[SPARK-38388] Repartition + Stage retries could lead to incorrect data

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.4.0, 3.1.1
• Component/s: SQL
• Labels: correctness data-loss

Environment

Spark 2.4 and 3.x

• Assignee: Unassigned
• Reporter: Jason Xu
• Created: 02/Mar/22 08:32
• Updated: 27/Dec/23 03:25

Description

Spark repartition uses RoundRobinPartitioning, the generated results is non-deterministic when data has some randomness and stage/task retries happen.

The bug can be triggered when upstream data has some randomness, a repartition is called on them, then followed by result stage (could be more stages).
As the pattern shows below:
upstream stage (data with randomness) -> (repartition shuffle) -> result stage

When one executor goes down at result stage, some tasks of that stage might have finished, others would fail, shuffle files on that executor also get lost, some tasks from previous stage (upstream data generation, repartition) will need to rerun to generate dependent shuffle data files.
Because data has some randomness, regenerated data in upstream retried tasks is slightly different, repartition then generates inconsistent ordering, then tasks at result stage will be retried generating different data.

This is similar but different to https://issues.apache.org/jira/browse/SPARK-23207, fix for it uses extra local sort to make the row ordering deterministic, the sorting algorithm it uses simply compares row/record hash. But in this case, upstream data has some randomness, the sorting algorithm doesn't help keep the order, thus RoundRobinPartitioning introduced non-deterministic result.

The following code returns 986415, instead of 1000000:

import scala.sys.process._
import org.apache.spark.TaskContext

case class TestObject(id: Long, value: Double)

val ds = spark.range(0, 1000 * 1000, 1).repartition(100, $"id").withColumn("val", rand()).repartition(100).map { 
  row => if (TaskContext.get.stageAttemptNumber == 0 && TaskContext.get.attemptNumber == 0 && TaskContext.get.partitionId > 97) {
    throw new Exception("pkill -f java".!!)
  }
  TestObject(row.getLong(0), row.getDouble(1))
}

ds.toDF("id", "value").write.mode("overwrite").saveAsTable("tmp.test_table")

spark.sql("select count(distinct id) from tmp.test_table").show

Command: 

spark-shell --num-executors 10 (--conf spark.dynamicAllocation.enabled=false --conf spark.shuffle.service.enabled=false)

To simulate the issue, disable external shuffle service is needed (if it's also enabled by default in your environment),  this is to trigger shuffle file loss and previous stage retries.
In our production, we have external shuffle service enabled, this data correctness issue happened when there were node losses.

Although there's some non-deterministic factor in upstream data, user wouldn't expect  to see incorrect result.