Hyukjin Kwon created SPARK-61355:
------------------------------------
Summary: Retry the requests that a connection closed before
sending them, as PySpark does
Key: SPARK-61355
URL: https://issues.apache.org/jira/browse/SPARK-61355
Project: Spark
Issue Type: Bug
Components: Connect
Affects Versions: connect-rust-4.3.0
Reporter: Hyukjin Kwon
After a connection reset, a request that the Rust client had queued on the
HTTP/2 connection but not sent yet failed with CANCELLED "operation was
canceled". hyper cancels such a request when its connection closes, tonic
reports hyper's canceled error as CANCELLED, and RetryPolicy::can_retry does
not retry CANCELLED. For example,
collect_reattaches_after_three_connection_resets in e2e_reattach_drop.rs failed
at its collect in 1 of 40 runs with 2 test threads. In PySpark 4.3.0,
grpc-python fails a call whose connection breaks with UNAVAILABLE, which
DefaultPolicy retries; such a request was never sent, so sending it again is
safe.
The fix retries a CANCELLED whose source is hyper's canceled error, which hyper
raises only for a request that its connection closed before sending. A
CANCELLED that the server sent, which has no source, and tonic's timeout of a
request, which is CANCELLED too, are still not retried.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]