[tahoe-dev] erasure coding makes files more fragile, not less
warner at lothar.com
Wed Mar 28 20:54:19 UTC 2012
On 3/27/12 12:06 PM, Zooko Wilcox-O'Hearn wrote:
> In fact, I think approximately 90% of all files that have ever been
> stored on a Tahoe-LAFS grid have died. (That's excluding all of the
> files of all of the customers of allmydata.com, which went out of
That.. is a pretty broad and potentially disingenuous statement, and
feels unsupported. I think I know what you mean, but it's a bit like
saying "everybody dies" or "all programs crash eventually": maybe true,
but kinda useless, and kinda deceptive, or at least distracting.
[BTW folks: Zooko and I talk about this stuff all the time, and we know
each other's opinions pretty well, so please don't misinterpret my words
as indicating anger or annoyance. We're old pals, and this is a
well-worn comfortable argument.]
The metric that I'd find useful is what percentage of files *that people
actively tried to keep around* were lost. Tahoe is a system for
multiplying the durability of your servers, but it's not magic, and if
you start with lousy unmaintained servers, then you aren't likely to
have good results.
> I came up with this provocative slogan (I know Brian loves my
> provocative slogans): "erasure coding makes files more fragile, not
Ah, you do know how to provoke me :). That's like saying "seat belts,
airbags, and helmets kill people", invoking haunting images of demonic
safety gear stalking the last remaining humans through the forest,
seatbelts to snare, airbags to suffocate, and helmets to keep watch for
the desperate counterattack. Sometimes (I'm reminded of your alien
toaster example) you imply things like "we should outlaw seatbelts, and
mandate that car seats must be attached to the front bumper, so people
feel scared enough to drive slower", which, although it might reduce the
rate of car crashes, is not a workable solution.
And what I think you should mean instead is "seat belts certainly save
lives, but we should pay attention to whether people might be tempted to
drive faster because of the feeling-of-safety they provide, and think
about how to mitigate that".
> The idea behind that is that erasure coding lulls people into a false
> sense of security.
That's the important part (if it's even true), and provocative
soundbites which omit it are delivering the wrong message.
> If K=N=1, or even if K=1 and N=2 (which is the same fault tolerance as
> RAID-1), then people understand that they need to constantly monitor
> and repair problems as they arise.
You know, I'm not sure that's actually true. I have a feeling that any
folks who lose data in a Tahoe grid (and I'm not accepting that claim
yet: I haven't personally heard many examples of loss, although you talk
to more users than I do) would be just as likely to lose data in a RAID
array, or in a single-disk server. I.e. to study this properly, I'd want
to separate the user population into "careful sysadmins" and "casual
end-users", and examine failure rates (and experiences) in the two camps
> But if K=3 and N=10, then the beautiful combinatorial math tells you
> that your file has lots of "9's" of reliability. The beautiful
> combinatorial math lies!
Hrm, we've had this argument before and I'm never sure where to go it.
Yes, the math in our provisioning/reliability tool describes a somewhat
unrealistic model with the usual because-it-makes-the-math-easier
assumptions (Poisson processes, independent identically-distributed
failures). Should we get rid of it? No, I think it still has value.
Should we add some warning stickers that say "human error and
non-independent failure modes will probably limit how close you can get
to these numbers"? Sure. If people ignore those stickers and believe the
fairy-tale math and drive too fast and crash and burn, should we throw
out the math? No, I think the tools are still useful to people who
understand the limits of the model.
> If almost all of the files that have ever been stored on Tahoe-LAFS
> have died, this implies one of two things:
(ugh, just the way you phrase that claim makes it sound like Tahoe is a
fundamentally flawed technology and any file that comes into contact
with it catches the plague and falls deathly ill. How about "any
potential file loss in Tahoe must come about because of one of the
> 1. The "reliability" of the storage servers must have been below K/N.
> I.e. if a file was stored with 3-of-10 encoding, but if each storage
> server had a 75% chance of dying, then the file would be *more* likely
> to die due to the erasure coding, rather than less likely to die,
> because a 75% chance of dying, a.k.a. a 25% chance of staying alive,
> is worse than the 30% number of shares required to recover the file.
Wait wait, the details are somewhat correct but the conclusion is wrong
and the premise is off-base. Yes, k>1 on servers with <50% reliability
is worse than k=1 on those same servers: bad servers are bad, relying
upon more of them is worse. 3-of-10 on bad servers is worse than
1-of-10. But 1-of-10 is way better than 1-of-2 or 1-of-1. And 3-of-10 is
better than 3-of-3. 3-of-10 on good=25% servers has an 80.7% chance of
failure, nearly the same as 1-of-1's 75% chance of failure. 2-of-10 on
good=25% gets you a 24.4% chance of failure, way better than 75%, and
1-of-10 is down to 5.6% failure. So it's not
erasure-coding/replication that's causing the problem, it's the
combination of k>1 and horrifically bad servers.
I'd rewrite your conclusion to be that the reliability must have been
below 50%, not below k/N. I think we've always assumed that servers will
have better than 50% reliability (you'd never pay a hosting provider for
anything worse than that). Tahoe is a tool for making a great grid out
of good servers, not for making a good grid out of lousy servers. If
you're stuck with that, use 1-of-N and hope for the best.
(If our goal had been to support lots-of-lousy-servers, we'd probably
have built something else: implement replication and automatic repair
first, then get around to things like mutable files and a web interface.
The result would be inefficient on good servers.)
> 2. The behavior of storage servers must not have been *independent*.
But let's add some other possibilities, some of which we can improve
with code, some of which depend upon sysadmins doing their jobs and grid
members honoring commitments to their clients:
3: people got bored, wandered away, took their servers with them
4: hosts got rebooted and servers didn't automatically come back up
5: failed servers weren't replaced
6: files weren't repaired
When there's no incentive to keep your server running (which could be as
simple as knowing that other people will know when it's been offline),
servers tend to go away, and using replication or FEC to mitigate that
is expensive (the old problem of not knowing whether the server is
coming back or not, therefore needing to treat it as permanent,
triggering immediate repair, and eventually the repair bandwidth is so
high that you can't use the grid for actual work).
We need better OS-integration code to make it easy to get a server to
come up on each reboot. On OS-X that means a LaunchAgent or something.
On Debian/etc it means an init.d or upstart job. (we have this problem
all the time with buildslaves: it's pretty easy to get one running by
hand, but the energy barrier between that and having a real every-reboot
service is high enough that a lot of folks don't bother).
I've been saying forever that it's too hard/slow/inconvenient to get
periodic repair to run automatically (cron jobs are soo gross, and
suffer from the same energy-barrier problem). And I've been planning
(and failing to complete) to move this functionality into the Tahoe
client for nearly as long (#643, #483). I really think that having
automatic repair of everything reachable from your rootcap(s) is
necessary to get close to the enticing durability promise that
> My conclusion: if you care about the longevity of your files, forget
> about erasure coding and concentrate on monitoring.
My conclusion: we can't make serious claims about the benefits/etc of
erasure coding until we've eliminated the confounders, by fixing known
problems that hurt reliability and then collecting actual data. I
think that claiming "erasure coding makes files more fragile" is wrong.
Getting good reliability out of a Tahoe grid requires just as much
attention and effort as any other storage technology: it can't conjure
reliability out of nothing. But I still believe that, for the same
effort, you'll get much more durability out of Tahoe than out of simple
replication/RAID (when Tahoe has the same level of automation as those
other tools, which it doesn't yet).
And I agree that building monitoring tools, and especially the automatic
repair agent, is just as important as Zooko says. Note that monitoring
by itself isn't enough: you need to take action when it's required,
either on the small scale (repair) or on the large (replacing servers).
But good tools to tell you when action is needed is the first step.
yay provocation! :-)
: 3-of-10 on p=25%, chance of failure is the sum of three cases:
Num(good servers)=0: 25%^10 = 5.6%
Num(good servers)=1: 10*25%^1*75%^9 = 18.8%
Num(good servers)=2: 10*9*25%^2*75%^8 = 56.3%
2-of-10 on p=25%: sum of the first two terms (Num=0,Num=1)
: this was in the plans for the "repair agent" at Allmydata before it
shut down: that would have been a good place to collect long-term
statistics on drive failure, share loss, repair bandwidth, and file
decay curves. The idea was to collect that data for a year or two,
including failures due to motherboards breaking and upgrades going
wrong, and then use it to justify a more deliberate choice of k and
N (minimizing cost while still meeting the reliability goals). This
project lost a lot of momentum when we lost the centralized place
to do that research, and the paycheck to build the supporting
More information about the tahoe-dev