Garret Wilson had a terrible experience upgrading to SVN 1.9 (http://subversion.tigris.org/issues/show_bug.cgi?id=4588), ranging from build problems over svnadmin load inconsistency to a low-level FSFS format 7 read error.
This post focuses on the FSFS issue. The good news is that confidence is high that the repo itself is correct and there has been no apparent data loss. The error message indicates that something like Offset=L2P[L2P[Item]] or Offset=L2P[Offset] instead of Offset=L2P[item] happened. Trying to reproduce the issue, I discovered bugs in our debug logging code (only affects FSFS devs) and in svnfsfs load-index (fix yet to be committed, work-around is possible). None of these explain the behavior Garret has been seeing. I've also been looking for places where svn_fs_fs__item_offset() gets called but those invocations seem to never possibly do a double lookup. My current hypothesis is that the server did not get restarted after replacing the repository. Because we decided not to make the instance ID part of the cache key, we could easily have picked up cached format 6 data for the format 7 repository. The lookup would start with a hit at the L2 DAG cache, mapping path+rev onto a noderev. From there, everything is accessed using physical offsets until some item is not cached. At that point, we would be trying to use an Offset instead of the Item index to address the f7 data. Error. -- Stefan^2.
