[premise-users] extending downtime today

[Robert E. Anderson]

Things are better this morning.

For clarification the problem we've been working on is only on the
secondary metadata mirror for the BeeGFS volume mounted on Premise as
"/mnt/home".  There is also a failed SSD drive in the "old Lustre"
storage array mounted on Premise as "/mnt/lustre".  A replacement Lustre
SSD drive was previously purchased and was on campus late Monday, we
should have the new drive in Morse today.



Last nights 2nd fix completely solved the BeeGFS metadata resync error.
It took about 6 hours to complete, but we now have a good BeeGFS
metadata mirror.

The original metadata backup process ran out of memory (more than a
day in).  The rsync backup of the same data is still
running, just taking longer than predicted to complete using the size
of data on the destination.

The current plan is to:

1. Wait for the rsync backup to complete (best guess is 2 more hours)

2. Bring Premise back online for users.

3. Start an beegfs-fsck (BeeGFS-File System ChecK) on the live BeeGFS
storage (which will take longer than the offline beegfs-fsck, possibly 
days to complete).

Ideally we can also replace the failed SSD drive in the "old Lustre"
metadata storage while the rsync completes.

We are holding off on migrating the BeeGFS secondary metadata server to
ZFS until after a clean beegfs-fsck.  Knowing that it takes 6 hours to
resync, makes doing this work first just extend further this
unscheduled downtime.  The rsync ZFS destination is also showing us how
much space our metadata consumes on a ZFS storage target.  So we are
gaining insights on this option. 

I'll provide either a messages that the system is back up, or yet
another status on or before 3pm.


On Tue, 13 Oct 2020 09:39:13 -0400
Robert Anderson <rea at sr.unh.edu> wrote:

> The metadata storage problem is proving more difficult to resolve. I
> hate keeping Premise down longer, but it's a lot better than having
> to say we lost all your stored data.
> 
> The issue is that the primary metadata server is failing to mirror
> it's data to the secondary. We beleive a full parallel file system
> check is needed, but that fails when the secondary metadata mirror is
> out of sync. A catch 22, that has no simple solution.
> 
> The underlying low level file storage currently utilizes a hardware 
> mirrored pair of SSD drives. This lowest level storage passes
> individuall file system checks.
> 
> To ensure the safety of the stored data we want to get at least one
> good offline copy of the current metadata. That process is going much
> slower than anticipated. Since we have no successful backup of this
> metadata, we have nothing to base a completion estimate on.
> 
> Once the backup safety net is in place we need to: reconfigure things
> to remove the secondary metadata mirror, perform the parallel file
> system check, and then add back the secondary meta data sever.
> Getting a good backup requires no users. The parallel file system
> check and re-mirror would be faster without users, but can also be
> done with normal user workloads.
> 
> While we are down we will also try to reconfigure the storage on the
> idle server to avoid this meta data backup issue in the future.
> During active use the metadata is always changing, and backups take
> many hours to complete. So we can never able capture a stable backup
> with active users on Premise.  Changing the low level storage to a
> file system that supports snapshots and replication, will provide us
> with a few viable backup options moving forward.
> 
> I will send another status update email this evening.
> 
> On October 12, 2020 17:17:45 Robert Anderson <rea at sr.unh.edu> wrote:
> > There was a filesystem error on Premise /mnt/home.
> >
> > A filesystem check on the main /mnt/home space needs to complete
> > without active users.
> >
> > Hopefully the system will be back online later this evening or in
> > the morning if the check takes all night.
> >
> >
> > On Mon, 2020-10-12 at 08:33 -0400, Thomas J. Baker wrote:
> >> We lost chilled water this morning and the Premise cluster has
> >> been shut down. The back up chiller appears to be misbehaving as
> >> well so no ETA yet on Premise return to service.
> >>
> >> Thanks,
> >>
> >> tjb
> >> --
> >> =======================================================================
> >>> Thomas Baker |
> >>> Systems Programmer email:
> >>> tjb at unh.edu
> >>> |
> >>> Research Computing Center office: Morse 206 |
> >>> University of New Hampshire voice: (603) 862-4490 |
> >>> 213 Morse Hall fax: (603) 862-1761 |
> >>> Durham, NH 03824 USA |
> >> =======================================================================
> >>
> >>
> >> _______________________________________________
> >> premise-users mailing list
> >> premise-users at lists.sr.unh.edu
> >>
> >> https://lists.sr.unh.edu/mailman/listinfo/premise-users
> 



-- 
Robert E. Anderson
Associate Director
UNH Research Computing Center