Hyukjin Kwon created SPARK-61184:
------------------------------------
Summary: Send the server-side session ID that the client observed,
and check that of the responses, so that a session ends when its server
restarts, as PySpark does
Key: SPARK-61184
URL: https://issues.apache.org/jira/browse/SPARK-61184
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: connect-rust-4.3.0
Reporter: Hyukjin Kwon
PySpark Connect keeps the server-side session ID of the responses and sends it
as client_observed_server_side_session_id on its requests (ExecutePlan,
AnalyzePlan, Config, AddArtifacts, ArtifactStatus, Interrupt, ReattachExecute,
ReleaseExecute, FetchErrorDetails). After a restart of the server, the server
raises INVALID_HANDLE.SESSION_CHANGED ("The existing Spark server driver
instance has restarted. Please reconnect."), and the client closes its session,
so the next calls raise NO_ACTIVE_SESSION. The client also checks the
server-side session ID of each response (_verify_response_integrity) and raises
a PySparkAssertionError when it changes.
The Rust client neither sends nor checks it. After a restart of the server, a
session silently goes on in a new session of the server: its configurations are
back to their defaults and its temporary views are gone. For example, after
spark.conf.set("spark.sql.shuffle.partitions", "7") and a restart of the
server, spark.conf.get("spark.sql.shuffle.partitions") gives "200", where
PySpark raises NO_ACTIVE_SESSION.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
