wzx140 opened a new pull request, #8051:
URL: https://github.com/apache/paimon/pull/8051

   ### Purpose
   
   `firstRowIdToPartitionMap` is used by data evolution MERGE INTO writes to 
locate the original partition and row count for each data file from its 
`firstRowId`. This mapping is built from scanned `DataSplit`s and then used by 
each Spark task while writing evolved rows.
   
   For merge plans that affect many files across many partitions, the map can 
become large. Capturing it directly in every task closure duplicates the 
serialized map for each task and can make Spark task serialization very large.
   
   ### Solution
   
   Broadcast `firstRowIdToPartitionMap` once through Spark instead of embedding 
the whole map in every task closure.
   
   The map stores partition `BinaryRow`s as serialized bytes. `BinaryRow` keeps 
its memory segments in transient fields and relies on Java serialization hooks 
to restore them, while Spark Kryo may serialize the object fields directly. 
Serializing the partition with `SerializationUtils.serializeBinaryRow` keeps 
the broadcast value Kryo-safe, and the writer deserializes the bytes before use.
   


-- 
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