L. C. Hsieh created SPARK-61764:
-----------------------------------
Summary: Fix Int overflow in the socket read timeout for the
Python worker idle timeout in pipelined mode
Key: SPARK-61764
URL: https://issues.apache.org/jira/browse/SPARK-61764
Project: Spark
Issue Type: Bug
Components: Spark Core
Affects Versions: 5.0.0
Reporter: L. C. Hsieh
The pipelined Python UDF path (BasePythonRunner.createPipelinedDataIn,
SPARK-56642) sets the socket read timeout used for the Python worker idle
timeout as
{code}
worker.channel.socket().setSoTimeout(if (idleTimeoutSeconds > 0)
idleTimeoutSeconds.toInt * 1000 else 0)
{code}
This is Int arithmetic, so it overflows when
spark.python.worker.idleTimeoutSeconds (or
spark.sql.execution.pyspark.udf.idleTimeoutSeconds) is above 2147483 seconds
(about 24.8 days). Depending on the value, the result is:
* negative (e.g. 2147484 gives -2147483296): Socket.setSoTimeout throws
IllegalArgumentException and the task fails;
* a much shorter timeout (e.g. 4294968 gives 704 ms): spurious idle timeouts
are logged, and healthy workers are killed when killOnIdleTimeout is set;
* 0 (e.g. 536870912 or 4294967296): the idle timeout is silently disabled.
The synchronous path and PythonPlannerRunner convert with
TimeUnit.SECONDS.toMillis on a Long and are not affected.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]