We've seen bugs in the past with the order of fsync calls required for durability across system crash. A way we can test for bugs like that is to make copies of the database directory using the dd with iflags=direct set to bypass the filesystem cache.
We should construct such a test and automate it. We've seen a case recently (see WT-3829( where older versions of WiredTiger saw failures using such a scheme around opening a backup cursor - so we should be sure to test with copies tightly around backup cursor creation.