Ceph

int HashIndex::_remove(const vector<string> &path,
               const ghobject_t &oid,
               const string &mangled_name) {
  int r;
  r = remove_object(path, oid);
  if (r < 0)
    return r;
  subdir_info_s info;
  r = get_info(path, &info);
  if (r < 0)
    return r;
  info.objs--;
  r = set_info(path, info);
  if (r < 0)
    return r;
  if (must_merge(info)) {
  ...

See above, if we delete an object from the current dirent, a check for whether this dirent's containing objects dropping below the merge_threshold is performed. And if it does, a merge of current dirent to it's parent dirent will be kicked off.

I think this code logic can be a little bit problematic for the following reason:
Suppose we have a dirent-A, say, which consists of:
1. subdir-0( which has 0 subdir and 9 objects)
2. subdir-1( which has 0 subdir and 100 objects)
3. 154 objects

Now we reset merge_threshold to 10 and split_multiplier to 1, and we know if we delete a object from subdir-0, a merge shall occur. As a result, subdir-0 is removed and all its containing objects goes into dirent-A which now has 1 subdir and 162(154 + 8) objects.

However, if we try to add another new object to dirent-A later, it will trigger a split immediately and the subdir-0 is recreated and all it's previous objects are re-claimed.

From the above scenario, we now see the cost of the merge may be very expensive and thus shall be avoided.