[
https://issues.apache.org/jira/browse/FLINK-40482?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Rui Fan updated FLINK-40482:
----------------------------
Fix Version/s: 2.4.0
> DefaultCheckpointPlanCalculator misclassifies FAILED/CANCELED tasks as
> finished
> -------------------------------------------------------------------------------
>
> Key: FLINK-40482
> URL: https://issues.apache.org/jira/browse/FLINK-40482
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Checkpointing
> Reporter: Rui Fan
> Assignee: Rui Fan
> Priority: Major
> Fix For: 2.4.0
>
>
> h3. Problem
> DefaultCheckpointPlanCalculator uses Execution.isFinished() (which returns
> state.isTerminal()) to decide whether a task is finished. This treats any
> terminal task — FINISHED, FAILED, or CANCELED — as finished.
> When a failover is in progress, some tasks may be FAILED/CANCELED. The
> calculator then classifies them as finished and excludes them from
> tasksToWaitFor, producing a checkpoint plan that expects fewer acks than
> it should (in the extreme, an empty plan).
> Today PendingCheckpoint rejects an empty tasksToWaitFor, so the worst case
> is masked. But the classification is wrong at its source: "terminal" is not
> the same as "finished", and relying on the downstream guard leaves the issue
> latent.
> h3. Fix
> - Classify explicitly in collectTaskRunningStatus: a non-terminal task is
> running; a terminal task must be genuinely FINISHED.
> - Abort the checkpoint with a graceful CheckpointException when a task is
> FAILED/CANCELED (i.e. a failover is underway) instead of building a plan
> from a transient state.
> - Remove the misleading Execution.isFinished() (its only caller).
> This is condition-based, not mode-based, so it is correct for both streaming
> and batch. calculateCheckpointPlan and all task state transitions run on the
> JM main thread, so the pre-check and plan computation are atomic.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
