Re: MOO Meeting 970305

Duane Wessels wrote:
> Issues I have with ICP invalidation:
>
> 1) UDP is unreliable. Do you require ACKs of the invalidation?

I don't think ACKs needed. I think of this more as a notice than a
criteria.

> 2) Would you establish a peer relationship with someone who had
> the power to invalidate your entire cache?

I don't think it is that bad. Objects should only be marked stale (to
require a revalidation with the source), not expunged completely.

> My cache access lists currently accept ICP from hundreds of
> subnets around the world. How can I trust the invalidation is
> legitimate?

You can't. And you can't trust all to not flood you with requests
either.

> 3) How do you limit the scope? Does a single PUT at one
> cache cause a flood of invalidation messages all over
> the world?

Not that practical. I think that the main issue is to minimize the
possibility that the same person gets a old object back after a PUT, not
to forse a update all over the world.

> Practially speaking, we could ad an opcode and it would be
> ignored by caches running prior versions.

Good.

I don't know much about the details in ICP protocol. How does it handle
extensions? Is there any standard way to make a (non-standard)
extension? That is almost guaranteed to not be misinterpreated by other
servers, possibly using a later version of ICP or different extensions?
Received on Tue Jul 29 2003 - 13:15:40 MDT