[
https://issues.apache.org/jira/browse/FLINK-40504?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18117790#comment-18117790
]
yongfu.gao commented on FLINK-40504:
------------------------------------
[~martijnvisser] The PR for this issue is up:
https://github.com/apache/flink/pull/29239
It addresses the root cause of the two mechanisms described in the ticket:
* WatermarkOutputMultiplexer now reports the idle -> active transition of the
combined status to
the underlying output (previously it only reported markIdle()), and a
watermark on a previously
idle multiplexed output triggers the combined update, so the re-activation is
no longer delayed
until the next periodic emit.
* registerNewOutput() reports the combined status immediately, which covers the
new-split case of
the same defect described in FLINK-22926.
* WatermarkToDataOutput.emitWatermark() marks the output active before the
monotonicity guard, as
required by the WatermarkOutput contract.
Verification: 7 new tests (4 in WatermarkOutputMultiplexerTest, 1 in
WatermarkToDataOutputTest and
2 end-to-end in SourceOperatorEventTimeTest). All 7 were verified to fail
without the corresponding
production change and to pass with it; a broader regression over 20
watermark-related test classes
(137 tests) passes on a clean build. Azure CI on ca8d2501 is green.
This is complementary to #29037 (FLINK-40499) and does not touch the files
changed there.
Could a committer please take a look, or point me to someone who can review it?
> FLIP-27 source never re-emits WatermarkStatus.ACTIVE after idleness unless
> its watermark strictly advances
> ----------------------------------------------------------------------------------------------------------
>
> Key: FLINK-40504
> URL: https://issues.apache.org/jira/browse/FLINK-40504
> Project: Flink
> Issue Type: Bug
> Components: API / Core, Runtime / Task
> Reporter: Martijn Visser
> Assignee: yongfu.gao
> Priority: Major
> Labels: pull-request-available
>
> Two compounding mechanisms leave a resumed FLIP-27 source announced-IDLE
> downstream:
> # {{WatermarkOutputMultiplexer.updateCombinedWatermark()}} calls
> {{underlyingOutput.markIdle()}} when combined-idle but has no
> {{underlyingOutput.markActive()}} counterpart; the per-output {{markActive}}
> implementations only mutate internal partial state.
> # {{WatermarkToDataOutput.emitWatermark}} returns at the monotonicity guard
> ({{newWatermark <= maxWatermarkSoFar}}) *before* {{markActiveInternally()}}
> ({{WatermarkToDataOutput.java:77-85}}).
> The only path back to ACTIVE is therefore a strictly larger watermark.
> Scenario: all
> splits of a subtask go idle; the combined watermark is flushed to the max M
> across
> splits (FLINK-38454). A split then resumes with backlog whose watermarks are
> <= M
> (e.g. a recovered Kafka partition catching up). Result — proven by red unit
> tests on
> both layers: *zero* calls reach the underlying output (no watermark, no
> ACTIVE), even
> via {{onPeriodicEmit}}; the edge stays IDLE downstream indefinitely while
> records flow.
> Downstream min-watermark is then driven by the other inputs, past the resumed
> backlog,
> and its records are dropped as late.
> Every sibling path re-activates eagerly on resume: {{StatusWatermarkValve}}
> (explicitly,
> since FLINK-40475), the legacy {{StreamSourceContexts}} (ACTIVE on every
> collect), the
> table {{WatermarkAssignerOperator}} (ACTIVE on first record), and DataStream
> V2
> (explicit idle=false watermark on next record). The FLIP-27 multiplexer path
> is the only
> one without a re-activation signal.
> FLINK-22926 (open since 2021) describes the new-split registration special
> case of this
> same defect; it should be closed into this ticket.
> Note: the fix is a semantics decision (emit ACTIVE eagerly on
> record/markActive, aligning
> with all other paths) rather than a local patch - it also composes with
> FLINK-40499: a
> resumed-but-not-advanced source is exactly the still-IDLE state in which its
> final
> MAX_WATERMARK is then dropped at drain.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
