Yicong Huang created SPARK-55505:
------------------------------------
Summary: NumberFormatException in SQLExecution.withNewExecutionId0
due to re-reading EXECUTION_ROOT_ID_KEY
Key: SPARK-55505
URL: https://issues.apache.org/jira/browse/SPARK-55505
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 4.2.0
Reporter: Yicong Huang
{{SQLExecution.withNewExecutionId0}} can throw {{NumberFormatException: Cannot
parse null string}} at the line:
{code:scala}
val rootExecutionId = sc.getLocalProperty(EXECUTION_ROOT_ID_KEY).toLong
{code}
The current code checks if {{EXECUTION_ROOT_ID_KEY}} is null, sets it if so,
then *re-reads* it from local properties assuming it is non-null:
{code:scala}
if (sc.getLocalProperty(EXECUTION_ROOT_ID_KEY) == null) {
sc.setLocalProperty(EXECUTION_ROOT_ID_KEY, executionId.toString)
sc.addJobTag(executionIdJobTag(sparkSession, executionId))
}
val rootExecutionId = sc.getLocalProperty(EXECUTION_ROOT_ID_KEY).toLong //
crashes here
{code}
This re-read can return null under high-concurrency scenarios involving nested
thread pools (e.g., {{CrossValidator(parallelism=4)}} with
{{OneVsRest(parallelism=2)}} running from a Python {{ThreadPoolExecutor}}).
The fix is to read the property once and use the value directly, avoiding the
re-read:
{code:scala}
val existingRootId = sc.getLocalProperty(EXECUTION_ROOT_ID_KEY)
val rootExecutionId = if (existingRootId != null) {
existingRootId.toLong
} else {
sc.setLocalProperty(EXECUTION_ROOT_ID_KEY, executionId.toString)
sc.addJobTag(executionIdJobTag(sparkSession, executionId))
executionId
}
{code}
CI failure demonstrating the flaky test: [GitHub Actions
Run|https://github.com/Yicong-Huang/spark/actions/runs/21961599500/attempts/1]
(Attempt 1 failed, attempt 2 passed. Failed test:
{{test_save_load_pipeline_estimator}} in {{CrossValidatorIOPipelineTests}})
Full stack trace:
{code}
java.lang.NumberFormatException: Cannot parse null string
at java.lang.Long.parseLong(Long.java:550)
at SQLExecution.scala:115
{code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
