[
https://issues.apache.org/jira/browse/FLINK-40115?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18095218#comment-18095218
]
Hancheng Deng commented on FLINK-40115:
---------------------------------------
Hi [~pmishra] , thanks for your interest in this issue.
I created this issue based on a watermark behavior I encountered, and I intend
to work on the documentation change myself.
The proposed change is a small documentation-only update to explain the
timestamp reliability assumption of bounded-out-of-orderness watermarks. No
code or API behavior will be changed.
Could a Flink committer please review the scope and assign this issue to me if
appropriate?
> Document timestamp reliability for bounded-out-of-orderness watermarks
> ----------------------------------------------------------------------
>
> Key: FLINK-40115
> URL: https://issues.apache.org/jira/browse/FLINK-40115
> Project: Flink
> Issue Type: Improvement
> Components: Documentation
> Reporter: Hancheng Deng
> Priority: Minor
>
> The documentation page "Generating Watermarks" includes an example of a
> bounded-out-of-orderness watermark generator that advances watermarks from
> the maximum observed event timestamp.
> This behavior is correct and efficient when event timestamps are reliable and
> the out-of-orderness bound matches the stream characteristics. However,
> streams may contain future timestamps, device clock errors, or backfilled
> records mixed with real-time records. In such cases, a single far-ahead
> timestamp can advance the watermark too aggressively and make subsequent
> records appear late.
> This issue proposes adding a short documentation note after the
> bounded-out-of-orderness watermark generator example to remind users that
> timestamps used for watermark generation should be reliable, and to mention
> mitigation options such as timestamp validation, a larger out-of-orderness
> bound, allowed lateness, late-record side outputs, or a custom
> WatermarkGenerator.
> This is a documentation-only improvement and does not change Flink behavior.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
