On Sep 20, 2011, at 3:02 PM, Derek Deeter wrote:
> We have set up a configuration that uses the collector sink with our own
> sink (called GlobalLogSink). Configuration is:
>
> collector (15000) { GlobalLogSink(...) }
>
> The GlobalLogSink connects to Cassandra which we have not fully tuned yet.
> We are running into an exception in the RollSink code that causes
> DirectDriver to close down, which occurs when Cassandra does a Garbage
> Collection or some other time intensive action that causes a response to be
> longer than necessary. I have pasted two tracebacks at the end of this
> email, one from the collector and the other from jstack.
>
> The code at RollSink:296 does a write with a wait set to 1000 millis; if
> the write takes more than that, an exception occurs that shuts the sink
> down.
>
> When testing in our development environment, we found that this problem went
> away, but there we were using the first flume-0.9.4 release, which has the
> 'synchronize' functionality before the ReentrantReadWriteLock was
> implemented, so we had to back down a release or two from flume-0.9.4+25.6,
> but then we ran into Ack problems.
>
> There are a couple fixes needed to correct this issue:
> We would like to add a fix to RollSink.java to make the 1000ms delay
> configurable.
> Also, when the timeout does occur, the Sink should fail gracefully and not
> just hang/close down.
> I would like to open a JIRA on this.
>
> To fix the problem I need to know what source/branch to use - this error
> occurs in the source associated with the flume-0.9.4+25 RPM file where we
> found source at 0.9.4-cdh3u1 that corresponds to that. Is there a branch
> label or tag that we should use to fix this?
>
> A couple of related questions we'd appreciate answers to:
> 1) Is there a way to set up a collector sink that doesn't use RollSink?
> 2) Is there a way to get acknowledgments without using RollSink?
>
> Ideally, we would like to have something like an 'AckSink' that's lets us
> utilize E2E with out having the RollSink functionality.
>
> Thanks in advance,
> Derek Deeter
I'd really like some information as to why the CollectorSink is hard coded to
have a RollSink. Although the use described above wants acknowledgements it
isn't clear why closing and reopening is necessary. I would think that in this
use case the locking wouldn't be necessary just to send the acknowledgements.
Can someone explain why the code works this way?
Ralph
