hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <tdunn...@maprtech.com>
Subject Re: Starting the Hadoop DataNode inside the HBase process?
Date Tue, 19 Jul 2011 00:13:49 GMT
On Mon, Jul 18, 2011 at 9:32 AM, Jason Rutherglen <
jason.rutherglen@gmail.com> wrote:

> > My gut is that this would be a maintenance headache
>
> What specifically do you think would cause a problem?
>

Tracking versions for one.  Everybody has a different favorite.  That is the
nice thing about standards.  There are so many to choose from.

Besides, how do you handle people who want the snapshots and higher
performance that you get from maprfs?

> Internal management of ZK is already an option (and I don't recommend that
> > either, for different reasons)
>
> What are the reasons?
>

The basic issue is that it is nice to use ZK to determine which services are
up and to avoid race conditions as services come up.  If some of the
services are actually running ZK, how do you distinguish that process
getting hung from not being up?

Also, ZK is very reliable and that is the primary virtue we are trying to
capitalize on when we use it as a coordination service.  Given that, how is
it a good thing to incorporate it into software that is inevitably less
stable?  Isn't that tantamount to  giving ZK's primary virtue?

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message