GitHub user PrithwishDas101 added a comment to the discussion: DagFileProcessorManager Death Spiral and Defunct Scheduler Processes
Because the processes reportedly remain defunct or in an uninterruptible sleep state, I'd investigate the operating system and filesystem behavior alongside the Airflow scheduler logs. A few pieces of evidence would be especially useful: - The Airflow version, executor, operating system, and whether DAGs are stored on NFS or another shared filesystem. - The scheduler and DagFileProcessorManager logs immediately before the first heartbeat failure. - Process state and wait-channel information for any process stuck in uninterruptible sleep. - Whether the failure reproduces with a minimal DAG that performs no filesystem operations at module import time. I'd also test whether expensive imports, network calls, or filesystem access at DAG-file top level contribute to the failure. Those operations run during DAG parsing and can destabilize parsing throughput, although they would not by themselves explain every kind of uninterruptible process state. I wouldn't conclude that Airflow itself is unable to terminate the processes without first identifying the OS-level process state and the operation they're blocked on. That evidence should help distinguish a DAG parsing issue from a filesystem or kernel-level problem. GitHub link: https://github.com/apache/airflow/discussions/63749#discussioncomment-18950016 ---- This is an automatically sent email for [email protected]. To unsubscribe, please send an email to: [email protected]
