HDFS-4305, which was fixed in 2.1.0-beta, clients could specify arbitrarily small preferred block size for a file including 0. This was normally done by faulty clients or failed creates, but it was possible.
Until 2.5, reading a fsimage containing inodes with 0 byte preferred block size was allowed. So if a fsimage contained such an inode, the namenode would come up fine. In 2.6, the preferred block size is required be > 0. Because of this change, the image that worked with 2.5 may not work with 2.6.
If a cluster ran a version of hadoop earlier than 2.1.0-beta before, it is under this risk even if it worked fine with 2.5.