Details
Major Description
In the following, the `indexer` UDF defined inside the `org.apache.spark.ml.feature.StringIndexerModel.transform` method throws an "Unseen label" error, despite the label not being present in the transformed DataFrame.
Here is the definition of the indexer UDF in the transform method:
val indexer = udf { label: String =>
if (labelToIndex.contains(label)) {
labelToIndex(label)
} else {
throw new SparkException(s"Unseen label: $label.")
}
}
We can demonstrate the error with a very simple example DataFrame.
scala> import org.apache.spark.ml.feature.StringIndexer import org.apache.spark.ml.feature.StringIndexer scala> // first we create a DataFrame with three cities scala> val df = List( | ("A", "London", "StrA"), | ("B", "Bristol", null), | ("C", "New York", "StrC") | ).toDF("ID", "CITY", "CONTENT") df: org.apache.spark.sql.DataFrame = [ID: string, CITY: string ... 1 more field] scala> df.show +---+--------+-------+ | ID| CITY|CONTENT| +---+--------+-------+ | A| London| StrA| | B| Bristol| null| | C|New York| StrC| +---+--------+-------+ scala> // then we remove the row with null in CONTENT column, which removes Bristol scala> val dfNoBristol = finalStatic.filter($"CONTENT".isNotNull) dfNoBristol: org.apache.spark.sql.Dataset[org.apache.spark.sql.Row] = [ID: string, CITY: string ... 1 more field] scala> dfNoBristol.show +---+--------+-------+ | ID| CITY|CONTENT| +---+--------+-------+ | A| London| StrA| | C|New York| StrC| +---+--------+-------+ scala> // now create a StringIndexer for the CITY column and fit to dfNoBristol scala> val model = { | new StringIndexer() | .setInputCol("CITY") | .setOutputCol("CITYIndexed") | .fit(dfNoBristol) | } model: org.apache.spark.ml.feature.StringIndexerModel = strIdx_f5afa23333fb scala> // the StringIndexerModel has only two labels: "London" and "New York" scala> str.labels foreach println London New York scala> // transform our DataFrame to add an index column scala> val dfWithIndex = model.transform(dfNoBristol) dfWithIndex: org.apache.spark.sql.DataFrame = [ID: string, CITY: string ... 2 more fields] scala> dfWithIndex.show +---+--------+-------+-----------+ | ID| CITY|CONTENT|CITYIndexed| +---+--------+-------+-----------+ | A| London| StrA| 0.0| | C|New York| StrC| 1.0| +---+--------+-------+-----------+
The unexpected behaviour comes when we filter `dfWithIndex` for `CITYIndexed` equal to 1.0 and perform an action. The `indexer` UDF in `transform` method throws an exception reporting unseen label "Bristol". This is irrational behaviour as far as the user of the API is concerned, because there is no such value as "Bristol" when do show all rows of `dfWithIndex`:
scala> dfWithIndex.filter($"CITYIndexed" === 1.0).count 17/11/04 00:33:41 ERROR Executor: Exception in task 1.0 in stage 20.0 (TID 40) org.apache.spark.SparkException: Failed to execute user defined function($anonfun$5: (string) => double) at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIterator.agg_doAggregateWithoutKey$(Unknown Source) at org.apache.spark.sql.catalyst.expressions.GeneratedClass$GeneratedIterator.processNext(Unknown Source) at org.apache.spark.sql.execution.BufferedRowIterator.hasNext(BufferedRowIterator.java:43) at org.apache.spark.sql.execution.WholeStageCodegenExec$$anonfun$8$$anon$1.hasNext(WholeStageCodegenExec.scala:395) at scala.collection.Iterator$$anon$11.hasNext(Iterator.scala:408) at org.apache.spark.shuffle.sort.BypassMergeSortShuffleWriter.write(BypassMergeSortShuffleWriter.java:125) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:96) at org.apache.spark.scheduler.ShuffleMapTask.runTask(ShuffleMapTask.scala:53) at org.apache.spark.scheduler.Task.run(Task.scala:108) at org.apache.spark.executor.Executor$TaskRunner.run(Executor.scala:335) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) at java.lang.Thread.run(Thread.java:745) Caused by: org.apache.spark.SparkException: Unseen label: Bristol. To handle unseen labels, set Param handleInvalid to keep. at org.apache.spark.ml.feature.StringIndexerModel$$anonfun$5.apply(StringIndexer.scala:222) at org.apache.spark.ml.feature.StringIndexerModel$$anonfun$5.apply(StringIndexer.scala:208) ... 13 more
To understand what is happening here, note that an action is triggered when we call `StringIndexer.fit()`, before the `CITYIndexed === 1` filter is applied, so the StringIndexerModel sees only London and New York, as expected. Now compare the query plans for `dfWithIndex` and `dfWithIndex.filter($"CITYIndexed" === 1.0)`:
scala> dfWithIndex.explain == Physical Plan == *Project [_1#3 AS ID#7, _2#4 AS CITY#8, _3#5 AS CONTENT#9, UDF(_2#4) AS CITYIndexed#159] +- *Filter isnotnull(_3#5) +- LocalTableScan [_1#3, _2#4, _3#5] scala> dfWithIndex.filter($"CITYIndexed" === 1.0).explain == Physical Plan == *Project [_1#3 AS ID#7, _2#4 AS CITY#8, _3#5 AS CONTENT#9, UDF(_2#4) AS CITYIndexed#159] +- *Filter (isnotnull(_3#5) && (UDF(_2#4) = 1.0)) +- LocalTableScan [_1#3, _2#4, _3#5]
Note that in the latter, the query plan has pushed the filter `$"CITYIndexed" === 1.0` back to be performed at the same stage as our null filter (`Filter (isnotnull(_3#5) && (UDF(_2#4) = 1.0))`).
With a debugger I have seen that both operands of `&&` are executed on each row of `df`: `isnotnull(_3#5)` and `UDF(_2#4) = 1.0`. Therefore, the UDF is passed the label `Bristol` despite isnotnull returning false for that row.
If we cache the DataFrame `dfNoBristol` immediately after creating it, then there is no longer an error because the optimizer does not attempt to call the UDF on unseen data. The fact that we get different results depending on whether or not we call cache is a cause for concern.
I have seen similar issues with pure SparkSql DataFrame operations when the DAG gets complicated (many joins, and aggregations). These are harder to isolate to such a simple example, but I plan to report them in the near future.
