Ok, thanks.  Quick Q:  Won't each sink consume the same data?  Do I need to set up the load balancing sink processor to keep that from happening?


On Jan 16, 2013, at 5:47 PM, Hari Shreedharan <hshreedharan@cloudera.com> wrote:

Also can you try adding more HDFS sinks reading from the same channel. I'd recommend using different file prefixes, or paths for each sink, to avoid collision. Since each sink really has just one thread driving them, adding multiple sinks might help. Also, keep an eye on the memory channel's sizes and see if it is filling up (there will be ChannelExceptions in the logs if it is). 


Hari

-- 
Hari Shreedharan

On Wednesday, January 16, 2013 at 2:34 PM, Brock Noland wrote:

Good to hear! Take five six thread dumps of it and then them our way.

On Wed, Jan 16, 2013 at 2:30 PM, Andrew Otto <otto@wikimedia.org> wrote:
Cool, thanks for the advice! That's a great blog post.

I've changed my ways (for now at least). I've got lots of disks to use once memory starts working, and this node has tooooons of memory (192G).

Here's my new flume.conf:

This is doing better, for sure. Note that I took out the timestamp regex_extractor just in case that was impacting performance. I'm using the regular ol' timestamp interceptor now.

I'm still not doing so great though. I'm getting about 300 Mb per minute in my HDFS files. I should be getting about 300G. That's better than before though. I've got 10% of the data this time, rather than 0.14% :)




On Jan 16, 2013, at 4:36 PM, Brock Noland <brock@cloudera.com> wrote:

Hi,

I would use memory channel for now as opposed to file channel. For
file channel to keep up with that you'd need multiple disks. Also your
checkpoint period is super-low which will cause lots of checkpoints
and slow things down.

However, I think the biggest issue is probably batch size. With that
much data you are likely going to want a large batch size for all
components involved. Something a low multiple of 1000. There is a good
article on this:

To re-cap would:

Use memory channel for now. Once you prove things work you can work on
tuning file channel (going to write larger batch sizes and multiple
disks).

Increase the batch size for your source/sink.

On Wed, Jan 16, 2013 at 1:22 PM, Andrew Otto <otto@wikimedia.org> wrote:
Ok, I'm trying my new UDPSource with Wikimedia's webrequest log stream. This is available to me via UDP Multicast. Everything seems to be working great, except that I seem to be missing a lot of data.

Our webrequest log stream consists of about 100000 events per second, which amounts to around 50 Mb per second.

I understand that this is probably too much for a single node to handle, but I should be able to either see most of the data written to HDFS, or at least see errors about channels being filled to capacity. HDFS files are set to roll every 60 seconds. Each of my files is only about 4.2MB, which is only 72 Kb per second. That's only 0.14% of the data I'm expecting to consume.

Where did the rest of it go? If Flume is dropping it, why doesn't it tell me!?

Here's my flume.conf:



Thanks!




On Jan 15, 2013, at 2:31 PM, Andrew Otto <otto@wikimedia.org> wrote:


Would love some reviews, thanks!
-Andrew


On Jan 14, 2013, at 1:01 PM, Andrew Otto <otto@wikimedia.org> wrote:

Thanks guys! I've opened up a JIRA here:



On Jan 14, 2013, at 12:43 PM, Alexander Alten-Lorenz <wget.null@gmail.com> wrote:

Hey Andrew,

for your reference, we have a lot of developer informations in our wiki:


cheers,
Alex

On Jan 14, 2013, at 6:37 PM, Hari Shreedharan <hshreedharan@cloudera.com> wrote:

Hi Andrew,

Really happy to hear Wikimedia Foundation is considering Flume. I am fairly sure that if you find such a source useful, there would definitely be others who find it useful too. I'd recommend filing a jira and starting a discussion, and then submitting the patch. We would be happy to review and commit it.


Thanks,
Hari

--
Hari Shreedharan


On Monday, January 14, 2013 at 9:29 AM, Andrew Otto wrote:

Hi all,

I'm an Systems Engineer at the Wikimedia Foundation, and we're investigating using Flume for our web request log HDFS imports. We've previously been using Kafka, but have had to change short term architecture plans in order to get data into HDFS reliably and regularly soon.

Our current web request logs are available for consumption over a multicast UDP stream. I could hack something together to try and pipe this into Flume using the existing sources (SyslogUDPSource, or maybe some combination of socat + NetcatSource), but I'd rather reduce the number of moving parts. I'd like to consume directly from the multicast UDP stream as a Flume source.

I coded up proof of concept based on the SyslogUDPSource, mainly just stripping out the syslog event header extraction, and adding in multicast Datagram connection code. I plan on cleaning this up, and making this a generic raw UDP source, with multicast being a configuration option.

My question to you guys is, is this something the Flume community would find useful? If so, should I open up a JIRA to track this? I've got a fork of the Flume git repo over on github and will be doing my work there. I'd love to share it upstream if it would be useful.

Thanks!
-Andrew Otto
Systems Engineer
Wikimedia Foundation

--
Alexander Alten-Lorenz
German Hadoop LinkedIn Group: http://goo.gl/N8pCF



--
Apache MRUnit - Unit testing MapReduce - http://incubator.apache.org/mrunit/



--
Apache MRUnit - Unit testing MapReduce - http://incubator.apache.org/mrunit/