[
https://issues.apache.org/jira/browse/BEAM-2467?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ismaël Mejía updated BEAM-2467:
-------------------------------
Component/s: (was: z-do-not-use-sdk-java-extensions)
io-java-kinesis
> KinesisIO watermark based on approximateArrivalTimestamp
> --------------------------------------------------------
>
> Key: BEAM-2467
> URL: https://issues.apache.org/jira/browse/BEAM-2467
> Project: Beam
> Issue Type: Improvement
> Components: io-java-kinesis
> Reporter: Paweł Kaczmarczyk
> Assignee: Paweł Kaczmarczyk
> Priority: Major
> Fix For: 2.2.0
>
>
> In Kinesis we can start reading the stream at some point in the past during
> the retention period (up to 7 days). With current approach for setting
> record's timestamp and watermark (both are always set to current time, i.e.
> Instant.now()), we can't observe the actual position in the stream.
> So the idea is to change this behaviour and set the record timestamp based on
> the
> [ApproximateArrivalTimestamp|http://docs.aws.amazon.com/kinesis/latest/APIReference/API_Record.html#Streams-Type-Record-ApproximateArrivalTimestamp].
> Watermark will be set accordingly to the last read record's timestamp.
> ApproximateArrivalTimestamp is still some approximation and may result in
> having records with out-of-order timestamp's which in turn may result in some
> events marked as late. This however should not be a frequent issue and even
> if it happens it should be a matter of milliseconds or seconds so can be
> handled even with a tiny allowedLateness setting
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
