[GitHub] [spark] viirya commented on a change in pull request #25442: [SPARK-28722][ML] Change sequential label sorting in StringIndexer fit to parallel

[GitBox <gi...@apache.org>]

viirya commented on a change in pull request #25442: [SPARK-28722][ML] Change sequential label sorting in StringIndexer fit to parallel
URL: https://github.com/apache/spark/pull/25442#discussion_r313719799
 
 

 ##########
 File path: mllib/src/main/scala/org/apache/spark/ml/feature/StringIndexer.scala
 ##########
 @@ -213,32 +221,36 @@ class StringIndexer @Since("1.4.0") (
     val labelsArray = $(stringOrderType) match {
       case StringIndexer.frequencyDesc =>
         val sortFunc = StringIndexer.getSortFunc(ascending = false)
-        countByValue(dataset, inputCols).map { counts =>
+        val orgStrings = countByValue(dataset, inputCols).toSeq
+        ThreadUtils.parmap(orgStrings, "sortingStringLabels", 8) { counts =>
           counts.toSeq.sortWith(sortFunc).map(_._1).toArray
-        }
+        }.toArray
       case StringIndexer.frequencyAsc =>
         val sortFunc = StringIndexer.getSortFunc(ascending = true)
-        countByValue(dataset, inputCols).map { counts =>
+        val orgStrings = countByValue(dataset, inputCols).toSeq
+        ThreadUtils.parmap(orgStrings, "sortingStringLabels", 8) { counts =>
           counts.toSeq.sortWith(sortFunc).map(_._1).toArray
-        }
+        }.toArray
       case StringIndexer.alphabetDesc =>
-        import dataset.sparkSession.implicits._
         dataset.persist()
-        val labels = inputCols.map { inputCol =>
-          dataset.select(inputCol).na.drop().distinct().sort(dataset(s"$inputCol").desc)
-            .as[String].collect()
-        }
+        val selectedCols = getSelectedCols(dataset, inputCols).map(collect_set(_))
+        val allLabels = dataset.select(selectedCols: _*)
+          .collect().toSeq.flatMap(_.toSeq).asInstanceOf[Seq[Seq[String]]]
 
 Review comment:
   distinct is done at executors by `collect_set` expression. Yes, sorting is done at the driver.
   
   This is a good point. I think it depends on the cardinality of input labels. For StringIndexer, the cardinality should not be very high, as suggested billion level.
   
   Actually, for frequency-based string order, sorting is also done at the driver, previously.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services

---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org