cshuo commented on PR #18958: URL: https://github.com/apache/hudi/pull/18958#issuecomment-4725207752
> What does not carry over is the planner rewrite. Spark gets batched/coalesced I/O by injecting ReadBlobRule through SparkSessionExtensions (HoodieAnalysis.scala:192); Flink has no equivalent resolution/planner-rule hook, so a per-row UDF would resolve each reference independently and lose that batching. Recovering it would mean manual buffering in the function or going through Flink's Module SPI for a built-in, which is heavier. @wombatu-kun Yes, that’s the main concern. If we use a Flink UDF to materialize OOL BLOB fields, we won’t have a safe way to leverage the same batched read optimization that Spark uses. Flink UDF is evaluated row by row and does not participate in operator lifecycle hooks, so it cannot reliably buffer records and flush them before a checkpoint, e.g. through `prepareSnapshotPreBarrier`. This could have a significant performance impact. -- 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]
