DanielLeens commented on PR #13060:
URL: https://github.com/apache/seatunnel/pull/13060#issuecomment-6851079531

   Thanks @nzw921rx and @SEZ9. I agree that a blanket `EXACTLY_ONCE` 
parallelism cap would remove existing functionality and would not repair Spark 
retry identity even at parallelism one.
   
   Some of the named concerns refer to code that is no longer present in 
`7c272c5a`. In `KafkaSinkWriter.java:243-296`, restore consumes the states 
assigned by the engine; there is no `validateRestoredStateOwner` 
modulo-ownership rejection. `parseRestoredTransaction` validates each ID 
against its own saved prefix (362-394), so different generated prefixes are not 
rejected merely for being different. Stateful recovery advances the adopted 
namespace revision (299-316). For a writer without retained state, sequence one 
is intentional: opening it fences its earlier holder; later abandoned 
transactions not represented by checkpoint state remain subject to Kafka's 
transaction timeout. That boundary is explicitly documented in 
`docs/en/connectors/sink/Kafka.md:104`, rather than claiming checkpoint-id 
alignment solves it. The constructor uses `getStableSubtaskIndex()` (104-106), 
and the Spark retry/speculation boundary is documented at line 108. The option 
row at line 66 describes the writer
 -specific ID shape.
   
   These pointers reconcile F1/F2/F3/F4/F7 with the author's ledger. F5/F6 
remain a separate efficiency follow-up. I am not dismissing @SEZ9's independent 
review or claiming new broker execution; this is a targeted reply on the 
unchanged head.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to