Re: icp bitrot?

On Thu, 2002-10-10 at 08:28, Duane Wessels wrote:
> > However, one thing I've been thinking about is allowing a 'memory only'
> > flag to the new getPublic* calls, to indicate we are in a latency
> > critical section, and only in-memory key tables should be queried. I'm
> > not convinced that we'll need that though.
>
> I would simply suggest, that if at all possible, users should have
> the option to keep some kind of ICP-searchable index of the whole
> cache in memory.

Absolutely. All they need to do is use cache_dirs that have an in-memory
index.

The question is: given the follow hypothetical configuration:
10Gb of ufs style cache-dir (in memory index).
30Gb of reiserfs/coss cache dir (on disk index).

Should ICP query both cachedirs (most accurate results, greatest
latency) or only the in-memory index providing cache-dir (lowest
latency, only provides hits on the 10Gb store).

This should be runtime configurable, and (potentially) self tuning IMO.
e.g. if the average icp service time rises above .5 seconds, stop
querying disk indexes for 1 hour.

Rob