The *good* tape drives can still stream linear data faster than any random access disk. Its hard convincing people of those statistics; most go with backup-to-disk. It’s like the “mainframe is death” myth rewritten for tapes.

In 2002, we multiplexed several disk clients just to keep the the tape drives spinning. This was despite requiring a dedicated backup NIC and many times a dedicated backup switch that datacenters funneled directly to the backup servers. (This was Netbackup 4.x with Solaris 8 servers in front of ADIC refrigerator-size libraries. LTO of course.) The block sizes used by network and disk filesystems were (and are) just not optimized for transferring large datasets in short time periods. I luv’ it when people don’t get this until they see a VLDB come to life from LTO.

And ditto on test restores. We charged extra to verify with Netbackup so most customers opted out. The customers with disk filesystems that didn’t support snapshots asked for a copy of the SLA/contract when a backup skipped files that were in use, or when their network mangled the data, and it was always an unsympathetic “told you so” moment.

These are age old debates. No one talks about cpio anymore because tar supposedly deprecated it. No; GNU tar (and most versions) don’t ship with internal verification. cpio almost always shipped with CRC via a flip of a switch. The per file verification that ole’ cpio did was more granular than per job/per dataset verification typically done via md5/sha1 by the big boys. Of course, restoring the whole backup client is best but nobody does that.

Bacula has very Netbackup-like tests for backup jobs, supports snapshots, supports a lot of OSes, and is free. Still use that and cpio against ole DDS