Hi,

Thank you for reply.

That topology doesn’t use reliable delivery
-> I thought I tested it with ack and backpressure function, but I will try
it again.

Try reducing the sleep time to something smaller like 5.
-> Thank you for advice.
Does "topology.max.spout.pending" help this situation?
In addition, I just wonder that people put sleeping time in spout in a real
deployment?

Thanks,
Junguk


On Wed, Oct 26, 2016 at 7:53 PM, P. Taylor Goetz <[email protected]> wrote:

> That topology doesn’t use reliable delivery, so there is nothing to
> throttle that spout without the `sleep()`. So the spout will emit as fast
> as it can, which is faster than the bolts in the topology can process them.
>
> Try reducing the sleep time to something smaller like 5.
>
> -Taylor
>
> > On Oct 25, 2016, at 10:51 PM, Junguk Cho <[email protected]> wrote:
> >
> > Hi, All.
> >
> > I evaluate Storm performance with WordCount example.
> >
> > In spout, I removed "Utils.sleep(100)" in RandomSentenceSpout.java.
> > So, it emitted sentences very fast.
> > Whenever I run WordCountTopology, "split" bolt dead with below messages
> and It is re-launched.
> >
> > 2016-10-25 20:37:39.395 STDERR [INFO] java.lang.OutOfMemoryError: GC
> overhead limit exceeded
> > 2016-10-25 20:37:39.398 STDERR [INFO] Dumping heap to artifacts/heapdump
> ...
> >
> > I increased worker.heap.memory.mb as 2GB.
> >
> > Obviously, split bolt was overloaded due to extensive inputs from a
> spout and simple execute logics (splitting a sentence into words and
> sending them to a count bolt).
> > So, I changed the number of splits and counts while I kept using 1
> spout, but I still saw this situation.
> >
> > Any advice will be welcome.
> >
> > Thanks,
> > Junguk
> >
> >
> >
>
>

Reply via email to