[
https://issues.apache.org/jira/browse/SPARK-22629?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16284166#comment-16284166
]
Marco Gaido commented on SPARK-22629:
-------------------------------------
This problem is related to SPARK-20586. I created a PR to add the same
interface in Python. When it will be available, you should then mark your UDF
as non-deterministic to avoid this weird behavior. You can find further details
in the description of the PR I am submitting and on the PR and JIRA I
previously referenced to.
Thanks.
> incorrect handling of calls to random in UDFs
> ---------------------------------------------
>
> Key: SPARK-22629
> URL: https://issues.apache.org/jira/browse/SPARK-22629
> Project: Spark
> Issue Type: Bug
> Components: PySpark
> Affects Versions: 2.1.0
> Reporter: Michael H
>
> {code:none}
> df_br = spark.createDataFrame([{'name': 'hello'}])
> # udf creates a random integer
> udf_random_col = udf(lambda: int(100*random.random()), IntegerType())
> # add a column to our DF using that udf
> df_br = df_br.withColumn('RAND', udf_random_col())
> df_br.show()
> +-----+----+
> | name|RAND|
> +-----+----+
> |hello| 68|
> +-----+----+
> # udf that adds 10 to an input column value
> random.seed(1234)
> udf_add_ten = udf(lambda rand: rand + 10, IntegerType())
> # unexpected result due to re-evaluation
> df_br.withColumn('RAND_PLUS_TEN', udf_add_ten('RAND')).show()
> +-----+----+-------------+
> | name|RAND|RAND_PLUS_TEN|
> +-----+----+-------------+
> |hello| 72| 87|
> +-----+----+-------------+
> # workaround: cache the resulst after using the random number generating udf
> df_br.withColumn('RAND',
> udf_random_col()).cache().withColumn('RAND_PLUS_TEN',
> udf_add_ten('RAND')).show()
> +-----+----+-------------+
> | name|RAND|RAND_PLUS_TEN|
> +-----+----+-------------+
> |hello| 68| 78|
> +-----+----+-------------+
> {code}
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
