[
https://issues.apache.org/jira/browse/SPARK-35089?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17353778#comment-17353778
]
Robert Joseph Evans commented on SPARK-35089:
---------------------------------------------
On window functions if the {{order by}} clause is ambiguous you can get
different results from one run to the next. This is because the order in which
shuffle data is read in is not deterministic, even though the sorting is. In
this example you are generating start with
{code}
val getRandomStart = udf((x:Int)=>{
x+scala.util.Random.nextInt(47)
})
{code}
The input to this {x} appears to be non-abiguous (0-some very large number),
but because of the + random(0 to 47) there is the possibility of multiple start
values being the same.
So for operations where order matters you can get ambiguous results. For lead
and lag a different lead/lag value can show up, because the one right after
this one {{lead(1)}} is different. For operations like rank, dense_rank, and
row_number the order of the values output is the same, but the rows are in a
different order so the value at each row/rank is different. This can also
impact operations like SUM, MIN, and MAX that use a row bounds on a windows
instead of value ranges. I'm not sure if this should be considered a bug or
not. Spark treats all window operations as deterministic, so in theory if
there is a crash you can get inconsistent results within the same query, but
that only happens if the end user put in a non-deterministic ordering.
> non consistent results running count for same dataset after filter and lead
> window function
> -------------------------------------------------------------------------------------------
>
> Key: SPARK-35089
> URL: https://issues.apache.org/jira/browse/SPARK-35089
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 3.0.1, 3.1.1
> Reporter: Domagoj
> Priority: Major
>
> **** edit 2021-05-18
> I have make it simpler to reproduce; I've put already generated data on s3
> bucket that is publicly available with 24.000.000 records
> Now all you need to do is run this code:
> {code:java}
> import org.apache.spark.sql.expressions.Window
> import org.apache.spark.sql._
> import org.apache.spark.sql.functions._
> val w = Window.partitionBy("user").orderBy("start")
> val ts_lead = coalesce(lead("start", 1) .over(w), lit(30000000))
> spark.read.orc("s3://dtonzetic-spark-sample-data/sample-data.orc").
> withColumn("end", ts_lead).
> withColumn("duration", col("end")-col("start")).
> where("type='TypeA' and duration>4").count()
> {code}
>
> this were my results:
> - run 1: 2547559
> - run 2: 2547559
> - run 3: 2547560
> - run 4: 2547558
> - run 5: 2547558
> - run 6: 2547559
> - run 7: 2547558
> This results are from new EMR cluster, version 6.3.0, so nothing changed.
> **** end edit 2021-05-18
> I have found an inconsistency with count function results after lead window
> function and filter.
>
> I have a dataframe (this is simplified version, but it's enough to reproduce)
> with millions of records, with these columns:
> * df1:
> ** start(timestamp)
> ** user_id(int)
> ** type(string)
> I need to define duration between two rows, and filter on that duration and
> type. I used window lead function to get the next event time (that define end
> for current event), so every row now gets start and stop times. If NULL (last
> row for example), add next midnight as stop. Data is stored in ORC file
> (tried with Parquet format, no difference)
> This only happens with multiple cluster nodes, for example AWS EMR cluster or
> local docker cluster setup. If I run it on single instance (local on laptop),
> I get consistent results every time. Spark version is 3.0.1, both in AWS and
> local and docker setup.
> Here is some simple code that you can use to reproduce it, I've used
> jupyterLab notebook on AWS EMR. Spark version is 3.0.1.
>
>
> {code:java}
> import org.apache.spark.sql.expressions.Window
> // this dataframe generation code should be executed only once, and data have
> to be saved, and then opened from disk, so it's always same.
> val getRandomUser = udf(()=>{
> val users = Seq("John","Eve","Anna","Martin","Joe","Steve","Katy")
> users(scala.util.Random.nextInt(7))
> })
> val getRandomType = udf(()=>{
> val types = Seq("TypeA","TypeB","TypeC","TypeD","TypeE")
> types(scala.util.Random.nextInt(5))
> })
> val getRandomStart = udf((x:Int)=>{
> x+scala.util.Random.nextInt(47)
> })
> // for loop is used to avoid out of memory error during creation of dataframe
> for( a <- 0 to 23){
> // use iterator a to continue with next million, repeat 1 mil times
> val x=Range(a*1000000,(a*1000000)+1000000).toDF("id").
> withColumn("start",getRandomStart(col("id"))).
> withColumn("user",getRandomUser()).
> withColumn("type",getRandomType()).
> drop("id")
> x.write.mode("append").orc("hdfs:///random.orc")
> }
> // above code should be run only once, I used a cell in Jupyter
> // define window and lead
> val w = Window.partitionBy("user").orderBy("start")
> // if null, replace with 30.000.000
> val ts_lead = coalesce(lead("start", 1) .over(w), lit(30000000))
> // read data to dataframe, create stop column and calculate duration
> val fox2 = spark.read.orc("hdfs:///random.orc").
> withColumn("end", ts_lead).
> withColumn("duration", col("end")-col("start"))
> // repeated executions of this line returns different results for count
> // I have it in separate cell in JupyterLab
> fox2.where("type='TypeA' and duration>4").count()
> {code}
> My results for three consecutive runs of last line were:
> * run 1: 2551259
> * run 2: 2550756
> * run 3: 2551279
> It's very important to say that if I use filter:
> fox2.where("type='TypeA' ")
> or
> fox2.where("duration>4"),
>
> each of them can be executed repeatedly and I get consistent result every
> time.
> I can save dataframe after crating stop and duration columns, and after that,
> I get consistent results every time.
> It is not very practical workaround, as I need a lot of space and time to
> implement it.
> This dataset is really big (in my eyes at least, aprox 100.000.000 new
> records per day).
> If I run this same example on my local machine using master = local[*],
> everything works as expected, it's just on cluster setup. I tried to create
> cluster using docker on my local machine, created 3.0.1 and 3.1.1 clusters
> with one master and two workers, and have successfully reproduced issue.
>
>
>
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
