Ok, so I just lost 40 GB of headers.. And it wasn't even by doing a stupid/uninformed selection (I would have been OK with that). I had the admin->control page open, monitoring load. I reach for the mouse, not looking at it, and manage accidentally to click the left mouse button with my pinky finger (misjudging where the mouse were, slightly). It slides slightly at the same time and the unlikely result (1 in 1000? 100000?) is that "Purge all newsgroups" gets selected. 0.5 seconds later all data is gone. Or, should be, but because of some bug URDD exits and a few groups are left untoched, leaving me with 23 GB of the 65+ I had. (There was one active set generating that had been running for a few hours, and one header update, so my guess is that one of those caused URDD to bomb when it wanted to nuke all data...)
I can only be happy this didn't happen when I have 300 GB of headers (and of course I will backup the data eventually). Still, a fatal command like this should ask for confirmation! I know this story sounds to freaky to be true, but it actually just happened. It'll take med a couple of days to get back to where I were so it's not disaster, but not fun. There are a few more commands that really should ask for confirmation too, but this is THE worst, by far.
Oh well. I guess my misfortune will make the product better in the end

EDIT: ...And there went the rest. I updated the one big group that was saved because of the bug described above. After it was done after a minute or so, it generated sets, but only for newly downloaded headers (195.000) (probably some counter was off, even though the groups list showed the correct numbers). Then "expire incomplete sets" kicked in. Poof, 53.000.000 gone.
Some days one should just stay in bed (or use LVM snapshots more often...). Jebus.
Oh well. Maybe it was for the best. On the flip side, I got to see that "optimize database" on 23GB take about 7 mins - this took hours before so at least the setup is fairly good now.
I'll move on to 1.0.4 instead and start over. I guess the DB format won't change from current SVN to 1.0.4, so this should be "safe", right?