LuciferYang commented on PR #1352:
URL: https://github.com/apache/paimon-rust/pull/1352#issuecomment-6816455723

   Thanks for the write-buffer repro. The contract I'd go with: maps match by 
key in every representation and at any depth. That is what Java's field-wise 
equaliser already does for top-level MAP fields and with 
`row-deduplicate-ignore-fields`, and what this PR does.
   
   The difference comes from Java's byte shortcut, so I fixed it there: 
apache/paimon#11110 (issue apache/paimon#11109) stops comparing `BinaryRow` / 
`BinaryArray` values by bytes when they hold maps or multisets, as #8534 did 
for FLOAT/DOUBLE. Equal bytes still return equal right away. As for the version 
boundary, `collect` `count-limit` is only on master, so with 
apache/paimon#11110 in, no released Java version persists the other result.
   
   To cover the buffered path, I replayed 15k random agg / aggReversed / DELETE 
/ retract sequences on `ARRAY<ARRAY<MAP<INT,INT>>>` with binary inputs, as the 
write buffer passes them, and 15k with generic inputs. With 
apache/paimon#11110, Java gives the same results for both, and Rust matches 
every step. Current Java master differs on 1087 of the binary-input sequences.
   


-- 
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