mbutrovich commented on code in PR #2578:
URL: https://github.com/apache/datafusion-comet/pull/2578#discussion_r2432627173
##########
native/core/src/execution/shuffle/row.rs:
##########
@@ -839,6 +839,7 @@ pub fn process_sorted_row_partition(
.open(&output_path)?;
output_data.write_all(&frozen)?;
+ output_data.flush()?;
Review Comment:
Does a partially written file read back successfully? Otherwise I don't
think flushing each batch, resulting in an unrecoverable, partially-written
file makes sense. If it's not readable, `flush()` should be outside of the loop
after all batches have been written. If it is readable, I think `flush()` where
you added it makes sense. Also it's misleading to say that this avoids data
loss, since `flush()` does not do an `fsync()`. I'm not proposing that we add
an `fsync()`. I do not know what sort of durability guarantees Spark typically
provides for its shuffle files when writing out batches.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
