On Wed, Feb 11, 2009 at 02:46, Dustin <[email protected]> wrote:

>
>  By overlapping, I meant building things from cache based on other
> things found from cache.  Where today's flush destroys everything, the
> negative flush would destroy *some* of everything.
>
>  It seems like a cache generation might be more easily understood.
> I'm still having trouble weighing the value of a negative time flush
> against the potential confusion it may cause.
>
> -~----------~----~----~----~------~----~------~--~---
>
Ok, I see your point.
However, I must disagree it is linked to the negative flush.
You can already face the problem with the current flush behaviour, with this
scenario for instance:
at t, process a reads data from memcached and continues its process
at t+1, process b flush in immediate mode the whole cache
at t+2, process a, continuing the process started at t, write some new data
in cache based on what it read earlier
basically, the data put by process a is still in the cache because set after
the flush while it is based on data that was supposed to be fully erased by
the flush, so we are in the situation you described.
The window of possible discrepency gets bigger if we use delay to perform
sequentially the flush on the whole cache.

Reply via email to