andygrove commented on PR #5724:
URL: 
https://github.com/apache/datafusion-comet/pull/5724#issuecomment-5779115136

   I caught up on everything since my last pass. Thanks for working through the 
earlier feedback. The Parquet 59 links, the zero-FPP guard, the unused 
max-bytes relaxation, the shared sizing table with the binary-search coverage, 
the explicit JVM wrapper and the overestimated-NDV test all look good. I've 
replied inline on the three things I think are still open: the name resolution 
on older runtimes, the folding opt-in and its docs, and the proto shape.
   
   A few other things:
   
   **Memory.** The native writers don't charge their buffers to any memory 
pool, and Bloom filters add a large up-front allocation per enabled column per 
open file. With fanout that multiplies by the number of open partitions, and 
folding doesn't return the memory because the vector keeps its capacity. That 
problem predates this PR and applies to both native writers, so I filed #6115 
to charge write buffers to the shared off-heap pool. It doesn't need to be 
solved here.
   
   **Benchmarks.** On sunchao's benchmark request, I think a native write with 
filters on vs off at the default cap is enough, reporting write time and peak 
RSS. The pinned iceberg-rust scan doesn't load Bloom filters, so native-reader 
pruning numbers don't really apply to this PR.
   
   **Rebase and CI.** Please rebase. Right now the only conflict is in 
`CometIcebergWriteActionSuite.scala`. After that we need CI to run the Spark 
3.4, 3.5 and 4.0 profiles, since no CI has run on any revision of this branch 
yet.
   


-- 
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]

Reply via email to