Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Flume >> mail # user >> Design problem while monitoring Flume


Copy link to this message
-
Re: Design problem while monitoring Flume
The most common cause of resending events from the source would be
failure to write to the channel. Most of the time this would be because
the channel is full.

An approach to collecting statistics will vary on what exactly you want
to do, but perhaps you could write metadata to headers in the
interceptor and than batch process the serialized headers after events
have actually been written. Or if you need to be realtime you can
replicate events to an additional path which leads to a custom sink that
collects statistics. So long as the sink doesn't "bounce"
events(rollback transactions) it shouldn't get any events resent.

One thing to keep in mind though is that flume in general only
guarantees delivery, it doesn't guarantee that stuff will only be
delivered once(though many components do only deliver once)

On 08/28/2013 04:09 PM, Anat Rozenzon wrote:
> Hi,
>
> I want to get some statistics out of Flume (For example, how many
> records were collected, How many files etc.).
> I've written my own interceptor that updates an MBean whenever records
> arrive.
>
> I've also written a MonitorServices that collects the data from the
> MBean every X minutes and send it to a database.
>
> My problem is that sometimes events are resent again from the source,
> I saw that while debugging.
> Not sure why... maybe because of a timeout while sending to the sink?
>
> Anyway, if this happens in production it will corrupt my statistics.
>
> Is there any way I can know that an event have failed reaching the
> sink eventhough it passed the interceptor?
> Is there a better place to collect such statistics than an interceptor?
>
> Thanks
> Anat