Hi Hugo Hope I do not come across as arguing for its own sake.
1.1 Agreed but I was not suggesting to use KafkaConsumer from multiple threads. We can use a single thread which is different from spout thread. That thread and spout will share a queue of ConsumerRecords. 1.2 Agreed again. I had taken the spout recommendation about non-blocking calls seriously, taking into account its critical role. I still believe it is a sound recommendation but may be there are other factors like topology throughput which might be more limiting than the spout making a Kafka consumer poll. 2.1 There is no room for an optimization which effects normal (non-failure) case negatively. I was just suggesting if we can optimize the failure case without a trade-off. Failure processing is quite independent of successfully acked messages. I am especially thinking of the failure case which, for some reason, is recurrent; the same message which does not get acked. I agree that it is much better to address the issue in the application. 2.2 As you have correctly mentioned, subsequently polled messages (previously acked/emitted) are no longer emitted. We cannot commit unless that failed message is finally acked after few retries. Due to which, we don't even process any new set of records. We anyway keep the current set of messages (polled repeatedly) in memory. All we are doing is polling the same set of messages and iterating till that one failure is rescheduled for N times (worse with exponential back off). Anyway, we are dealing with rare scenarios which should get the least priority especially if it introduces complexity. On a different note, sending arbitrarily (large) sized messages to Kafka is a problem in itself. @Stig I will think about it further and evaluate the need for optimization. Thanks Chandan On Tue, Jul 11, 2017 at 1:37 AM, chandan singh <cks07...@gmail.com> wrote: > 1) Agreed that short duration blocks are less likely to cause an issue but > it will be much better to avoid them. Anyway, a conclusion is not easy > without some benchmark. I will let You know if I am able to do some volume > testing on both options and observe significant difference. > > The auto spout block/wait when no messages are emitted makes absolute > sense but a different scenario, less likely in a mostly loaded topology. > > You are absolutely correct about decoupling the consumption from > processing when you see it from the perspective of Storm. I think there is > a very subtle difference when we keep Kafka in mind. It is the spout thread > which is polling (consuming) and iterating (processing) over the polled > messages. If we separate the consumption in another thread and push > messages in a queue, iterating (processing) is now concurrent and > decoupled. Anyway, I too feel there should not be any significant > difference in throughput but it will be interesting to measure the > difference. > > I went overboard while mentioning seek before every poll when I had the > failure scenario in mind. > > Thanks a lot. I will keep you guys updated on the switch from custom spout > to KafkaSpout. > > Thanks for the amazing work. > > Chandan > > On Tue, Jul 11, 2017 at 12:56 AM, chandan singh <cks07...@gmail.com> > wrote: > >> Sorry about being cryptic there. What I meant is that it will be much >> better if we don't make assumptions about frequency of failure rates in >> topologies. I know it is more of a commonsense but out of curiosity, can >> you point me to any Storm documentation which makes a comment on preferable >> failure rates. I was suggesting if we can offer the user an optimization >> through clean API, the user will be free to decide on the rationale of >> using it. >> >> On Tue, Jul 11, 2017 at 12:06 AM, Bobby Evans < >> ev...@yahoo-inc.com.invalid> wrote: >> >>> I'm not sure what assumptions you want to make that this is preventing, >>> or why they would be helpful. >>> >>> - Bobby >>> >>> >>> On Monday, July 10, 2017, 12:14:53 PM CDT, chandan singh < >>> cks07...@gmail.com> wrote: >>> >>> Hi Stig & Bobby >>> >>> Thanks for confirming my understanding. >>> >>> 1) Ensuring that calls to nexTuple(), ack() and fail() are non-blocking >>> has been a guideline on http://storm.apache.org/releas >>> es/1.1.0/Concepts.html >>> for long. Copying verbatim here : "The main method on spouts is >>> nextTuple. >>> nextTuple either emits a new tuple into the topology or simply returns if >>> there are no new tuples to emit. It is imperative that nextTuple does not >>> block for any spout implementation, because Storm calls all the spout >>> methods on the same thread." I admit that there is some chance my >>> interpretation is partially incorrect but I have been following it in a >>> custom spout till now. Even though the objective is different, there is a >>> similar hint on Kafka official documentation. Please see under heading >>> "2. >>> Decouple Consumption and Processing" on >>> https://kafka.apache.org/0110/javadoc/index.html?org/apache/ >>> kafka/clients/consumer/KafkaConsumer.html. >>> Essentially, a thread polls Kafka and spout thread gets the messages >>> through a shared queue. If pre-fetching is present in Kafka (I will read >>> about it further), I assume we do not have to fetch in another thread >>> but I >>> am not sure how does the pre-fetching behave with re-seeking before every >>> poll. >>> >>> 2) @Bobby, you are correct in pointing what needs to be optimized but the >>> facts, sometimes, prevent us from making assumptions. We do optimize our >>> retry loop such that we don't poll the messages again. I especially see >>> problems when combined with exponential back off. I am not sure how >>> difficult or clean will it be to expose some sort of configuration to >>> allow >>> such optimization? Do you think it will be worth trying out something? >>> >>> Thanks >>> Chandan >>> >> >> >
