[
https://issues.apache.org/jira/browse/FLINK-5017?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15644877#comment-15644877
]
Aljoscha Krettek commented on FLINK-5017:
-----------------------------------------
Oh dear, this is turning out to be quite tricky.
My current thinking is that we can completely bypass operators if we have IDLE
as watermark input. For example, consider this chain of operators:
Kafka Source -> op1 -> op2 -> timestamp extractor
op1 and op2 can be any operation (map,flatmap, what have you...).
None of the operations op1, op2 or the timestamp extractor would actually need
to be informed of the IDLE watermark. The intermediate operations cannot do
anything with it, event the timestamp extractor cannot do anything but forward
it. Therefore I'm thinking to just bypass it.
The only place where it is relevant is in the watermark advancement code in the
input processors. If all inputs are IDLE, then the input processor needs to
bypass the chain and directly output to all real outgoing connections, which in
itself is quite tricky.
> Introduce WatermarkStatus stream element to allow for temporarily idle
> streaming sources
> ----------------------------------------------------------------------------------------
>
> Key: FLINK-5017
> URL: https://issues.apache.org/jira/browse/FLINK-5017
> Project: Flink
> Issue Type: New Feature
> Components: Streaming
> Reporter: Tzu-Li (Gordon) Tai
> Assignee: Tzu-Li (Gordon) Tai
> Priority: Blocker
> Fix For: 1.2.0
>
>
> A {{WatermarkStatus}} element informs receiving operators whether or not they
> should continue to expect watermarks from the sending operator. There are 2
> kinds of status, namely {{IDLE}} and {{ACTIVE}}. Watermark status elements
> are generated at the sources, and may be propagated through the operators of
> the topology using {{Output#emitWatermarkStatus(WatermarkStatus)}}.
> Sources and downstream operators should emit either of the status elements
> once it changes between "watermark-idle" and "watermark-active" states.
> A source is considered "watermark-idle" if it will not emit records for an
> indefinite amount of time. This is the case, for example, for Flink's Kafka
> Consumer, where sources might initially have no assigned partitions to read
> from, or no records can be read from the assigned partitions. Once the source
> detects that it will resume emitting data, it is considered
> "watermark-active".
> Downstream operators with multiple inputs (ex. head operators of a
> {{OneInputStreamTask}} or {{TwoInputStreamTask}}) should not wait for
> watermarks from an upstream operator that is "watermark-idle" when deciding
> whether or not to advance the operator's current watermark. When a downstream
> operator determines that all upstream operators are "watermark-idle" (i.e.
> when all input channels have received the watermark idle status element),
> then the operator is considered to also be "watermark-idle", as it will
> temporarily be unable to advance its own watermark. This is always the case
> for operators that only read from a single upstream operator. Once an
> operator is considered "watermark-idle", it should itself forward its idle
> status to inform downstream operators. The operator is considered to be back
> to "watermark-active" as soon as at least one of its upstream operators
> resume to be "watermark-active" (i.e. when at least one input channel
> receives the watermark active status element), and should also forward its
> active status to inform downstream operators.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
