[SPARK-27892] Saving/loading stages in PipelineModel should be parallel

• Type: Improvement
• Status: In Progress
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 3.1.0
• Component/s: ML
• Labels: easyfix performance
• Assignee: Unassigned
• Reporter: Jason Wang
• Created: 31/May/19 00:19
• Updated: 10/Jul/20 19:15

Description

When a PipelineModel is saved/loaded, all the stages are saved/loaded sequentially. When dealing with a PipelineModel with many stages, although each stage's save/load takes sub-second, the total time taken for the PipelineModel could be several minutes. It should be trivial to parallelize the save/load of stages in the SharedReadWrite object.

 

To reproduce:

import org.apache.spark.ml._
import org.apache.spark.ml.feature.VectorAssembler
val outputPath = "..."
val stages = (1 to 100) map { i => new VectorAssembler().setInputCols(Array("input")).setOutputCol("o" + i)}
val p = new Pipeline().setStages(stages.toArray)
val data = Seq(1, 1, 1) toDF "input"
val pm = p.fit(data)
pm.save(outputPath)

Issue Links

• links to: [Github] Pull Request #29068 (Moovlin)
• links to: [Github] Pull Request #29068 (Moovlin)