Jiwon Park created SPARK-58447:
----------------------------------
Summary: Preserve operation tags in Spark Connect Python
foreachBatch callbacks
Key: SPARK-58447
URL: https://issues.apache.org/jira/browse/SPARK-58447
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: 5.0.0
Reporter: Jiwon Park
Spark Connect does not preserve operation tags from the `writeStream.start()`
request in Python `foreachBatch` callbacks.
The callback runs in a separate Python worker that creates a local Connect
client. Because operation tags are client thread-local, the callback session
does not contain the tags from the originating request. Consequently, nested
actions such as `count`, writes, and SQL commands are executed without those
tags.
A minimal reproduction is:
spark.addTag("foreach-batch")
def process_batch(df, _):
assert "foreach-batch" in df.sparkSession.getTags()
df.count()
query = stream.writeStream.foreachBatch(process_batch).start()
query.processAllAvailable()
The assertion fails with Spark Connect. The callback session should inherit the
operation tags captured when the streaming query starts.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
