Fault Tolerance?

[Dustin Sallings]

On Sep 27, 2007, at 21:56, Paul Scott wrote:

>> Agreed. Wouldn=92t it be great though to have a mem-based HA =
datastore?
>
> I would certainly vote +1 on that idea!

	You do realize you wouldn't have anything remotely like the =20
performance of memcached, don't you?  You'd need something along the =20
lines of two-phase-commit if you want any kind of correctness.  If =20
you don't want correctness, then why are you worried about HA?

	If you lose a node, how do you plan on rematerializing?  A =
complete =20
synchronization would block both nodes in a two-node cluster.

	How would you handle conflicts during rematerialization after a =20=

netsplit?

	Is it acceptable to block all clients during a netsplit (pending =
=20
some sort of magical synchronization that knows what to do when =20
conflicts occur)?

	After you get all of the pieces in place, are you sure you'd =
have =20
something that would be any faster than any solution that isn't =20
completely in-memory?

--=20
Dustin Sallings