Mailing-List: contact user-help@hbase.apache.org; run by ezmlm
Precedence: bulk
Reply-To: user@hbase.apache.org
MIME-Version: 1.0
In-Reply-To: 
 <CAAg3a2pKfWwz-p5-tO5BYEXOPGU401KaE1amMJKpKAyzFz1DVQ@mail.gmail.com>
References: 
 <CAK7dMtBn_sxJ9S93b-Ab53vUtQGaQ-mskXKNcDf4kxcDwWqY0A@mail.gmail.com>
	<CAAg3a2pKfWwz-p5-tO5BYEXOPGU401KaE1amMJKpKAyzFz1DVQ@mail.gmail.com>
Date: Mon, 25 Apr 2016 01:38:07 -0700
Message-ID: 
 <CAK7dMtD5hkV2CGJFTMWPXp+CANu7jwUyd+x5Y9HgweqT9WYMDA@mail.gmail.com>
Subject: Re: Balancing reads and writes
From: Kevin Bowling <kevin.bowling@kev009.com>
To: user@hbase.apache.org
Content-Type: multipart/alternative; boundary=047d7b5d5ffc83863805314b1811

--047d7b5d5ffc83863805314b1811
Content-Type: text/plain; charset=UTF-8

Yeah here's the config I ended up with.  Out of the box it had really
severe blocking for write bursts, reads are much better with this and
handlers turned up a bit:

  <!-- Read perf:  split reads from writes in the call queue -->
  <property> <name>hbase.ipc.server.callqueue.read.ratio</name>
<value>0.4</value> </property>
  <!-- We could probably keep reads and scans on same queues but it
triggers startup bug (HBASE-14252) -->
  <property> <name>hbase.ipc.server.callqueue.scan.ratio</name>
<value>0.5</value> </property>
  <!-- Increase the number of queues to threads for less contention
(default 0.1) -->
  <property> <name>hbase.ipc.server.callqueue.handler.factor</name>
<value>0.5</value> </property>

Regards,
Kevin

On Sat, Apr 16, 2016 at 9:27 PM, Vladimir Rodionov <vladrodionov@gmail.com>
wrote:

> There are separate RPC queues for read and writes in 1.0+ (not sure about
> 0.98). You need to set sizes of these queues accordingly.
>
> -Vlad
>
> On Sat, Apr 16, 2016 at 4:23 PM, Kevin Bowling <kevin.bowling@kev009.com>
> wrote:
>
> > Hi,
> >
> > Using OpenTSDB 2.2 with its "appends" feature, I see significant impact
> on
> > read performance when writes are happening.  If a process injects a few
> > hundred thousand points in batch, the call queues on on the region
> servers
> > blow up and until they drain a new read request is basically blocked at
> the
> > end of the line.
> >
> > Any recommendations for keeping reads balanced vs writes?
> >
> > Regards,
> > Kevin
> >
>

--047d7b5d5ffc83863805314b1811--