[jira] [Commented] (BEAM-7978) ArithmeticExceptions on getting backlog bytes

["Kenneth Knowles (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/BEAM-7978?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16923775#comment-16923775 ] 

Kenneth Knowles commented on BEAM-7978:
---------------------------------------

I just happened to dig in to this, and really Minutes.between(instant, instant) is not safe unless you handle the overflow and convert to UNKNOWN behavior. Localizing to this is, in the simplified kinesis client, seems best. That keeps the troublesome behavior to a minimum. It is better than forcing a complex spec outside this local function.

> ArithmeticExceptions on getting backlog bytes 
> ----------------------------------------------
>
>                 Key: BEAM-7978
>                 URL: https://issues.apache.org/jira/browse/BEAM-7978
>             Project: Beam
>          Issue Type: Bug
>          Components: io-java-kinesis
>    Affects Versions: 2.14.0
>            Reporter: Mateusz
>            Assignee: Alexey Romanenko
>            Priority: Major
>             Fix For: 2.16.0
>
>          Time Spent: 40m
>  Remaining Estimate: 0h
>
> Hello,
> Beam 2.14.0
>  (and to be more precise [commit|https://github.com/apache/beam/commit/9fe0a0387ad1f894ef02acf32edbc9623a3b13ad#diff-b4964a457006b1555c7042c739b405ec]) introduced a change in watermark calculation in Kinesis IO causing below error:
> {code:java}
> exception:  "java.lang.RuntimeException: Unknown kinesis failure, when trying to reach kinesis
> 	at org.apache.beam.sdk.io.kinesis.SimplifiedKinesisClient.wrapExceptions(SimplifiedKinesisClient.java:227)
> 	at org.apache.beam.sdk.io.kinesis.SimplifiedKinesisClient.getBacklogBytes(SimplifiedKinesisClient.java:167)
> 	at org.apache.beam.sdk.io.kinesis.SimplifiedKinesisClient.getBacklogBytes(SimplifiedKinesisClient.java:155)
> 	at org.apache.beam.sdk.io.kinesis.KinesisReader.getTotalBacklogBytes(KinesisReader.java:158)
> 	at org.apache.beam.runners.dataflow.worker.StreamingModeExecutionContext.flushState(StreamingModeExecutionContext.java:433)
> 	at org.apache.beam.runners.dataflow.worker.StreamingDataflowWorker.process(StreamingDataflowWorker.java:1289)
> 	at org.apache.beam.runners.dataflow.worker.StreamingDataflowWorker.access$1000(StreamingDataflowWorker.java:149)
> 	at org.apache.beam.runners.dataflow.worker.StreamingDataflowWorker$6.run(StreamingDataflowWorker.java:1024)
> 	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
> 	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
> 	at java.lang.Thread.run(Thread.java:745)
> Caused by: java.lang.ArithmeticException: Value cannot fit in an int: 153748963401
> 	at org.joda.time.field.FieldUtils.safeToInt(FieldUtils.java:229)
> 	at org.joda.time.field.BaseDurationField.getDifference(BaseDurationField.java:141)
> 	at org.joda.time.base.BaseSingleFieldPeriod.between(BaseSingleFieldPeriod.java:72)
> 	at org.joda.time.Minutes.minutesBetween(Minutes.java:101)
> 	at org.apache.beam.sdk.io.kinesis.SimplifiedKinesisClient.lambda$getBacklogBytes$3(SimplifiedKinesisClient.java:169)
> 	at org.apache.beam.sdk.io.kinesis.SimplifiedKinesisClient.wrapExceptions(SimplifiedKinesisClient.java:210)
> 	... 10 more
> {code}
> We spotted this issue on Dataflow runner. It's problematic as inability to get backlog bytes seems to result in constant recreation of KinesisReader.
> The issue happens if the backlog bytes are retrieved before watermark value is updated from initial default value. Easy way to reproduce it is to create a pipeline with Kinesis source for a stream where no records are being put. While debugging it locally, you can observe that the watermark is set to the value on the past (like: "-290308-12-21T19:59:05.225Z"). After two minutes (default watermark idle duration threshold is set to 2 minutes) , the watermark is set to value of [watermarkIdleThreshold|https://github.com/apache/beam/blob/9fe0a0387ad1f894ef02acf32edbc9623a3b13ad/sdks/java/io/kinesis/src/main/java/org/apache/beam/sdk/io/kinesis/WatermarkPolicyFactory.java#L110]), so the next backlog bytes retrieval should be correct. However, as described before, running the pipeline on Dataflow runner results in KinesisReader being closed just after creation, so the watermark won't be fixed.
> The reason of the issue is following: The introduced watermark policies are relying on [WatermarkParameters|https://github.com/apache/beam/blob/9fe0a0387ad1f894ef02acf32edbc9623a3b13ad/sdks/java/io/kinesis/src/main/java/org/apache/beam/sdk/io/kinesis/WatermarkParameters.java] which initialises currentWatermark and eventTime to [BoundedWindow.TIMESTAMP_MIN_VALUE|https://github.com/apache/beam/commit/9fe0a0387ad1f894ef02acf32edbc9623a3b13ad#diff-b4964a457006b1555c7042c739b405ecR52]. This result in watermark being set to new Instant(-9223372036854775L) at the KinesisReader creation. Calculated [period between the watermark and the current timestamp|https://github.com/apache/beam/blob/9fe0a0387ad1f894ef02acf32edbc9623a3b13ad/sdks/java/io/kinesis/src/main/java/org/apache/beam/sdk/io/kinesis/SimplifiedKinesisClient.java#L169] is bigger than expected causing the ArithmeticException to be thrown.
> The maximum retention on Kinesis streams is  [7 days|https://aws.amazon.com/kinesis/data-streams/faqs/] so it should be safe to initialise the affected watermark parameters with new Instant().minus(MAX_KINESIS_STREAM_RETENTION_PERIOD) where MAX_KINESIS_STREAM_RETENTION_PERIOD is the duration of 7 days.
> Remark: seems that in the past there was similar issue present (fixed in 2.4). Please look into the [ticket|https://issues.apache.org/jira/browse/BEAM-3881]. 
> Best regards,
>  Mateusz



--
This message was sent by Atlassian Jira
(v8.3.2#803003)