hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrew Purtell <apurt...@apache.org>
Subject Re: ZK rethink?
Date Sun, 12 Apr 2009 02:24:58 GMT

Hi Chad,

> -- Won't we soon be creating a lot less objects? This
> may reduce the pressure on the GC substantially. 

Yes, and yes. But, in write heavy environments I think the
new edits represent new memcache entries that are then
freed after flush, etc., and still put pressure on the GC.

> -- Have a Java ZK session handler with a longer timeout for
> ease-of-use and accessibility to novices. Have a native C ZK
> session handler with a shorter timeout for folks who get
> serious about things.
> -- We definitely should avoid swapping - but doing an
> mlockall() seems like it might be going too far. Is swapping
> actually part of the problems that Andrew has seen or are
> his issues strictly related to the GC operating within
> physical memory bounds?

Swapping is not actually part of my issues but it easily
could cause a number of problems for unfortunate users. 
Mlocking datanodes might be something worth exploring along
with regionservers if someone has the time and inclination.
These components are most likely to cause trouble if they
are partially swapped out during a GC cycle. (Namenode also,
as a separate case, just for sake of completeness as it is
always given special attention during deployment -- or
should, at least.)

   - Andy


View raw message