danny0405 commented on code in PR #12960: URL: https://github.com/apache/hudi/pull/12960#discussion_r2005155909
########## rfc/rfc-87/rfc-87.md: ########## @@ -24,9 +24,329 @@ ## Approvers - @danny0405 -- @xiarixiaoyao -- @yuzhaojing +- @cshuo ## Status: Claim -JIRA: [HUDI-8934](https://issues.apache.org/jira/browse/HUDI-8934) +Umbrella ticket: [HUDI-9075](https://issues.apache.org/jira/browse/HUDI-9075) + +## Abstract + +Building on RFC-84, which removed Avro from Flink’s pre-write operators, RFC-87 eliminates Avro from the write path to improve performance. +Current writes suffer from excessive Avro serialization/deserialization and in-memory storage of List<HoodieRecord>, causing high GC overhead. +This RFC replaces DataBucket’s list storage with Flink’s BinaryInMemorySortBuffer, enabling efficient sorting and iterator-based writes. +HoodieLogBlock is also refactored to separate deserialization from buffering. +Precombine field deduplication will now occur after sorting. Review Comment: We can sort if the cost is low, but we should keep at least the sequence for same record key records if the ordering value are the same. The current gains may be low because in the merging phase (with base file), the sequence would be broken anyway. -- 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]
