steveniemitz commented on PR #17783:
URL: https://github.com/apache/beam/pull/17783#issuecomment-1143881332

   > The other part of the change makes sense to reduce byte[] copies by using 
ByteString.
   > 
   > CC: @tudorm
   
   Maybe I'll pull the ByteString refactoring stuff out into another review 
just to make this easier?  Do you have any particular issues with it using 
ByteString there?
   
   The downsides with using Output/Input stream are really too big to ignore 
here, the performance differences are orders of magnitude in our tests.  The 
main problem is that most "stream" compressor implementations are designed to 
compress a large amount of data, but in this case we're usually only 
compressing a few 100-1KB.  It makes the overhead from creating/destroying the 
compressor streams very high (comparatively at least).  We ran into this 
problem both with deflate and zstd, and its one of the reasons we ended up with 
an interface like this.  If its really a non-starter putting this on OSS with a 
similar interface that's fine though, we can continue maintaining this in our 
own fork for the time being. 
   
   The PipelineVisitor idea is interesting, although I'm skeptical how well 
it'd work in practice.  For example with a Combine the coder for the data being 
shuffled is the accumulator coder, not the value coder of the KV.  I bet you'd 
need a bunch of special cases to pick the "right" coder to wrap for various 
transforms.


-- 
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