Come to think of it, I believe that that is false.
It is only true if the added drive is one of the failing population and that is only
a few percent of the total population of drives. So if you add a drive that isn't
in the failing group the MTBF is still the same.
That doesn't really make any sense. MTBF doesn't seek to describe
individual drives and the percentile of failing drives is going to
change over time - so you can't approach it that way. Furthermore
adding drives adds complexity to the disk subsystem and increased
chances of infant mortality, premature death, etc. as well as a single
failing drive (whether of mature years or not) messing up the data on
other healthy drives. Maybe MTBF is not an ideal way of describing
this, but if it is totally wrong a whole lot of people need to be
reeducated as the prevailing way of describing array reliability is
array MTBF= drive MTBF/n drives in array.
Sure if you add a drive from a "failing group" it will die prematurely
& if you have all good drives they will last past their estimated
service life and probably go around the same time. But the whole
point of descriptive tools like MTBF is to account for a
_whole_population_ (whether it does this well or not).
But doesn't describe "your" drive. It describes the mean time between failure
for the whole population of drives (but only the ones expected to fail).
We're not talking about "my drive." We are talking about the chances
of suffering problems with an array of drives as compared to a single
drive. When you put drives together in an array a more complex animal
is created.
Unless you
own the whole population or a significant part of the population you can't say any-
thing about whether you will be hit and if so *when* you will be hit with a failure.
MTBF describes the mean time between one drive failure and the next, averaged,
if you know how to break it down.
MTBF is an extrapolated largely theoretical number so it never
accurately correlates to real world use of the whole population.
Disclosed MTBF is generally _theoretical_MTBF so it provides a
guesstimate based on a theoretical model of reality. It is something
to be taken with a grain of salt anyway. _Operational_MTBF is
generally not disclosed by manufacturers as it tends to be different
than the number they used to sell the product. MTBF is an
extrapolation based on the aggregate analysis of large numbers of
drives over (generally) a short period of time & says nothing about a
particular disk.
Rather crude:
With a single drive that is going to die at moment x there are two possibilities
when adding another one to form an array: there is a 50% chance that that
drive dies sooner and a 50% chance that it dies later than the first one.
Yet the chance that the array will have died at moment x is still 100%.
Rather crude indeed. So the array is always dead? You messed that up
pretty bad.
Array reliability doesn't correlate to flipping a two sided coin on
the first day of "statistics" as covered in grade school math. You're
missing that the point is to characterize the shift in extrapolated
up-time/reliability NOT whether an array or individual disk is up at a
particular time. Characterizing drive health in a binary way: either
it is up or it is dead is just restating the obvious and explains
nothing useful.
But you can't say that the second drive makes the array less reliable because
you might have used that drive as a single drive as well and not have complained.
Sure I would have complained, it's just a question of when. Nothing
lasts forever. And just because it may not suffer an infant death
doesn't mean it has a 100% chance of running 100% reliably for the
full estimated service life.
More importantly, when you use them together the array goes down as
soon as the first one to fail does so. Therefore up-time RAID0 =
actual life of shortest living drive of the array- but you can't
accurately predict actual lifespan so we are forced to deal with more
complex and somewhat problematic statistical
explanations/extrapolations.
I don't rate the chance of being 'lucky' as a rate of reliability.
Rather convoluted.
You are indeed counting on being lucky, that is to say you expect to
not have a mixture of good and sub par/subtly damaged drives in an
array and that either drives work 100% correctly or they are offline.
The real world doesn't operate that way.
Your problem with the prevailing MTBF calculation of an array is "But
that still says nothing about when or if your drives will die" and
your explanation centers around a binary view of health at "moment x"
(without limitation). These are incompatible insofar as one does not
explain the other. A point in time is different than a time
frame/length of time.
Because when the single drive dies, it too is dead. Dead is dead, period.
Not true. It's just not the case that either a drive works 100%
accurately or it is completely dead. That happens sometimes but what
is far more likely is a period where a drive is failing and errors are
creeping in before it has to be removed. This is unacceptable in any
environment and difficult to combat without intelligent use of
redundant data/drives. When no redundant data is available individual
drives are preferred because it is easier to contain/limit such
problems. In raid0, because the drives depend on each other, the
reliability of the good drives in the array is limited by the first
one to fail/begin to fail. Because that is hard to predict for many
reasons you can neither say something like array service life is
halved with a 2 drive raid 0 nor say expected reliability is identical
because I always have a 50% chance that any particular drive I choose
will work.
Your idea that adding drives to an array does not affect reliability &
up-time of a disk subsystem is completely unique - and it is that way
for a reason. To assert that a more complex system has the same
likelihood for something to go wrong as a much simpler one makes no
intuitive or practical sense. It's curious that while you despise
newbs and are eager to slam them many of your arguments such as this
one tend to pander to them, as no one who knows any better would
believe you. I've been giving you the benefit of the doubt, so far,
but come on- enough trolling, Folkert.
