Martijn Visser created FLINK-41077:
--------------------------------------
Summary: RpcEndpoint.MainThreadExecutor#schedule swallows
exceptions thrown by the scheduled task
Key: FLINK-41077
URL: https://issues.apache.org/jira/browse/FLINK-41077
Project: Flink
Issue Type: Bug
Components: Runtime / RPC
Affects Versions: 2.2.1, 2.3.0, 1.20.6, 2.4.0
Reporter: Martijn Visser
Assignee: Martijn Visser
{{RpcEndpoint.MainThreadExecutor#schedule}} wraps the task in a {{FutureTask}}
and runs that in the main thread. {{FutureTask#run}} catches every
{{Throwable}}, so an exception thrown by a scheduled task is stored in the
returned future and never reaches {{PekkoRpcActor#handleRunAsync}}, which logs
failures of {{execute()}}-submitted runnables and rethrows fatal errors. A
scheduled task that throws, including an {{OutOfMemoryError}}, is silent unless
the caller calls {{get()}} on the returned future, and no in-tree caller does:
{{HeartbeatManagerSenderImpl}} (a throwing heartbeat tick ends the heartbeat
loop with no log line), {{DefaultBlocklistHandler}},
{{DeclarativeSlotPoolBridge}}, {{DefaultDeclarativeSlotPool}},
{{PhysicalSlotRequestBulkCheckerImpl}}, {{ExecutionTimeBasedSlowTaskDetector}},
{{AdaptiveScheduler}}, and the driver executor in {{ActiveResourceManager}}.
{{schedule()}} should surface the failure to the actor the same way
{{execute()}} does while keeping it in the returned future. After the change a
fatal error in any of these tasks exits the JVM like one in an
{{execute()}}-submitted runnable does today.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
