Ühel kenal päeval (kolmapäev, 19. jaanuar 2005, 00:39-0500), kirjutas
Tom Lane:
> What this really boils down to is whether we think we have
> order-of-a-year before the patent is issued. I'm nervous about
> assuming that. I'd like to have a plan that will produce a tested,
> credible patch in less than six months.
Can't this thing be abstracted out like so many other things are (types,
functions, pl-s) or should be/were once (storage managers) ?
Like different scheduling algorithms in the linux kernel ?
What makes this inherently so difficult to do ?
Is it just testing or something for fundamental?
Most likely also the gathering of information needed to decide on
replacement policy.
If just testing, we could move fast to supplying two algos LRU/ARC ,
selectable at startup.
This has extra benefit of allowing easily testing other algorithms - I
guess that for unpredictable workloads a random policy in 80% tail of
LRU cache should not do too badly, probably better than 7.x's seqscan
polluteable LRU ;)
--
Hannu Krosing <hannu@tm.ee>