[
https://issues.apache.org/jira/browse/SPARK-29823?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dong Wang updated SPARK-29823:
------------------------------
Description:
In mllib.clustering.KMeans.run(), the rdd {color:#de350b}_norms_{color} is
persisted. But {color:#de350b}_norms_ {color}only has a single child, i.e., the
rdd {color:#de350b}_zippedData_ {color:#172b4d}which was not
persisted{color}{color}{color:#172b4d}.{color} So all the actions that reply on
_norms_ also reply on _{color:#de350b}zippedData{color}._ The rdd
{color:#de350b}_zippedData_{color} will be used by multiple times in
_runAlgorithm()._ Therefore _{color:#de350b}zippedData{color}_ should be
persisted, not _{color:#de350b}norms{color}_.
{code:scala}
private[spark] def run(
data: RDD[Vector],
instr: Option[Instrumentation]): KMeansModel = {
if (data.getStorageLevel == StorageLevel.NONE) {
logWarning("The input data is not directly cached, which may hurt
performance if its"
+ " parent RDDs are also uncached.")
}
// Compute squared norms and cache them.
val norms = data.map(Vectors.norm(_, 2.0))
norms.persist() // Unnecessary persist. Only used to generate zippedData.
val zippedData = data.zip(norms).map { case (v, norm) =>
new VectorWithNorm(v, norm)
} // needs to persist
val model = runAlgorithm(zippedData, instr)
norms.unpersist() // Change to zippedData.unpersist()
{code}
This issue is reported by our tool CacheCheck, which is used to dynamically
detecting persist()/unpersist() api misuses.
was:
In mllib.clustering.KMeans.run(), the rdd {color:#de350b}_norms_{color} is
persisted. But {color:#de350b}_norms_ {color}only has a single child, i.e., the
rdd {color:#de350b}_zippedData_ {color:#172b4d}which was not
persisted{color}{color}{color:#172b4d}. S{color}o all the actions that reply on
_norms_ also reply on _{color:#de350b}zippedData{color}._ The rdd
{color:#de350b}_zippedData_{color} will be used by multiple times in
_runAlgorithm()._ Therefore _{color:#de350b}zippedData{color}_ should be
persisted, not __ _{color:#de350b}norms{color}._
{code:scala}
private[spark] def run(
data: RDD[Vector],
instr: Option[Instrumentation]): KMeansModel = {
if (data.getStorageLevel == StorageLevel.NONE) {
logWarning("The input data is not directly cached, which may hurt
performance if its"
+ " parent RDDs are also uncached.")
}
// Compute squared norms and cache them.
val norms = data.map(Vectors.norm(_, 2.0))
norms.persist() // Unnecessary persist. Only used to generate zippedData.
val zippedData = data.zip(norms).map { case (v, norm) =>
new VectorWithNorm(v, norm)
} // needs to persist
val model = runAlgorithm(zippedData, instr)
norms.unpersist() // Change to zippedData.unpersist()
{code}
This issue is reported by our tool CacheCheck, which is used to dynamically
detecting persist()/unpersist() api misuses.
> Improper persist strategy in mllib.clustering.KMeans.run()
> ----------------------------------------------------------
>
> Key: SPARK-29823
> URL: https://issues.apache.org/jira/browse/SPARK-29823
> Project: Spark
> Issue Type: Bug
> Components: MLlib
> Affects Versions: 2.4.3
> Reporter: Dong Wang
> Priority: Major
>
> In mllib.clustering.KMeans.run(), the rdd {color:#de350b}_norms_{color} is
> persisted. But {color:#de350b}_norms_ {color}only has a single child, i.e.,
> the rdd {color:#de350b}_zippedData_ {color:#172b4d}which was not
> persisted{color}{color}{color:#172b4d}.{color} So all the actions that reply
> on _norms_ also reply on _{color:#de350b}zippedData{color}._ The rdd
> {color:#de350b}_zippedData_{color} will be used by multiple times in
> _runAlgorithm()._ Therefore _{color:#de350b}zippedData{color}_ should be
> persisted, not _{color:#de350b}norms{color}_.
> {code:scala}
> private[spark] def run(
> data: RDD[Vector],
> instr: Option[Instrumentation]): KMeansModel = {
> if (data.getStorageLevel == StorageLevel.NONE) {
> logWarning("The input data is not directly cached, which may hurt
> performance if its"
> + " parent RDDs are also uncached.")
> }
> // Compute squared norms and cache them.
> val norms = data.map(Vectors.norm(_, 2.0))
> norms.persist() // Unnecessary persist. Only used to generate zippedData.
> val zippedData = data.zip(norms).map { case (v, norm) =>
> new VectorWithNorm(v, norm)
> } // needs to persist
> val model = runAlgorithm(zippedData, instr)
> norms.unpersist() // Change to zippedData.unpersist()
> {code}
> This issue is reported by our tool CacheCheck, which is used to dynamically
> detecting persist()/unpersist() api misuses.
--
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
