Hello All, 

Recent Samza Runner tests failure in python/xlang [1][2] reveals an interesting 
fact that Reshuffle Transform in classic pipeline requires the input to be KV 
while portable pipeline does not, where Reshuffle in portable mode it has an 
extra step to append a random key [3].

This suggests that Reshuffle in classic mode is, sort of, equivalent to 
ReshufflePerKey in potable mode instead of Reshuffle itself. Couple of 
questions on this:

1. Is such SDK/API discrepancy expected?
2. If Yes, then, what are the advised approach for runners to implement 
translators for such transforms?
3. If No, is this something we can improve?

Best,
Ke


[1] https://ci-beam.apache.org/job/beam_PostCommit_Python_VR_Samza/288/ 
<https://ci-beam.apache.org/job/beam_PostCommit_Python_VR_Samza/288/> 
[2] https://ci-beam.apache.org/job/beam_PostCommit_XVR_Samza/285/ 
<https://ci-beam.apache.org/job/beam_PostCommit_XVR_Samza/285/> 
[3] 
https://github.com/apache/beam/blob/master/sdks/python/apache_beam/transforms/util.py#L730
 
<https://github.com/apache/beam/blob/master/sdks/python/apache_beam/transforms/util.py#L730>
 

Reply via email to