Mailing-List: contact common-user-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Reply-To: common-user@hadoop.apache.org
Received-SPF: neutral (nike.apache.org: local policy)
Message-ID: <4BC364CA.3050309@apache.org>
Date: Mon, 12 Apr 2010 11:22:02 -0700
From: Patrick Hunt <phunt@apache.org>
User-Agent: Mozilla/5.0 (X11; U; Linux i686 (x86_64); en-US;
 rv:1.9.1.9) Gecko/20100317 Thunderbird/3.0.4
MIME-Version: 1.0
To: zookeeper-user@hadoop.apache.org
CC: Thomas Koch <thomas@koch.ro>, Mahadev Konar <mahadev@yahoo-inc.com>,
        common-user@hadoop.apache.org, gearman@googlegroups.com
Subject: Re: feed queue fetcher with hadoop/zookeeper/gearman?
References: <C7E89C75.33584%mahadev@yahoo-inc.com>
 <201004121949.01427.thomas@koch.ro>
In-Reply-To: <201004121949.01427.thomas@koch.ro>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

See this environment http://bit.ly/4ekN8G. Subsequently I used the 3 
server setup, each configured with 8gig of heap in the jvm and 4 
CPUs/jvm (I think I used 10second session timeouts for this) for some 
additional testing that I've not written up yet. I was able to run ~500 
clients (same test script) in parallel. So that means about 5million 
znodes and 25million watches.

The thing to watch out for is:
1) most important is you need to tune the GC, in particular you need to 
turn on CMS and incremental GC. OTW the GC pauses will cause high 
latencies and you will see session timeouts
2) you need a stable network, esp for the serving ensemble
3) sufficient memory available in the JVM heap
4) no IO issues on the serving hosts (VM's, overloaded disk, swapping, 
etc...)

In your case you've got less going on with only 30 or so writes per 
second. The performance page shows that your going to be well below the 
max ops/sec we see in our testing harness.

btw, gearman would also be a good choice imo. I've looked at integrating 
ZK with gearman, there are two potentials. 1) as an additional backend 
persistent store for gearman, 2) as a way of addressing gearman 
failover. 1 is pretty simple to do today, 2 is harder, would require 
some changes to gearman itself but I think it would be useful (automatic 
failover of persistent tasks if a gearman server fails).

Patrick

On 04/12/2010 10:49 AM, Thomas Koch wrote:
> Mahadev Konar:
>> Hi Thomas,
>>    There are a couple of projects inside Yahoo! that use ZooKeeper as an
>> event manager for feed processing.
>>
>> I am little bit unclear on your example below. As I understand it-
>>
>> 1. There are 1 million feeds that will be stored in Hbase.
>> 2. A map reduce job will be run on these feeds to find out which feeds need
>> to be fetched.
>> 3. This will create queues in ZooKeeper to fetch the feeds
>> 4.  Workers will pull items from this queue and process feeds
>>
>> Did I understand it correctly? Also, if above is the case, how many queue
>> items would you anticipate be accumulated every hour?
> Yes. That's exactly what I'm thinking about. Currently one node processes like
> 20000 Feeds an hour and we have 5 feed-fetch-nodes. This would mean ~100000
> queue items/hour. Each queue item should carry some meta informations, most
> important the feed items, that are already known to the system so that only
> new items get processed.
>
> Thomas Koch, http://www.koch.ro