If you're looking for some kind of instrumentation finer than at batch boundaries, you'd have to do something with the individual messages yourself. You have full access to the individual messages including offset.
On Thu, Apr 27, 2017 at 1:27 PM, Dominik Safaric <dominiksafa...@gmail.com> wrote: > Of course I am not asking to commit for every message. But instead of, > seeking to commit the last consumed offset at a given interval. For example, > from the 1st until the 5th second, messages until offset 100.000 of the > partition 10 were consumed, then from the 6th until the 10th second of > executing the last consumed offset of the same partition was 200.000 - and so > forth. This is the information I seek to get. > >> On 27 Apr 2017, at 20:11, Cody Koeninger <c...@koeninger.org> wrote: >> >> Are you asking for commits for every message? Because that will kill >> performance. >> >> On Thu, Apr 27, 2017 at 11:33 AM, Dominik Safaric >> <dominiksafa...@gmail.com> wrote: >>> Indeed I have. But, even when storing the offsets in Spark and committing >>> offsets upon completion of an output operation within the foreachRDD call >>> (as pointed in the example), the only offset that Spark’s Kafka >>> implementation commits to Kafka is the offset of the last message. For >>> example, if I have 100 million messages, then Spark will commit only the >>> 100 millionth offset, and the offsets of the intermediate batches - and >>> hence the questions. >>> >>>> On 26 Apr 2017, at 21:42, Cody Koeninger <c...@koeninger.org> wrote: >>>> >>>> have you read >>>> >>>> http://spark.apache.org/docs/latest/streaming-kafka-0-10-integration.html#kafka-itself >>>> >>>> On Wed, Apr 26, 2017 at 1:17 PM, Dominik Safaric >>>> <dominiksafa...@gmail.com> wrote: >>>>> The reason why I want to obtain this information, i.e. <partition, >>>>> offset, timestamp> tuples is to relate the consumption with the >>>>> production rates using the __consumer_offsets Kafka internal topic. >>>>> Interestedly, the Spark’s KafkaConsumer implementation does not auto >>>>> commit the offsets upon offset commit expiration, because as seen in the >>>>> logs, Spark overrides the enable.auto.commit property to false. >>>>> >>>>> Any idea onto how to use the KafkaConsumer’s auto offset commits? Keep in >>>>> mind that I do not care about exactly-once, hence having messages >>>>> replayed is perfectly fine. >>>>> >>>>>> On 26 Apr 2017, at 19:26, Cody Koeninger <c...@koeninger.org> wrote: >>>>>> >>>>>> What is it you're actually trying to accomplish? >>>>>> >>>>>> You can get topic, partition, and offset bounds from an offset range like >>>>>> >>>>>> http://spark.apache.org/docs/latest/streaming-kafka-0-10-integration.html#obtaining-offsets >>>>>> >>>>>> Timestamp isn't really a meaningful idea for a range of offsets. >>>>>> >>>>>> >>>>>> On Tue, Apr 25, 2017 at 2:43 PM, Dominik Safaric >>>>>> <dominiksafa...@gmail.com> wrote: >>>>>>> Hi all, >>>>>>> >>>>>>> Because the Spark Streaming direct Kafka consumer maps offsets for a >>>>>>> given >>>>>>> Kafka topic and a partition internally while having enable.auto.commit >>>>>>> set >>>>>>> to false, how can I retrieve the offset of each made consumer’s poll >>>>>>> call >>>>>>> using the offset ranges of an RDD? More precisely, the information I >>>>>>> seek to >>>>>>> get after each poll call is the following: <timestamp, offset, >>>>>>> partition>. >>>>>>> >>>>>>> Thanks in advance, >>>>>>> Dominik >>>>>>> >>>>> >>> > --------------------------------------------------------------------- To unsubscribe e-mail: user-unsubscr...@spark.apache.org
