On 07.10.2011 21:20, Stan Hoeppner wrote:
If I may make a purely subjective comment: 2.5m spooled emails on a single host is insane.
I'm not arguing that. In the end the system is supposed to cope with 300k mails in 24h, balanced on two servers, which I think can be achieved without a lot of tuning. This was merely a fat-fingered burn-in test, which failed in a way I did not expect. And I think it is a nice DoS vector, probably for most installations.
I'm testing ext4 now (which was not easy, given how hard it is to use this filesystem on SLES11.1 at all) and it's worse in the default settings. In the default settings it created only around 650k inodes (one inode for every four blocks of 4k), after inode exhaustion it fails the same horrible deadlock way XFS did. On the pro side, "df -i" is reliable, so you can actually monitor the situation before it happens.
Of course it's admin 101 you can tune the filesystem. ext4 with one inode per block works nicely, cannot run out of inodes and subjectively feels a lot faster than xfs. I think I need to benchmark it again.
Again, this is not postfix' fault, I don't see how it could handle running out-of-inodes gracefully at all. One could check for free inodes (something like queue_mininodes), but this would have failed in the XFS case, too. It's just a filesystem problem you might see more often with mailservers.
Bernhard
