknownTypeError("Unknown key type: %s" % type)
+            reader.send_command(p, key)
+            r.pttl_or_ttl_pipeline(p, key)
+            p.type(key)
+        results = p.execute()
+
+        for i in range(len(do_keys)):
+            key = decoded_keys[i]
+            original_type = type_results[i]
+            if original_type == 'none':
+                # this is where we actually skip a key that was deleted
+                # by concurrent operations
+                continue
+            final_type = results[i*3+2].decode('ascii')
+            if original_type != final_type:
+                # type changed, will retry
+                next_do_keys.append(key)
+                # need to update expected type
+                next_type_results.append(final_type)
+                continue
+            reader = readers.get(original_type)
+            value = reader.handle_response(results[i*3], pretty, encoding)
+            ttl = r.decode_pttl_or_ttl_pipeline_value(results[i*3+1])
+            yield key, final_type, ttl, value
+        retries -= 1
+        do_keys = next_do_keys
+        type_results = next_type_results
+
+    if len(do_keys) > 0:
+        raise ConcurrentModificationError('Keys %s are being concurrently modified' % ', '.join(do_keys))
 
 def _empty(r):
     for key in r.keys():
@@ -372,14 +420,14 @@ def ijson_top_level_items(file, local_streaming_backend):
 class TextReadWrapper(object):
     def __init__(self, fp):
         self.fp = fp
-        
+
     def read(self, *args, **kwargs):
         return self.fp.read(*args, **kwargs).decode()
 
 class BytesReadWrapper(object):
     def __init__(self, fp):
         self.fp = fp
-        
+
     def read(self, *args, **kwargs):
         return self.fp.read(*args, **kwargs).encode('utf-8')