[1619] in SIPB-AFS-requests
Avoiding rebuilding the VLDB
daemon@ATHENA.MIT.EDU (ghudson@MIT.EDU)
Wed Nov 16 17:12:13 1994
From: ghudson@MIT.EDU
Date: Wed, 16 Nov 94 17:10:21 -0500
To: sipb-afsreq@MIT.EDU
Cc: usenet@MIT.EDU
Chad brought up an interesting point. We should be able to do avoid
rebuilding the VLDB and avoid causing any outages as follows:
1. Put the 4GB and 2GB disk on opus, bring it up as new-ronald-ann.
2. Migrate ronald-ann data to new-ronald-ann.
3. Remove ronald-ann from server CellServDBs, take it down.
4. Bring up ronald-ann as new-rosebud.
5. Migrate rosebud data to new-rosebud (leaving behind vlserver).
6. Update client CellServDBs to point at 18.181.
7. Swap names new-rosebud/rosebud and new-ronald-ann/ronald-ann
8. Take down rosebud after a few weeks.
Step 6 is a little unclear to me, since clients will try to resolve
the hostnames before looking at the IP addresses, so there may be
timeouts on either side of the name switch; I'm not sure. I guess if
we just change client CellServDBs to:
18.181.0.XX #rosebud.mit.edu
18.181.0.YY #ronald-ann.mit.edu
and XX < YY, then we're set to do the name switch without timeouts.
If clients resolve the names, they choose rosebud's old IP address and
win; if clients use the IP address, they choose one of the new servers
and win.
My only nervousness about this is that, until step 6, client
CellServDBs point to ronald-ann as a volume location server, and there
isn't one there. I don't think this will cause any problems, though.
I think as long as ronald-ann's disks stay with ronald-ann, the disks
go to the right places (the disk currently on opus, and the extra disk
on picayune go to news). This does restrict us to waiting for the
RTFM cutover before cutting over news, but if we take a disk off of
ronald-ann before bringing it up as new-rosebud, then we have an
outage when we put a 1GB disk back on.
We only have one Delni port free right now, which is kind of a drag.
We can bring down sipb-server-1 and bring it up as opus on 18.70,
since it will be changing its IP address during the news cutover
anyway.
