On Thu, May 14, 2009 at 1:07 PM, Joey Echeverria <[email protected]> wrote:

> With a 5 server zk ensemble and an 80% write ratio, you should be able
> to support about 10,000 operations per second[1]. That sounds
> reasonable to me for most uses that require locks. If you require
> higher performance than that, then locking probably isn't for you.
> Taking advantage of versioning or other optimistic concurrency control
> is probably necessary.
>


Sounds like it won't scale (smile).

IIRC, currently, the THBase implementation has client run transaction.  It
uses OCC.   I suppose ZK would help guard against case where client dies
mid-transaction.


> At a minimum, I think it makes sense to have zk-based locks as an
> alternative to the current locks which tie up an RPC thread. Testing
> will probably required to see how it performs under various
> assumptions.
>


We're tying up a RPC listener at moment because client just waits on lock
instead of first checking and if taken, retrying later.  Shouldn't be too
hard to fix, hopefully in a way that doesn't impinge on the default
row-based-lock-for-the-single-access case.

St.Ack

Reply via email to