Re: [TLS] Collisions (Re: Nico's suggestions - Re: Consensus Call:

[Nicolas Williams <[e-mail]>]

On Tue, May 11, 2010 at 05:11:23PM +0200, Stefan Santesson wrote:
> On 10-05-11 5:57 PM, "Nicolas Williams" <[e-mail]> wrote:
> > b) the failure will eventually happen again (since the user will most
> > likely want to talk to the same servers whose objects' checksums
> > collided.
> > 
> 
> Yes, and with the cash for that server flushed, the client will be bound to
> retry without caching (since it has no data cached anymore).

It's not a forgone conclusion that the client will retry.  The retry
will have to be at the application layer, and the application might not
be prepared to retry (since it won't know that the fatal error isn't
quite so fatal).

> The client will then (typically) update it's cache on the next successful
> handshake and problem should be fixed.

If one collision arises once, it will arise again, and again, and again
until all or all-but-one of the colliding objects are retired from use.
The client will have to detect collisions and then mark colliding
objects as must-not-use-from-cache.

Detecting collisions is hard.  If the application does not retry then
you can only heuristically decide that fatal errors imply collisions,
which will often not be the case.  If the application does retry then
the collision can be detected.  The server can't detect collisions any
more reliably nor sooner than the client (the server may have less
context than the client), I think.

Nico
--