[HCoop-Discuss] hardware advice

[Adam Megacz]

Adam Chlipala <adamc at hcoop.net> writes:
> What about taking frequent snapshots?  Is that too expensive for us to 
> be able to do it every hour or so to a machine in the same data center?  

I was doing that for a while with krunk.  Unfortunately AFS's snapshot
mechanism is only incremental at file granularity.  So if you change
one byte in a multi-megabyte/gigabyte file, the snapshot mechanism
will transfer the whole file across the network.  I don't think it
would scale well.  Also, it seemed to be causing weird problems for a
while.  In fact, I think you (adamc) were one of the users who got
bitten by it (I only tried it on the volumes belonging to admins).

I'll be honest here, high availability is an area where AFS is
severely lacking.  But it's not like NFS or Samba are any better.  The
only filesystems that do HA are either research prototypes (Coda,
Ceph), proprietary/commercial (Netapp), or not suitable for
general-purpose user-home-directory storage (GoogleFS/Hadoop, Lustre).

  - a