Can one of the admins verify this patch?

ok to test

@sunfch on the surface the suggestion is fine. Note that there might be a case where some objects will miss the commit entry, e.g., if rgw crashed between the prepare and the commit. There should be some mechanism to handle that (e.g., handle entries that were in the prepare state but without commit following).

@yehudasa if rgw crashed between the moment that a object is uploaded completely and the commit entry is not still written to disk, radosgw-agent that applied this pull request will miss the object. Have I understood correctly?
If the entry in prepare state and the entry in complete state for the object are all exist, the prepare entry should been ignored. If there is no entry in complete state for a object, how should the entry in prepare state for the object be processed? Have you good idea for that?
Thanks

Good job @sunfch !! I meet the same problem as you did, the usage of secondary zone is twice to the master.
I have one concern to your change, is there any necessary to add entry.op == 'write' into if statement? there are also duplicated log entries with op =='del' which will cause an object to be delete twice, and this change will not take effect to them.

@goukiwang
You are right!
I did not delete objects at master zone, so I did not run into that case.