On Thu, Jul 3, 2008 at 3:09 PM, Aaron Blew <[EMAIL PROTECTED]> wrote: > My take is that since RAID-Z creates a stripe for every block > (http://blogs.sun.com/bonwick/entry/raid_z), it should be able to > rebuild the bad sectors on a per block basis. I'd assume that the > likelihood of having bad sectors on the same places of all the disks > is pretty low since we're only reading the sectors related to the > block being rebuilt. It also seems that fragmentation would work in > your favor here since the stripes would be distributed across more of > the platter(s), hopefully protecting you from a wonky manufacturing > defect that causes UREs on the same place on the disk. > > -Aaron
The per-block statement above is important - zfs will only rebuild the blocks that have data. A 100TB pool with 1 GB in use will rebuild 1 GB. As such, it is more a factor of the amount of data rather than the size of the RAID device. A periodic zpool scrub will likely turn up read errors before you have a drive failure AND unrelated read errors. Since ZFS merges the volume management and file system layers such an uncorrectable read would turn into zfs saying "file /a/b/c is corrupt - you need to restore it" rather than traditional RAID5 saying "this 12 TB volume is corrupt - restore it". ZFS already makes multiple copies of metadata so if you were "lucky" and the corruption happens to the metadata it should be able to get a working copy from elsewhere. Of course, raidz2 further decreases your chances of losing data. I would highly recommend reading Richard Elling's comments in this area. For example: http://blogs.sun.com/relling/entry/raid_recommendations_space_vs_mttdl http://blogs.sun.com/relling/entry/zfs_raid_recommendations_space_performance http://blogs.sun.com/relling/entry/a_story_of_two_mttdl http://opensolaris.org/jive/thread.jspa?threadID=65564#255257 -- Mike Gerdts http://mgerdts.blogspot.com/ _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss