Slow (~0.3s) get request

[Yossi Ittach <yo...@gmail.com>]

Hi all

I Have Hbase with 1 RegionServer , Hadoop with 1 DataNode (on diffrent
machines).
The Hbase contains ~ 3M rows. It takes approxametly 0.3 sec to reply to a
get request (from the moment it's submitted to the moment a reply is
recieved) - this is way too long. Is it reasonable , or is it just a
function of "add more region servers"?

If I'll decrease the hregion file size , will it help?

Thanks!


Vale et me ama
Yossi

Re: Slow (~0.3s) get request

[Yossi Ittach <yo...@gmail.com>]

Thank you very much!

Vale et me ama
Yossi


On Tue, Oct 28, 2008 at 1:48 PM, Jean-Daniel Cryans <jd...@apache.org>wrote:

> Yossi,
>
> You should run both regionserver and datanode on the same machine, this
> reduces network overhead.
>
> A typical random read on our cluster takes 0.00x ms or sometimes 0.0x ms
> under normal circumstances. What changes that is:
> the size of the rows
> the number of deletes that row had (though this is much better starting in
> 0.18.1 released yesterday)
> the number of version that row has (>100 000, it may become a problem)
> if it's the first time that that region is seen (requires finding that
> region first)
>
> Also maybe your region is just badly needing a major compaction which
> happens once everyday.
>
> Decreasing the file size shouldn't really affect the speed of that request.
>
> Adding more region servers will help only if your current region server is
> overloaded. However, if your table only has 1 region, adding a new server
> won't change much.
>
> J-D
>
> On Tue, Oct 28, 2008 at 4:47 AM, Yossi Ittach <yo...@gmail.com> wrote:
>
> > Hi all
> >
> > I Have Hbase with 1 RegionServer , Hadoop with 1 DataNode (on diffrent
> > machines).
> > The Hbase contains ~ 3M rows. It takes approxametly 0.3 sec to reply to a
> > get request (from the moment it's submitted to the moment a reply is
> > recieved) - this is way too long. Is it reasonable , or is it just a
> > function of "add more region servers"?
> >
> > If I'll decrease the hregion file size , will it help?
> >
> > Thanks!
> >
> >
> > Vale et me ama
> > Yossi
> >
>

Re: Slow (~0.3s) get request

[Jean-Daniel Cryans <jd...@apache.org>]

Yossi,

You should run both regionserver and datanode on the same machine, this
reduces network overhead.

A typical random read on our cluster takes 0.00x ms or sometimes 0.0x ms
under normal circumstances. What changes that is:
the size of the rows
the number of deletes that row had (though this is much better starting in
0.18.1 released yesterday)
the number of version that row has (>100 000, it may become a problem)
if it's the first time that that region is seen (requires finding that
region first)

Also maybe your region is just badly needing a major compaction which
happens once everyday.

Decreasing the file size shouldn't really affect the speed of that request.

Adding more region servers will help only if your current region server is
overloaded. However, if your table only has 1 region, adding a new server
won't change much.

J-D

On Tue, Oct 28, 2008 at 4:47 AM, Yossi Ittach <yo...@gmail.com> wrote:

> Hi all
>
> I Have Hbase with 1 RegionServer , Hadoop with 1 DataNode (on diffrent
> machines).
> The Hbase contains ~ 3M rows. It takes approxametly 0.3 sec to reply to a
> get request (from the moment it's submitted to the moment a reply is
> recieved) - this is way too long. Is it reasonable , or is it just a
> function of "add more region servers"?
>
> If I'll decrease the hregion file size , will it help?
>
> Thanks!
>
>
> Vale et me ama
> Yossi
>