[jira] [Comment Edited] (SPARK-38388) Repartition + Stage retries could lead to incorrect data

["Tom Sisso (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-38388?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17556051#comment-17556051 ] 

Tom Sisso edited comment on SPARK-38388 at 6/19/22 10:18 AM:
-------------------------------------------------------------

Hi
We want to share that we experience a variant of this issue as well (at Taboola, using Spark 3.1.3), and stress that this issue can occur in any scenario where the recalculation won't produce exactly the same rows, not only randomness related or the ones mentioned here.
In our case we suffered from it in multiple jobs that perform simple aggregation over some double/float type column and then repartition it to control the number of output files, for example:
{code:java}
sqlContext.sql(
" SELECT integerColumn, SUM(someDoubleTypeValue) AS value
FROM data
GROUP BY integerColumn "
).repartition(3)
.write()
…{code}

In cases where we perform such calculation, we might get slightly different value on a retry due to precision issues.
Similar to what was already explained here, such case will again lead to different hashcode for this row, which will later cause multiple rows to be mapped to different partitions, which results in incorrect data that contains duplicate & missing rows.

It can be reproduced in a similar manner - add failure after some parts already succeeded (& disable external shuffle service, speculation and dynamic allocation):
{code:java}
sqlContext.sql(
" SELECT integerColumn, SUM(someDoubleTypeValue) AS value
FROM data
GROUP BY integerColumn "
).repartition(3)
.map((MapFunction<Row, Row>) row -> {
if (TaskContext.get().stageAttemptNumber() == 0 &&
TaskContext.get().attemptNumber() == 0 &&
TaskContext.get().partitionId() > 1)
{ Thread.sleep(60_000); System.exit(1); }
return row;
}, RowEncoder.apply(schema))
.write()
…
{code}
 

We will appreciate if it can get higher priority and think that it should be communicated to users until a proper solution will be implemented, maybe with a warning on the repartition(numPartitions) api, we stopped using it.
Thanks


was (Author: JIRAUSER291139):
Hi
We want to share that we experience a variant of this issue as well (at Taboola, using Spark 3.1.3), and stress that this issue can occur in any scenario where the recalculation won't produce exactly the same rows, not only randomness related or the ones mentioned here.
In our case we suffered from it in multiple jobs that perform simple aggregation over some double/float type column and then repartition it to control the number of output files, for example:
sqlContext.sql(
" SELECT integerColumn, SUM(someDoubleTypeValue) AS value
  FROM data
  GROUP BY integerColumn "
).repartition(3)
.write()
…
In cases where we perform such calculation, we might get slightly different value on a retry due to precision issues.
Similar to what was already explained here, such case will again lead to different hashcode for this row, which will later cause multiple rows to be mapped to different partitions, which results in incorrect data that contains duplicate & missing rows.

It can be reproduced in a similar manner - add failure after some parts already succeeded (& disable external shuffle service, speculation and dynamic allocation):
sqlContext.sql(
" SELECT integerColumn, SUM(someDoubleTypeValue) AS value
  FROM data
  GROUP BY integerColumn "
).repartition(3)
.map((MapFunction<Row, Row>) row -> \{
                    if (TaskContext.get().stageAttemptNumber() == 0 &&
                            TaskContext.get().attemptNumber() == 0 &&
                            TaskContext.get().partitionId() > 1) {
                        Thread.sleep(60_000);
                        System.exit(1);
                    }

                    return row;
                }, RowEncoder.apply(schema))
.write()
…

We will appreciate if it can get higher priority and think that it should be communicated to users until a proper solution will be implemented, maybe with a warning on the repartition(numPartitions) api, we stopped using it.
Thanks

> Repartition + Stage retries could lead to incorrect data 
> ---------------------------------------------------------
>
>                 Key: SPARK-38388
>                 URL: https://issues.apache.org/jira/browse/SPARK-38388
>             Project: Spark
>          Issue Type: Bug
>          Components: SQL
>    Affects Versions: 2.4.0, 3.1.1
>         Environment: Spark 2.4 and 3.1
>            Reporter: Jason Xu
>            Priority: Major
>              Labels: correctness, data-loss
>
> 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:
> {code:java}
> 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{code}
> Command: 
> {code:java}
> spark-shell --num-executors 10 (--conf spark.dynamicAllocation.enabled=false --conf spark.shuffle.service.enabled=false){code}
> 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.



--
This message was sent by Atlassian Jira
(v8.20.7#820007)

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