Re: Search over referrals (Re: LAST CALL: draft-ietf-ldapext-referral-00.txt)

[Leif Hedstrom <leif@netscape.com>]

David Boreham wrote:

>  Alan Lloyd wrote:
>
>>         Well no - some systems use files and RAM cache to get
>> performance. So I think that customers should know that and do a
>> "pull
>> the plug test" and see what the restart times are and what state the
>> DIB
>> is in afterwards.
>
> Good idea. Next time you tell a customer to do this, also tell
> them to make sure the server is subject to write operations
> while they pull the plug---it makes the test rather more realistic :)

I think this thread is about databases using a file system for data
storage, right? I'm no DB expert, but I believe several modern file
systems can improve your database's reliability significantly, if they
use the file system for storage (and not the raw device). I know
Solaris-2.6 comes with a "direct-I/O" option, to bypass memory caching
etc. (basically synchronous I/O I think?).

I'm personally biased towards the Veritas VxFS file system. It's very
fast, and more importantly, very robust. The best feature with VxFS is
that I can tune a particular file system for say optimal reliability and
consistency, sacrificing performance. The intent log feature means
recovering after a crash is much (*much*) faster then UFS, which is also
an important consideration (e.g. compare 3 minutes downtime to ~1/2 hour
to fsck a disk).

We use VxFS internally for all mission critical servers, including our
two "main" LDAP servers. I'd never use UFS again on something that needs
to survive an "ACID" test. :).

Just my $.02,

-- Leif