Michael.Schuster@XXX>
  • Subject: Re: [zfs-discuss] Performance with tens of thousands of zfs filesystems
  • From: Miles Nordin <4909D6F6.2020806@XXX> (Michael Schuster's message of "Thu, 30 Oct 2008 08:47:02 -0700")
  • References: <4909C32C.6080906@XXX> <4909D6F6.2020806@XXX>
  • User-agent: T-gnus/6.17.2 (based on No Gnus v0.2) SEMI/1.14.6 (Maruoka) FLIM/1.14.7 (Sanjō) APEL/10.6 Emacs/21.4 (alpha--netbsd) MULE/5.0 (SAKAKI)
  • Xref: castrovalva.Ivy.NET mail.list.archive:2302

  • >>>>> "ms" == Michael Schuster <Michael.Schuster@XXX writes:
    
        ms> (oh, btw: wasn't IPS created in part to get away from the
        ms> whole patch ...  ermm ... issue? ;-)
    
    hah!  yeah, it sounds like it's working extremely well for people,
    especially with the ZFS root.  The speed and casualness with which
    they upgrade is stunning.  The combination of IPS, ZFS, and the LiveCD
    idea seems to be just awesome.
    
    but this is just UI-optimisation and showmanship.  IPS is the small
    part of what people like---it could almost be just 'tar', like BSD.
    I do not like IPS.
    
    I concede the idea of a single arrow of time is not stupid.  I was
    briefly an HPUX sysadmin and got yelled at for applying all patches I
    could find because apparently doing this makes your system unstable.
    You have to apply patches only after your system starts crashing.
    ``It's not Solaris'' the more senior SA said.  The
    single-arrow-of-time kool-aid retreats even farther from the HPUX
    best-practice: ``we are going to make something like fake dependencies
    among our patches, to force them into a linear progression, which will
    make them easier for us to regression-test.''  Great idea!
    
    Regression testing is expensive, is an ugly trade-off against
    security, and is extremely unpleasant frustrating repetitive work
    duplicated at each site that makes the whole system feel like rickety
    crap.  Maybe if Sun regression-testing gets really good, people at the
    edge can start being lazier about it.  And the more software
    bundles---not just solaris---that can come under this
    we-regression-test-it umbrella, the more valuable the whole ball of
    wares becomes.  I'm not sure how many years a person needs to watch
    RedHat winning to learn this lesson, but it's still a good lesson.
    
    The problem is, OpenSolaris has gone much too far!  It has taken away
    not just the spaghetti created by playing dim-sum with patches, but
    also the two levels of stability branches: sol8 sol9 sol10, and
    sol10u2 sol10u3, sol10u4.  ``Single arrow of time'' should mean that
    the binary revision control tree follows the source revision control
    tree exactly, so every customer's system is in a state that could be
    built from a 'nightly' script.  It should not mean there are no
    branches in the scm!
    
    Also AIUI, IPS isn't a source packaging system.  It doesn't make it
    easier for me to get from installed-package back to source, then make
    one tiny tweak and rebuild.  or to maintain site-local changes while
    still tracking upstream fixes.  BSD ports/pkgsrc and Gentoo do this,
    and it's fantastic especially for web sites dependent on low-quality
    software written by screaching web2.0 monkeys with lots of
    dependencies where the code in the package system to build them is
    much more elegant than the code-feces being built.  I need the package
    system just to document the monkey-spaghetti.  Leaving this gap---I
    mean, creating a perhaps-deliberate mysteriousness about how the
    source tree is transformed into the binary packages---is a mistake.
    
    
    My sneaking suspicion based on hints at the bottom of the opensolaris
    Download web page, is of a plan is to offer only OpenSolaris for free.
    The old Sol10/SXCE will go back inside the wall, available by
    subscription only.  And SXCE will, just like Sol10, lack the freedom
    to rebuild your system with small local changes.  I don't worry as
    much that fear of CentOS-like ``cannibals'' will put Solaris back
    behind a wall because I don't think that's what'll do it.  (but don't
    you love the way these people who believe in free markets like a
    religion equate `giving customers choices' with `cannibalizing other
    products'?)
    
    Anyway, not that.  I think CYA-ist attitude is what's going to do it.
    There seems to be a concept built into all parts of Solaris OS, and
    internalized by the older customers, that of course They cannot LET
    you do _____ with your system because that will Generate Calls.  There
    are people who believe a system that gives meaningful software freedom
    to end users cannot be supported because people will ``tamper'' with
    the sacred binary releases passed down from the tower.  It's not
    enough to tell customers ``we won't support that'' because some of
    them will pay a lot of money, expect to get something for it, and make
    your life hell by bargaining.  It follows in this binary-software
    advocate's mind that making money on support means trying to create
    many situations in which you can use phrases like, ``oh Gosh i WISH i
    could help you but i just CANT because our lawyers/licensing/property
    simply will not allow, you have to understand so so much code and
    relationships and <hazy confusing things behind wall> hands are tied
    nothing I can do no matter how desperately i want to help you is there
    anything ELSE i can help you with (not this)?''
    
    but this is exactly why I rarely pay for support.  I can't think of
    many times when I've had a problem and actually gotten help.  When I
    do pay, it's usually in truth a payment to download software, and the
    person taking my money enjoys his private fantasy from his
    business-college days that he's selling me a relationship rather than
    a ball of bits.  But if I have to talk to someone on the phone to get a
    URL I'm annoyed, think this MBA-heavy corporation has a shitty
    website, and am angry at them for wasting my time with their effort to
    control my future behavior by manipulating my ``experience''---if I
    want a mind-fuck or a snow-job I'll go to the theater where at least
    the players are pretty, speak well, and have some style and talent.
    
    It's happening right here in front of us---ZFS users are fleeing to
    OpenSolaris because the formal Sun support is slow to deliver patches,
    unhelpful (restore pool from backup), suspect (blame-the-sysadmin,
    weirdly absurd stories delivered with icy certainty).  My favorite
    advice was ``get support that allows escalation to engineering.
    patch-only support is worthless.''  In other words, there's nothing on
    the menu they've written that will fill your belly.  If you can't
    afford to order off the menu, don't bother.  Just get a hot dog from
    the cart in front of their building.
    
    Before, my main disincentive to buying support was that the time I
    spend on the phone talking to front-line PR-techs telling me to jump
    through hoops would be better-spent googling forums, or digging in
    myself offline and trying to solve the problem.  Now, added to that
    incentive, I have to give up source to get access to the stable branch
    with the best support.
    
    No, fuck that.  better to wait for a slightly-more-free alternative to
    Nexenta, or to stick with RedHat/CentOS and use single-branch ``arrow
    of time'' OpenSolaris for the unstable stuff---exactly the opposite of
    how I imagine Sun hoped to grab some territory from Linux.
    

    Attachment: pgpiYk6QHxgZd.pgp
    Description: PGP signature