It should not be very sensitive to drift. I think timestamp is meant mostly as an increasing number. implementation might have a bug. will check.
Could paste a larger portion of the log? Raghu. On Tue, Oct 13, 2009 at 10:47 PM, Jason Venner <[email protected]>wrote: > The actual body of the error contains a statement to the effect that the > > Namenode has an edit log with timestamp of ..... but new checkpoint > was created using editlog with timestamp .... > > This is from FSImage.validateCheckpointUpload. The drift is about the difft > in time in the edit logs. > > THe unique situation in my experience here is that one of the dfs.name.dir > directories is on an nfs mounted file system, and that particular file has a > mode time that is often 0.5+/-0.5 seconds different. > > Any simple fixes? Even with the best ntp, we are going to be +/- a few msec > and the test in validateCheckpointUpoad is exact equals comparison. > The other thing that confuses me is why, the time stamp is differenet, the > code from a reading, appears to always pick the same file for modtime > comparisons. > -- > Pro Hadoop, a book to guide you from beginner to hadoop mastery, > http://www.amazon.com/dp/1430219424?tag=jewlerymall > www.prohadoopbook.com a community for Hadoop Professionals >
