[Gluster-devel] Client side AFR race conditions?

[Anand Babu Periasamy]

Hi Martin,

If application doesn't use locking in a multi-user mode, data
can be corrupted with or without AFR. With AFR in place, corruption
can also result in disparate set of data, other than losing the order
of writes. No file system can guarantee integrity, if applications
do not synchronize writes in multiuser mode.

Even if we introduce atomic writes within AFR, it still doesn't
fix application's bugs. It will only slow down writes for well
behaved applications.

--
Anand Babu Periasamy
GPG Key ID: 0x62E15A31
Blog [http://ab.freeshell.org]
The GNU Operating System [http://www.gnu.org]
Z RESEARCH Inc [http://www.zresearch.com]



Martin Fick wrote:
> --- Anand Babu Periasamy <ab at gnu.org.in> wrote:
> 
>> When multiple applications write to a same file, it
>> is really application's responsibility to handle 
>> coherency using POSIX locks or some form of IPC/RPC 
>> mechanism.  Even without AFR, file system's do not 
>> guarantee order of writes and hence integrity of 
>> data. When AFR is inserted, this corruption may
>> lead to disparate set of data other than overwrites.
>>
>> It shouldn't be seen as an issue with AFR. If
>> applications handle coherency, AFR will work fine.
> It
>> is possible to introduce atomic-write option (locked
> 
>> writes) in AFR, but it is useless, because it still 
>> cannot avoid corruption because one application 
>> overwrote the data of the other, without holding a 
>> lock.
>>
>> In summary, AFR doesn't have race condition.
> 
> I'm sorry, but I disagree.  I do not think that you
> understand the full problem.  The issue is not one of
> application coherency but simply one of ensuring that
> both copies have the same data.  With the current
> implementation I cannot guarantee that I get my good
> or my bad data consistently, on every read I could get
> a different answer, as many different answers as there
> are subvolumes.  It really is a split brain issue.  I
> am not asking for ordered writes, simply that the
> writes be the same on all subvolumes.
> 
> In fact, I really should extend the question and ask,
> how does AFR handle locking?  Does it ensure that all
> subvolumes are locked?  I assume it does, but the next
> question is, what does it do about potential infinite
> locks if clients die?  Does is detect a downed client
> and release the lock?
> 
> -Martin
> 
> 
> 
>       ____________________________________________________________________________________
> Be a better friend, newshound, and 
> know-it-all with Yahoo! Mobile.  Try it now.  http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ