Here's what I'm thinking: Above the ObjectStore, we clear everything from temp on restart anyway.
We always write bits of a temp object in pieces, and then at the end copy/move it into the main collection. On replay, we should *only* do that final move/rename if the temp object was replayed in its entirety. So: - clear out temp collections in the filestore on startup. - give temp objects unique names so that they don't collide with non-temp object fd caching (or whatever else). for the DBObjectMap part there is probably some futzing though to make this work right. - add a new 'move_from_temp' type operation that renames an object a temp (coll_t::is_temp()) collection to a non-temp one. it will succeed iff the temp source exists. - all operations that write to temp objects fail if the object doesn't already exist, except an explicit 'create' op - all transactions the osds generate that write to temp object start with that explicit create. The combination of these thigns means that we will only have a temp source for the move_from_temp op if it is complete. Which I think means we can avoid any of the fsync guard stuff entirely. The DBObjectMap I'm very fuzzy on, so I suspect that's where the tricky part will be. Maybe the temp object name includes the intended hobject_t in it somewhere, or something, so that the rename can be reflected in leveldb at the end. Thoughts? Maybe we can do a quick hangout this afternoon to make sure this will work before I start putting it together... sage -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
