"Greg Sabino Mullane" <greg@turnstep.com> writes:
> Spending time on this is silly, IMO, unless there is a technical reason
> why the feature should be replaced.
Well, people can validly have different opinions on how critical it is
to dodge the upcoming patent (and surely whether you live in the US or
not affects your viewpoint). But as to the second part of your comment,
the fact is that the ARC buffer management code has been underwhelming
and we were already looking around for something better. I believe Jan
already admitted that his original testing was flawed and that the
algorithm is not so much better than LRU as he thought. We are also
staring at the fact that ARC is not at all helpful when it comes to the
problem of reducing contention for the BufMgrLock. It uses an inherently
centralized, serialized data structure and the operations it requires
aren't notably cheap. So I was already feeling dissatisfied even before
the patent issue came up, and I'm all for getting rid of ARC as soon as
we can find (and test) something better.
regards, tom lane
