Hi Fran, did you observe actual data loss due to the problem you are describing or are you discussing a possible issue based on your observations?
AFAIK, Flink's Kafka consumer keeps track of the offsets itself and includes these in the checkpoints. In case of a recovery, it does not rely on the offsets which were committed back to Kafka but only on the offsets it checkpointed itself. Gordon (in CC) is familiar with all details of Flink's Kafka consumer and can give a more detailed answer. Best, Fabian 2017-07-19 16:55 GMT+02:00 Francisco Blaya <francisco.bl...@hivehome.com>: > Hi, > > We have a Flink job running on AWS EMR sourcing a Kafka topic and > persisting the events to S3 through a DateTimeBucketer. We configured the > bucketer to flush to S3 with an inactivity period of 5 mins.The rate at > which events are written to Kafka in the first place is very low so it is > easy for us to investigate how the Flink job would recover in respect to > Kafka offsets after the job gets cancelled or the Yarn session killed. > > What we found is that Flink acks Kafka immediately before even writing to > S3. The consequence of this seems to be that if the job gets cancelled > before the acked events are flushed to S3 then these are lost, they don't > get written when the job restarts. Flink doesn't seem to keep in its > checkpointed state the fact that it acked those events but never flushed > them to S3. Checkpoints are created every 5 seconds in S3. > > We've also tried to configure externalized checkpoints throught > "state.checkpoints.dir" configuration key and "env.getCheckpointConfig. > enableExternalizedCheckpoints(ExternalizedCheckpointCleanup.RETAIN_ON_CANCELLATION)" > in the job so that they don't automatically get cleaned up when the job > gets cancelled or the Yarn session killed. We can see the job uses a > restored checkpoint upon restart but still we get missing events in S3. > > Has anyone come across this behaviour before? Are we assuming something > wrong? > > We're using EMR 5.4.0 and Flink 1.2.0. > > Regards, > Fran > > hivehome.com <http://www.hivehome.com> > > > > Hive | London | Cambridge | Houston | Toronto > The information contained in or attached to this email is confidential and > intended only for the use of the individual(s) to which it is addressed. It > may contain information which is confidential and/or covered by legal > professional or other privilege. The views expressed in this email are not > necessarily the views of Centrica plc, and the company, its directors, > officers or employees make no representation or accept any liability for > their accuracy or completeness unless expressly stated to the contrary. > Centrica Connected Home Limited (company no: 5782908), registered in > England and Wales with its registered office at Millstream, Maidenhead > Road, Windsor, Berkshire SL4 5GD. >
