[
https://issues.apache.org/jira/browse/FLINK-13537?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Flink Jira Bot updated FLINK-13537:
-----------------------------------
Labels: auto-deprioritized-major stale-minor (was:
auto-deprioritized-major)
I am the [Flink Jira Bot|https://github.com/apache/flink-jira-bot/] and I help
the community manage its development. I see this issues has been marked as
Minor but is unassigned and neither itself nor its Sub-Tasks have been updated
for 180 days. I have gone ahead and marked it "stale-minor". If this ticket is
still Minor, please either assign yourself or give an update. Afterwards,
please remove the label or in 7 days the issue will be deprioritized.
> Changing Kafka producer pool size and scaling out may create overlapping
> transaction IDs
> ----------------------------------------------------------------------------------------
>
> Key: FLINK-13537
> URL: https://issues.apache.org/jira/browse/FLINK-13537
> Project: Flink
> Issue Type: Bug
> Components: Connectors / Kafka
> Affects Versions: 1.8.1, 1.9.0
> Reporter: Nico Kruber
> Priority: Minor
> Labels: auto-deprioritized-major, stale-minor
>
> The Kafka producer's transaction IDs are only generated once when there was
> no previous state for that operator. In the case where we restore and
> increase parallelism (scale-out), some operators may not have previous state
> and create new IDs. Now, if we also reduce the {{poolSize}}, these new IDs
> may overlap with the old ones which should never happen! Similarly, a
> scale-in + increasing {{poolSize}} could lead the the same thing.
> An easy "fix" for this would be to forbid changing the {{poolSize}}. We could
> potentially be a bit better by only forbidding changes that can lead to
> transaction ID overlaps which we can identify from the formulae that
> {{TransactionalIdsGenerator}} uses. This should probably be the first step
> which can also be back-ported to older Flink versions just in case.
> ----
> On a side note, the current scheme also relies on the fact, that the
> operator's list state distributes previous states during scale-out in a
> fashion that only the operators with the highest subtask indices do not get a
> previous state. This is somewhat "guaranteed" by
> {{OperatorStateStore#getListState()}} but I'm not sure whether we should
> actually rely on that there.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
