On Sun, Jun 21, 2009 at 11:57 AM, Matt W. Benjamin<[email protected]> wrote:
> So the AFS protocol should provide access to the best available snapshot
> available relative to a point in time. Does that mandate that snapshot
> generations are best regarded in the protocol as a timestamp per se? I don't
> believe so. Doing so appears to conflate a temporal search criteria with an
> ordinal (or opaque identifier). It misbehaves near the maximum resolution,
> and seems to carry the assumption that all implementations will be producing
> effectively continuous snapshots.
Agreed. I definitely like the idea of carrying a finely granular
timestamp as metadata, but I specifically see the following pitfalls
to using said timestamp as the primary keying material:
* sparse namespace is undesirable since it makes enumeration of all
version numbers non-trivial
* time nonlinearity (e.g. across reboot, ntpdate, sysadmin blunder, etc.)
* following on from nonlinearity, how do we define this key in a
multi-master replication scheme? MAX({timestamps}) seems logical, but
it is tenuously dependent on system time being sane everywhere in the
cluster.
-Tom
_______________________________________________
AFS3-standardization mailing list
[email protected]
http://michigan-openafs-lists.central.org/mailman/listinfo/afs3-standardization
