flume-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Paul Chavez <pcha...@verticalsearchworks.com>
Subject LoadBalancing Sink Processor question
Date Fri, 29 Mar 2013 20:20:21 GMT
I am curious about the observed behavior of a set of agents configured with a Load Balancing
sink processor.

I have 4 'tier1' agents receiving events directly from app servers that feed into 2 'tier2'
agents that write to HDFS. They are connected up via Avro Sink/Sources and a Load Balancing
Sink Processor.

Both 'tier2' agents write the same directory and I have observed they occasionally step on
each other and one of the tier2 agents at that point 'loses' and gets hung up on a file lease
exception. I'm not concerned with that at the moment as I know it's not best practices and
this is more of a pilot architecture.

My concern is that once a tier2 agent gets stuck it obviously fills it's channel in time,
and then stops accepting put requests from the Avro source. At this point my *expectation*
is that the upstream tier1 agents will continue to round-robin to the tier2 nodes with every
other 'put' request failing. Assuming the remaining tier2 node can handle the throughput (which
it can) I would not expect the tier1 agents to ever fill their channels.

In actuality what happens is the tier1 agents slowly fill the channel and eventually start
refusing put attempts from the application servers. It seems that once a given batch has been
allocated to the bad sink, it won't ever get released to be processed by the other, working
sink.

Is this the way it should work? Is this a defect or as designed? I will probably switch to
a failover processor because I really only need one HDFS writer to keep up with my data, but
I do think this isn't working as intended.

thanks,
Paul


Mime
View raw message