glabel, gpart and zfs confusion.

Peter Maloney peter.maloney at brockmann-consult.de
Sat Feb 25 21:43:34 UTC 2012


Am 25.02.2012 16:56, schrieb Bob Friesenhahn:
> On Sat, 25 Feb 2012, Peter Maloney wrote:
>
>> In Solaris, I've read that the IO system is designed such that a some
>> commands (eg. flush of a partition) does not necessarily flush the
>> disk's write cache... like the command can't move up the chain. So if
>> you put zfs on a partition, you can get data loss (eg. transaction
>> rollback required and probably no corruption).
>
> I wonder where you read that since it seems like bad information?  In
> Solaris, if zfs uses a partition (rather than the whole disk), the
> disk write cache is not enabled by default due to the possibility that
> some other partition uses a legacy filesystem like UFS, which could
> become inconsistent and corrupted if the write cache is enabled.  The
> drawback then becomes that zfs writes are likely to incur more latency.
No idea. I was just trying to point out where this recommendation to
keep it separate comes from... but I don't know the details. But what
you said makes sense. But I am sure that among the random things I read
that sounded semi-credible (eg. by some guy claiming to be a ZFS
engineer), it wasn't only about performance; it was more about
corruption. (but then again, there are lots of doomsayers saying ZFS
will somehow fail you, even though when they explain it, it is usually
user error)

And thanks for your criticism; looking back at this document:
http://www.solarisinternals.com/wiki/index.php/ZFS_Best_Practices_Guide
It looks like they just talk about the cache and not corruption, even if
I look at very old versions of the page. So either what I read before
was likely quite wrong, or just opinion based eg. some bad experience of
some tester or admin.


More information about the freebsd-fs mailing list