BInwei Yang

To bypass the result mismatch issue of https://github.com/facebookincubator/velox/issues/9219, you may set a very large MaxSpillRunRows and maxSpillFileSize

@mbasmanova @aditi-pandit Do you have more comments? The PR is to fix the issue of TPCDS Q67.

@mbasmanova Any more comments? The function is used by Gluten columnar shuffle.

> @mbasmanova @pedroerp I created a benchmark in [9265b97](https://github.com/facebookincubator/velox/commit/9265b975152586ebdda5576314133a3c24120f53) > > Benchmark Result at commit [569500d](https://github.com/facebookincubator/velox/commit/569500df1ffb26642c44233af75941561cd10c13) > > ``` > ============================================================================ > [...]hmarks/ExpressionBenchmarkBuilder.cpp relative time/iter iters/s > ============================================================================ > hash_ARRAY##hash...

> > I didn't go through your code. If I understand correctly, it's the comparison between an indirect branch prediction vs. direct branch prediction. > > If the prediction is...

@mbasmanova , @marin-ma collected below data. conclusion is to use the virtual function call. Virtual function call has 15% better performance. The root cause is it has 20% less instructions....

@marin-ma Let's use virtual function way for now. Can you update the PR?

>       virtual/switch > Instruction per loop 14,714 18,584 0.792 > IPC 2.54 2.78   > loop per second 533,513 462,064 1.155 > branch misprediction ratio 2.8% 2.0% > branch misprediction/1K inst...

@yma11 When the call client_->HeadObject or client_->GetObject fail, can we get the the retry number from outcome? If so, let's print the number in error message. So user can know...

Thank you! It's what we need. @yma11 can add more tests result from parquet-mr later.