Review - Glacier - Highly durable, decentralized storate despite massive correlated failures

This paper introduces Glacier, the first distributed system that
provides hard durability guarantees. Glacier is based on DHTs and
employs erasure codes and massive replication to store fragments of an
object on different nodes such that these objects can be retrieved at
any time by the owner.

The strength of this paper lies in the design and implementation of
Glacier, backed by a fairly good evaluation. The authors clearly present
their assumptions (usage, failures, clock) and describe some neat
mechanisms of Glacier that replicates and fragments objects as well as
recover the objects. Glacier shows very low maintenance traffic during
normal mode, shows that traffic increases linearly with increasing load
and the system scales well. The authors also present some attacks that
could be carried out on the system and how their design mitigate those
risks.

The system however lacks delete and update primitives which I believe
are important. It is unclear from the paper as to why these
functionalities were not implemented. Instead the system relies on
timeout of fragments. The authors do not discuss how heterogeneity
affects their system. Even in a corporate LAN, it is fairly common to
have a heterogeneous system, old systems and laptops have smaller
storage and processing ability. This results in some nodes storing far
more fragments (possibly of the same object).

Although this system is robust, scalable and efficient (assuming 11X
disk space overhead is not a problem) in my opinion and represents a
good use of DHTs in a practical environment, I don't believe it is
essential. Since disks are so cheap and storage is unlimited today, we
could imagine a much simpler implementation of a redundant system using
a RAID file server. It solves the same problem and is much simpler, and
effective since we are considering a corporate LAN.
Received on Thu Nov 24 2005 - 09:55:44 EST