When logging an OP_CLOSE to the edit log, the NN writes out an updated file mtime and atime. However, when reading in an OP_CLOSE from the edit log, the NN does not apply these values to the in-memory FS data structure. Because of this, a file's mtime or atime may appear to go back in time after an NN restart, or an HA failover.
Most of the time this will be harmless and folks won't notice, but in the event one of these files is being used in the distributed cache of an MR job when an HA failover occurs, the job might notice that the mtime of a cache file has changed, which in MR2 will cause the job to fail with an exception like the following:
Credit to Sujay Rau for discovering this issue.