Use StringIO for pure python legacy batch inserts.

This reduces memory fragmentation when encoding large batches of documents for insert. Testing shows improvements in memory usage in pure python and reduced GC overhead in Jython with no reduction in performance.

Author: behackett

pymongo/message.py

@@ -219,10 +219,10 @@ def _insert_message(insert_message, send_safe):
 
  send_safe = safe or not continue_on_error
  last_error = None
- begin = struct.pack("<i", int(continue_on_error))
- begin += bson._make_c_string(collection_name)
- message_length = len(begin)
- data = [begin]
+ data = StringIO()
+ data.write(struct.pack("<i", int(continue_on_error)))
+ data.write(bson._make_c_string(collection_name))
+ message_length = begin_loc = data.tell()
  has_docs = False
  for doc in docs:
  encoded = bson.BSON.encode(doc, check_keys, uuid_subtype)
@@ -231,14 +231,14 @@ def _insert_message(insert_message, send_safe):
 
  message_length += encoded_length
  if message_length < client.max_message_size and not too_large:
- data.append(encoded)
+ data.write(encoded)
  has_docs = True
  continue
 
  if has_docs:
  # We have enough data, send this message.
  try:
- client._send_message(_insert_message(_EMPTY.join(data),
+ client._send_message(_insert_message(data.getvalue(),
  send_safe), send_safe)
  # Exception type could be OperationFailure or a subtype
  # (e.g. DuplicateKeyError)
@@ -261,13 +261,15 @@ def _insert_message(insert_message, send_safe):
  " bytes." %
  (encoded_length, client.max_bson_size))
 
- message_length = len(begin) + encoded_length
- data = [begin, encoded]
+ message_length = begin_loc + encoded_length
+ data.seek(begin_loc)
+ data.truncate()
+ data.write(encoded)
 
  if not has_docs:
  raise InvalidOperation("cannot do an empty bulk insert")
 
- client._send_message(_insert_message(_EMPTY.join(data), safe), safe)
+ client._send_message(_insert_message(data.getvalue(), safe), safe)
 
  # Re-raise any exception stored due to continue_on_error
  if last_error is not None: