[BEAM-12701] Converting two deferred dataframes to csv in the same pipeline causes a PCollection label collision

• Type: Bug
• Status: Resolved
• Priority: P2
• Resolution: Fixed
• Affects Version/s: 2.31.0, 2.32.0, 2.33.0
• Fix Version/s: 2.34.0
• Component/s: dsl-dataframe
• Assignee: Eduardo Sánchez López
• Reporter: Jérémie Bigras-Dunberry
• Created: 31/Jul/21 20:07
• Updated: 16/Sep/21 17:29
• Resolved: 16/Sep/21 17:28

Description

 

If you use  the to_csv of the DeferredDataFrame twice in a single pipeline like this : 

df1 = pd.DataFrame.from_records({"a":"b"}, index=[0])
df2 = pd.DataFrame.from_records({"a":"b"}, index=[0])

with beam.Pipeline() as p:
 df1 = to_dataframe(to_pcollection(df1, pipeline=p), label="df1")
 df2 = to_dataframe(to_pcollection(df2, pipeline=p), label="df2")

 df1.to_csv("test.csv")
 df2.to_csv("test2.csv")

You get this error on the second to_csv call

 

RuntimeError: A transform with label "ToPCollection(df)" already exists in the pipeline. To apply a transform with a specified label write pvalue | "label" >> transform

I think it comes from the fact that to_csv  is calling a  to_pcollection without any label, causing to infer an identical label for both to_csv function calls. 

 

 

Issue Links

• links to: GitHub Pull Request #15450