Ceph » RADOS

If, during peering, last_update moves backwards, we may rewrite the full info but leave a fastinfo record in place with a newer last_update.

From ML:

In the scenario of EC deployment, suppose we done a peering process for a pg and
changed one shard's last_update from lu1(e1'3) to lu2(e1'2) .lu1 was written as
fastinfo and lu2 was written as info. After that we restarted this osd and
loaded pgs again. when we read pg info from disk, we will find the pg info is
lu1 applied to lu2, which becomes incorrect. the true value should be lu2. That
may cause the coming peering execute incorrectly and result in unfound objects.
I currently considered below two options:
1. delete fastinfo when we need to change info;
2. add extra sequence number to fastinfo and info structure to make it keep them
in the right order.