tustvold commented on issue #4764:
URL: https://github.com/apache/arrow-rs/issues/4764#issuecomment-1704391735

   > feature gating the IPC readers 
   
   The IPC readers are necessary to read the embedded arrow schema data, I'm 
also surprised if they are a meaningful bottleneck in compile times.
   
   > and take kernels out
   
   The parquet crate doesn't make use of the take kernel, it does make use of 
the cast kernels though which indirectly use the cast kernel. This would be 
non-trivial to remove, especially for decimals
   
   > This should not hurt your test times as I will ensure the default features 
remain the same
   
   Yes until someone uses disabled default features :sweat_smile: I dunno, I 
have found features to not really justify their downsides.
   
   > But I don't think it's needed here as I don't plan to mess with the 
defaults.
   
   We still need to ensure compilation with non-default sets of feature flags? 
We currently already have a fairly long list of combinations in CI...
   
   > The compile times in polars are already way longer than we want. That's 
why I'd like to cherry pick on functionality
   
   Perhaps we might do this empirically? If there are kernels that represent an 
outsize impact on compile times we then work out mitigation strategies for 
them? I strongly suspect that there are couple of particularly problematic 
sub-kernels (likely dictionaries) that have an outsize impact, and which it 
might be able to address this without having to resort to feature flags? FWIW I 
have some ideas on how to make the take kernel less expensive.
   


-- 
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