Roshan,
Following your suggestion that is how I am addressing it at this moment :
——
import org.apache.storm.kafka.spout.KafkaSpout;
import org.apache.storm.kafka.spout.KafkaSpoutConfig;
import org.apache.storm.utils.Utils;
public class CustomKafkaSpout<K, V> extends KafkaSpout<K, V> {
private static final long serialVersionUID = 1L;
private Long spoutSleepInterval = null;
public CustomKafkaSpout(KafkaSpoutConfig<K, V> kafkaSpoutConfig, Long
sleepinterval) {
super(kafkaSpoutConfig);
if (sleepinterval != null) {
spoutSleepInterval = sleepinterval;
}
}
@Override
public void nextTuple() {
super.nextTuple();
if (spoutSleepInterval != null) {
Utils.sleep(spoutSleepInterval);
} // end if
}
}
---
IPVP
On July 21, 2017 at 9:41:18 PM, I PVP
([email protected]<mailto:[email protected]>) wrote:
Could the Resource Aware Scheduler be a way to solve this issue/use case?
On July 21, 2017 at 4:40:11 PM, Roshan Naik
([email protected]<mailto:[email protected]>) wrote:
I am thinking another way to handle this is to introduce a sleep inside the
nextTuple() for the case there was no emit ? … but that’s only if you have a
custom spout.
-roshan
From: Bobby Evans <[email protected]>
Reply-To: "[email protected]" <[email protected]>
Date: Friday, July 21, 2017 at 7:55 AM
To: I PVP <[email protected]>
Cc: "[email protected]" <[email protected]>
Subject: Re: Utils.sleep on KafkaSpout
Yes I would look at balancing them. You use case is not one that we have
thought about much and we might be able to make things much more efficient for
idle spouts, but it would take some work.
I would start off just by doubling the sleep time to 2ms instead of 1 and see
how that impacts your CPU usage. 10 or 20 ms probably would be fine in most
cases. You might even be able to get away with 100ms sleeps if the throughput
of your topologies is very low. My real concern was the 3 second sleep as it is
so very long that I would want you to be careful with it.
- Bobby
On Friday, July 21, 2017, 9:47:12 AM CDT, I PVP <[email protected]> wrote:
Thanks very much for the explanation.
So considering that my application is a MVP on beta usage ( very low traffic)
and I cannot afford to have all the servers needed to have all the +40
topologies running without starving CPU even when everything is idle, should I
focus on balancing these two settings (
TOPOLOGY_SLEEP_SPOUT_WAIT_STRATEGY_TIME_ and topology.max.spout.pending) or is
there a better way to adjust the resource consumption to the low usage that
my application has at this moment ?
Thanks
IP VP
On July 21, 2017 at 11:36:36 AM, Bobby Evans
([email protected]<mailto:[email protected]>) wrote:
That would slow down all of your spouts by a lot. If you just want it for a
single kafkaspout then you would want to set it only for that spout by calling
`addConfiguration(Config.TOPOLOGY_SLEEP_SPOUT_WAIT_STRATEGY_TIME_MS, 3000)` on
the SpoutDeclarer for that spout.
The issue is that the spout sleeps that amount of time when there is an empty
emit or if max spout pending was hit, or if back pressure said that the
topology should be throttled. Not emitting things is very common, even when
you are processing a normal amount of data. So for normal spouts you are
likely to see the spout pause for 3 seconds (your setting), then get a big
burst of data to process and if all 3 seconds of data cannot fit into
topology.max.spout.pending the spout will sleep again for 3 seconds and now you
have more then 3 seconds of data to process, which it is likely to now be able
to do.
- Bobby
On Friday, July 21, 2017, 7:17:16 AM CDT, Stig Rohde Døssing <[email protected]>
wrote:
Yes, that should work too.
2017-07-21 13:35 GMT+02:00 I PVP <[email protected]<mailto:[email protected]>>:
Would defining it for each topology with the following code be also a option or
is there any disadvantage of doing it this way?
--
org.apache.storm.Config conf = new Config();
….
conf.put(Config.TOPOLOGY_ SLEEP_SPOUT_WAIT_STRATEGY_ TIME_MS, 3000);
--
best,
IPVP
On July 21, 2017 at 4:21:52 AM, Stig Rohde Døssing
([email protected]<mailto:[email protected]>) wrote:
When a call to nextTuple on the spout doesn't emit any tuples, the spout
executor will sleep for a bit. The duration is set here
https://github.com/apache/stor m/blob/e38f936077ea9b3ba5cd568
b69335e0aac8369dd/conf/
defaults.yaml#L247<https://github.com/apache/storm/blob/e38f936077ea9b3ba5cd568b69335e0aac8369dd/conf/defaults.yaml#L247>,
you could increase it if you want.
2017-07-21 3:18 GMT+02:00 I PVP <[email protected]<mailto:[email protected]>>:
I am experiencing very High CPU usage with storm-kafka spout even when idle
for hours.
I changed all my Kafka Spouts to the new org.apache.storm.kafka.spo
ut.KafkaSpout but the issue continues.
How to tune it ?
Is there something like a Utils.sleep for KafkaSpout?
Thanks
IP VP
