deepakpanda93 commented on issue #8474:
URL: https://github.com/apache/hudi/issues/8474#issuecomment-4844850117

   Closing this as stale — it was never reproduced (the requested 
duplicate-record meta fields were never shared, and it only manifested 
transiently on executor loss).
   
   That said, on the substance: your hypothesis is right — on a stage re-run 
after executor loss, a BLOOM-index tagLocation recompute can re-classify an 
update as an insert. Current Hudi mitigates this by persisting the tagged 
records before building the workload profile and writing (so the classification 
and the write use the same materialized tagging), and the broader "recompute on 
retry" robustness is being hardened 
([#15017](https://github.com/apache/hudi/issues/15017), and retry test coverage 
in [#16717](https://github.com/apache/hudi/issues/16717) for 1.1.0).
   
   The most reliable way to avoid this class of issue is to use a 
deterministic, persisted index — the Record Level Index 
(hoodie.index.type=RECORD_INDEX, backed by the metadata table) — so tag lookups 
read stable index entries rather than recomputing a bloom scan that can vary 
across retries.
   
   If you can reproduce duplicate keys after executor loss on a recent 1.x 
release (ideally with RLI), please reopen with the _hoodie_record_key / 
_hoodie_file_name of the duplicates and the timeline, and we'll dig in. Thanks!


-- 
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