On Sun, Jun 9, 2013 at 12:46 PM, Stan Hoeppner <stan@xxxxxxxxxxxxxxxxx> wrote:
> In a recent linux-raid list thread here:
> seriously flawed arguments against the reliability of XFS, and even the
> performance of XFS, are made. The OP even quotes Dave's LCA
> presentation as a performance reason to avoid XFS. The party really
> gets started at paragraph 7.
> I made a brief effort to debunk his claims and explained that he can't
> have O_PONIES, that he should use fsync or O_DIRECT, etc for data
> safety. To non experts/advanced filesystem users, his long winded
> argument may be persuasive. Obviously none of you experts has time to
> debunk every such post, but this one may be worth a read at least,
> especially given the weight Google gives to vger lists.
The really unfortunate thing about this is that the bug which would
prevent transaction flushing from happening got imported and shipped
for a rather long time in RHEL. It's one thing to get a file zeroed
that's a few seconds old, but having the same happen to files which
haven't been touched in hours, even before issuing manual sync, is
certainly not very reassuring.
As a very satisfied user of XFS on CentOS 6, I have been nervous
enough about that to go through the trouble of rebooting our main
server for a kernel upgrade a few weeks ago. Thanks to RedHat's
deceptive tactics regarding kernel patches, I have also not been able
to pin-point the exact range of kernel versions affected by this in a
reasonable amount of time and hence have not found out (thankfully not
the hard way) if it was even necessary.