yihua opened a new issue, #20084: URL: https://github.com/apache/hudi/issues/20084
Spark serializes the upsert shuffle partitioner into every task on both sides of the shuffle, and the write and record creation functions into every task of their stages. They carry state the tasks never read: `UpsertPartitioner` holds the `HoodieTable`, the write config and the whole `WorkloadProfile` although `getPartition` only needs the bucket maps; the commit action executors captured by the write function keep their input records; and `HoodieCreateRecordUtils` captures the DataFrame, the write config and schemas. The partitioner also looks up small files with one Spark task per partition path, each building its own file system view. Proposal: keep driver-only state out of these closures, look up small files on the driver for partitions with inserts, and build record creation state on the driver. part of #20064 -- 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]
