[
https://issues.apache.org/jira/browse/SPARK-61440?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
L. C. Hsieh updated SPARK-61440:
--------------------------------
Fix Version/s: connect-rust-4.3.0
> Make the retries of the calls of the Spark Connect Rust client in one place,
> and do not panic on a stream deadline too far to be a time
> ---------------------------------------------------------------------------------------------------------------------------------------
>
> Key: SPARK-61440
> URL: https://issues.apache.org/jira/browse/SPARK-61440
> Project: Spark
> Issue Type: Improvement
> Components: Connect
> Affects Versions: connect-rust-4.3.0
> Reporter: L. C. Hsieh
> Priority: Major
> Fix For: connect-rust-4.3.0
>
>
> The Rust client (crates/spark-connect-core/src/client.rs) retries its calls
> in four hand-written loops: the unary and stream-opening calls (with_retry),
> the artifact upload, the background ReleaseExecute, and the reattach of a
> result stream. Each has its own copy of the decision "can the policy retry
> this failure, is a retry left, wait its backoff", so a change to the retry
> policy has to be made four times, and the loops can drift apart.
> This makes that decision in one place in retries.rs, as PySpark's Retrying: a
> retry of the policy with its backoff, and the retry made at once that PySpark
> raises as a RetryException, bounded by its elapsed time. The differences
> between the loops that are on purpose stay, as parameters: the calls of a
> client check that the client is still open before a retry and ReleaseExecute
> does not, an attempt of an upload can fail with an error that is not retried,
> and the reader of a reattachable stream shares the attempts of the policy
> among the failures of a response (SPARK-61415).
> Two related fixes in the same code:
> * The deadline of a stream is computed as Instant::now() + deadline at three
> places, which panics for a configured deadline too large to add to the time.
> A deadline that cannot be a time is never reached, so it is none.
> * INVALID_CURSOR.DISCONNECTED and INVALID_HANDLE.OPERATION_NOT_FOUND /
> SESSION_NOT_FOUND are recognized by their name in the message of the error,
> as PySpark does. The match stays, in one named function with tests.
> No other change of behavior.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]