AndriiSushko commented on issue #18758:
URL: https://github.com/apache/iceberg/issues/18758#issuecomment-6694662204

   Reproduced with DuckDB 2.0.0's Iceberg extension as the client, on 
`apache/iceberg-rest-fixture:latest` built from `b0df3ca`, using the JDBC 
catalog on SQLite.
   
   Two `MERGE` statements committed through `/v1/transactions/commit`, with a 
proxy releasing both commit requests at the same moment. Both returned 204, and 
one write was lost in 12 of 12 rounds:
   - guarded on empty `main`, same new row (4 of 4)
   - guarded on an existing `main` snapshot, different new rows (4 of 4)
   - guarded on an existing `main` snapshot, update to the same row (4 of 4)
   
   So this also affects commits on a table with existing data, including 
appends of unrelated rows, and not only a guard on empty `main`. The catalog 
log showed the same sequence: `Cannot commit ... metadata location has 
changed`, `Retrying task after failure: sleepTimeMs=100`, `Nothing to commit.`
   
   When the second commit was released after the first had finished, it got the 
expected 409.
   
   For comparison, AWS S3 Tables with the same client lost no writes in 36 
rounds of concurrent merges, though those rounds were timed from the client 
side and not forced through a proxy.


-- 
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]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to