[Gluster-devel] HA translator total failure in 2.0.0rc1

[Daniel Maher]

Hello all,

As noted in this email to the gluster-users list :
http://zresearch.com/pipermail/gluster-users/20090114/001389.html

I've got a simple and reproducible scenario to crash a Gluster client 
using the HA translator to access two AFR'd servers.  The scenario is 
identical to that described by Krishna Srinivas on the gluster-devel 
list on 08-01-2008 :
http://lists.gnu.org/archive/html/gluster-devel/2009-01/msg00059.html

        Client
          |
          HA
         /  \
        /    \
     AFR1    AFR2
      |        |
  Server1    Server2

Basically, if i stop glusterfsd on Server1, HA on Client switches to 
AFR2 as expected ; however, when i re-enable glusterfsd on Server1, then 
stop glusterfsd on Server2, one of two things occurs :
1. Client stops communicating entirely with the cluster (transport 
endpoint not connected), or
2. Client recovers and continues communicating with AFR1.
It appears to be random as to which one actually occurs.

If the client recovers and continues to communicate, and i re-enable 
glusterfsd on Server2, Client stops communicating immediately with the 
cluster - every time, guarunteed.

There are therefore two key questions :
1. In the first component, why doesn't the client switch gracefully 
between available subvolumes ?
2. In the second component, why does re-enabling a 
previously-unavailable subvolume crash the client ?

All relevant details are in the mail to the gluster-users list, linked 
above.

Any ideas what's going on here ?


-- 
Daniel Maher <dma+gluster AT witbe DOT net>