Anders Rydbirk created SPARK-36553:
--------------------------------------

             Summary: KMeans fails with NegativeArraySizeException for K = 
50000 after issue #27758 was introduced
                 Key: SPARK-36553
                 URL: https://issues.apache.org/jira/browse/SPARK-36553
             Project: Spark
          Issue Type: Bug
          Components: ML, MLlib, PySpark
    Affects Versions: 3.1.1
            Reporter: Anders Rydbirk


We are running KMeans on approximately 350M rows of x, y, z coordinates using 
the following configuration:
{code:java}
KMeans(
  featuresCol='features',
  predictionCol='centroid_id',
  k=50000,
  initMode='k-means||',
  initSteps=2,
  tol=0.00005,
  maxIter=20,
  seed=SEED,
  distanceMeasure='euclidean'
)
{code}
When using Spark 3.0.0 this worked fine, but  when upgrading to 3.1.1 we are 
consistently getting errors unless we reduce K.

Stacktrace:

 
{code:java}
An error occurred while calling o167.fit.An error occurred while calling 
o167.fit.: java.lang.NegativeArraySizeException: -897458648 at 
scala.reflect.ManifestFactory$DoubleManifest.newArray(Manifest.scala:194) at 
scala.reflect.ManifestFactory$DoubleManifest.newArray(Manifest.scala:191) at 
scala.Array$.ofDim(Array.scala:221) at 
org.apache.spark.mllib.clustering.DistanceMeasure.computeStatistics(DistanceMeasure.scala:52)
 at 
org.apache.spark.mllib.clustering.KMeans.runAlgorithmWithWeight(KMeans.scala:280)
 at org.apache.spark.mllib.clustering.KMeans.runWithWeight(KMeans.scala:231) at 
org.apache.spark.ml.clustering.KMeans.$anonfun$fit$1(KMeans.scala:354) at 
org.apache.spark.ml.util.Instrumentation$.$anonfun$instrumented$1(Instrumentation.scala:191)
 at scala.util.Try$.apply(Try.scala:213) at 
org.apache.spark.ml.util.Instrumentation$.instrumented(Instrumentation.scala:191)
 at org.apache.spark.ml.clustering.KMeans.fit(KMeans.scala:329) at 
java.base/jdk.internal.reflect.NativeMethodAccessorImpl.invoke0(Native Method) 
at java.base/jdk.internal.reflect.NativeMethodAccessorImpl.invoke(Unknown 
Source) at 
java.base/jdk.internal.reflect.DelegatingMethodAccessorImpl.invoke(Unknown 
Source) at java.base/java.lang.reflect.Method.invoke(Unknown Source) at 
py4j.reflection.MethodInvoker.invoke(MethodInvoker.java:244) at 
py4j.reflection.ReflectionEngine.invoke(ReflectionEngine.java:357) at 
py4j.Gateway.invoke(Gateway.java:282) at 
py4j.commands.AbstractCommand.invokeMethod(AbstractCommand.java:132) at 
py4j.commands.CallCommand.execute(CallCommand.java:79) at 
py4j.GatewayConnection.run(GatewayConnection.java:238) at 
java.base/java.lang.Thread.run(Unknown Source)
{code}
 

The issue is introduced by 
[#27758|[https://github.com/apache/spark/pull/27758/files#diff-725d4624ddf4db9cc51721c2ddaef50a1bc30e7b471e0439da28c5b5582efdfdR52]]
 which significantly reduces the maximum value of K:

[DistanceMeasure.scala|[https://github.com/zhengruifeng/spark/blob/d31d488e0e48a82fd5b43c406f07b8c7d27dd53c/mllib/src/main/scala/org/apache/spark/mllib/clustering/DistanceMeasure.scala#L52]]
{code:java}
val packedValues = Array.ofDim[Double](k * (k + 1) / 2)
{code}
 

*What we have tried:*
 * Reducing iterations
 * Reducing input volume
 * Reducing K

Only reducing K have yielded success.

 

*What we don't understand*:

**Given the line of code above, we do not understand why we would get an 
integer overflow:

For K=50,000, packedValues should be allocated with the size of 1,250,025,000 < 
(2^31) and not result in a negative array size.


Please let me know if more information is needed, this is my first time raising 
a bug for a OS.



--
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

Reply via email to