[LU-11007] Upgrading Lustre 2.8fe to 2.10.x - on-disk format change question

• Type: Bug
• Resolution: Done
• Priority: Minor
• Affects Version/s: Lustre 2.10.5
• Labels: llnl
• Severity: 3
• Assignee: Peter Jones
• Reporter: Olaf Faaland

Description

Are there on-disk format changes between Lustre 2.8fe and Lustre 2.10.x, even if one avoids use of new features to 2.10 like PFL and project quota, and avoids DNE2?

IE if we were to mount a Lustre 2.8fe file system with Lustre 2.10 kernel modules, and perform some mdtests and IORs, would there be on-disk changes that prevent the file system from being mounted again with 2.8fe kernel modules?

Is the answer different for ldiskfs than for ZFS backend?

The context would be performing tests before releasing the file system to users, after upgrading the software from 2.8fe to 2.10.x., and falling back to 2.8fe if the tests reveal a problem.

Comments

[Andreas Dilger]

One possible issue is the addition of the ZFS native dnode quota accounting (which is a zpool feature that needs to be enabled explicitly), which is enabled with ZFS 0.7.0 and Lustre 2.10. However, I don't think you use the Lustre dnode accounting because it was not working well before that update in any case. Similarly, the multimount feature is new in ZFS 0.7 and Lustre 2.10, but AFAIK you are already using ZFS 0.7, and it would need to be explicitly enabled in any case.

If you don't use project quota, PFL layouts, or DNE-2 there shouldn't be any issues downgrading. There is still an "mdt.*.enable_remote_dir" tunable that controls whether DNE striped or remote directories can be created by regular users or not, but it controls both DNE-1 and DNE-2 directories. There is not a tunable to prevent the creation of PFL files if some user tried to do that, but that would only affect those specific files from being accessed if the system was downgraded again.

[Peter Jones]

Olaf

While we have not tested this ourselves it seems possible that it would work. It seems quite likely that this will work fine for ZFS but we are less sure about ldiskfs so if you decide to explore this option you should focus your testing there.

Regards

Peter

[Olaf Faaland]

This would be a last resort, but it would be useful to know if it's even worth trying. That's especially true with ldiskfs since we cannot create snapshots.