Hyukjin Kwon created SPARK-61879:
------------------------------------

             Summary: Take and send the arguments of the native ML members as 
PySpark does: relation methods, qualnames and bindings, param-map order, the 
from_* class methods and the computeCost warning
                 Key: SPARK-61879
                 URL: https://issues.apache.org/jira/browse/SPARK-61879
             Project: Spark
          Issue Type: Bug
          Components: Connect
    Affects Versions: connect-rust-4.4.0
            Reporter: Hyukjin Kwon


The native ML members of the drop-in take and send their arguments otherwise 
than PySpark 4.3.0's:
- LDAModel.describeTopics, the GLR summary's residuals and 
Word2VecModel.getVectors bind a fixed signature, where PySpark's 
try_remote_attribute_relation sends only the positional arguments: 
describeTopics(maxTermsPerTopic=1) gives 1 term where PySpark's keyword is 
dropped and it gives 3, and residuals(residualsType='pearson') gives the 
pearson residuals where PySpark gives the deviance ones.
- The members' __qualname__ is the public class's, not that of the class that 
defines them in PySpark (JavaMLWritable.write, JavaPredictionModel.predict), 
and their wrappers bind calls with other messages.
- The param maps of the objects that the drop-in makes natively, such as the 
defaults of a CrossValidatorModel, are in another order than PySpark's.
- CountVectorizerModel.from_vocabulary and StringIndexerModel.from_labels / 
from_arrays_of_labels extract their arguments as Rust types first, so labels 
given as a str, a set, an iterator or dict keys raise TypeError, and the errors 
come in another order than PySpark's.
- BisectingKMeansModel.computeCost's FutureWarning is attributed to the 
drop-in's wrapper, not to pyspark/ml/clustering.py, so a filter by PySpark's 
module does not match.




--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to