[SPARK-25004][CORE] Add spark.executor.pyspark.memory limit.

## What changes were proposed in this pull request?

This adds `spark.executor.pyspark.memory` to configure Python's address space limit, [`resource.RLIMIT_AS`](https://docs.python.org/3/library/resource.html#resource.RLIMIT_AS). Limiting Python's address space allows Python to participate in memory management. In practice, we see fewer cases of Python taking too much memory because it doesn't know to run garbage collection. This results in YARN killing fewer containers. This also improves error messages so users know that Python is consuming too much memory:

```
  File "build/bdist.linux-x86_64/egg/package/library.py", line 265, in fe_engineer
    fe_eval_rec.update(f(src_rec_prep, mat_rec_prep))
  File "build/bdist.linux-x86_64/egg/package/library.py", line 163, in fe_comp
    comparisons = EvaluationUtils.leven_list_compare(src_rec_prep.get(item, []), mat_rec_prep.get(item, []))
  File "build/bdist.linux-x86_64/egg/package/evaluationutils.py", line 25, in leven_list_compare
    permutations = sorted(permutations, reverse=True)
  MemoryError
```

The new pyspark memory setting is used to increase requested YARN container memory, instead of sharing overhead memory between python and off-heap JVM activity.

## How was this patch tested?

Tested memory limits in our YARN cluster and verified that MemoryError is thrown.

Author: Ryan Blue <[email protected]>

Closes apache#21977 from rdblue/SPARK-25004-add-python-memory-limit.

Author: rdblue

core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala

@@ -49,9 +49,6 @@ private[spark] class PythonRDD(
     isFromBarrier: Boolean = false)
   extends RDD[Array[Byte]](parent) {
 
-  val bufferSize = conf.getInt("spark.buffer.size", 65536)
-  val reuseWorker = conf.getBoolean("spark.python.worker.reuse", true)
-
   override def getPartitions: Array[Partition] = firstParent.partitions
 
   override val partitioner: Option[Partitioner] = {
@@ -61,7 +58,7 @@ private[spark] class PythonRDD(
   val asJavaRDD: JavaRDD[Array[Byte]] = JavaRDD.fromRDD(this)
 
   override def compute(split: Partition, context: TaskContext): Iterator[Array[Byte]] = {
-    val runner = PythonRunner(func, bufferSize, reuseWorker)
+    val runner = PythonRunner(func)
     runner.compute(firstParent.iterator(split, context), split.index, context)
   }
 

core/src/main/scala/org/apache/spark/api/python/PythonRunner.scala

@@ -27,6 +27,7 @@ import scala.collection.JavaConverters._
 
 import org.apache.spark._
 import org.apache.spark.internal.Logging
+import org.apache.spark.internal.config.PYSPARK_EXECUTOR_MEMORY
 import org.apache.spark.security.SocketAuthHelper
 import org.apache.spark.util._
 
@@ -62,14 +63,20 @@ private[spark] object PythonEvalType {
  */
 private[spark] abstract class BasePythonRunner[IN, OUT](
     funcs: Seq[ChainedPythonFunctions],
-    bufferSize: Int,
-    reuseWorker: Boolean,
     evalType: Int,
     argOffsets: Array[Array[Int]])
   extends Logging {
 
   require(funcs.length == argOffsets.length, "argOffsets should have the same length as funcs")
 
+  private val conf = SparkEnv.get.conf
+  private val bufferSize = conf.getInt("spark.buffer.size", 65536)
+  private val reuseWorker = conf.getBoolean("spark.python.worker.reuse", true)
+  // each python worker gets an equal part of the allocation. the worker pool will grow to the
+  // number of concurrent tasks, which is determined by the number of cores in this executor.
+  private val memoryMb = conf.get(PYSPARK_EXECUTOR_MEMORY)
+      .map(_ / conf.getInt("spark.executor.cores", 1))
+
   // All the Python functions should have the same exec, version and envvars.
   protected val envVars = funcs.head.funcs.head.envVars
   protected val pythonExec = funcs.head.funcs.head.pythonExec
@@ -82,7 +89,7 @@ private[spark] abstract class BasePythonRunner[IN, OUT](
   private[spark] var serverSocket: Option[ServerSocket] = None
 
   // Authentication helper used when serving method calls via socket from Python side.
-  private lazy val authHelper = new SocketAuthHelper(SparkEnv.get.conf)
+  private lazy val authHelper = new SocketAuthHelper(conf)
 
   def compute(
       inputIterator: Iterator[IN],
@@ -95,6 +102,9 @@ private[spark] abstract class BasePythonRunner[IN, OUT](
     if (reuseWorker) {
       envVars.put("SPARK_REUSE_WORKER", "1")
     }
+    if (memoryMb.isDefined) {
+      envVars.put("PYSPARK_EXECUTOR_MEMORY_MB", memoryMb.get.toString)
+    }
     val worker: Socket = env.createPythonWorker(pythonExec, envVars.asScala.toMap)
     // Whether is the worker released into idle pool
     val released = new AtomicBoolean(false)
@@ -485,20 +495,17 @@ private[spark] abstract class BasePythonRunner[IN, OUT](
 
 private[spark] object PythonRunner {
 
-  def apply(func: PythonFunction, bufferSize: Int, reuseWorker: Boolean): PythonRunner = {
-    new PythonRunner(Seq(ChainedPythonFunctions(Seq(func))), bufferSize, reuseWorker)
+  def apply(func: PythonFunction): PythonRunner = {
+    new PythonRunner(Seq(ChainedPythonFunctions(Seq(func))))
   }
 }
 
 /**
  * A helper class to run Python mapPartition in Spark.
  */
-private[spark] class PythonRunner(
-    funcs: Seq[ChainedPythonFunctions],
-    bufferSize: Int,
-    reuseWorker: Boolean)
+private[spark] class PythonRunner(funcs: Seq[ChainedPythonFunctions])
   extends BasePythonRunner[Array[Byte], Array[Byte]](
-    funcs, bufferSize, reuseWorker, PythonEvalType.NON_UDF, Array(Array(0))) {
+    funcs, PythonEvalType.NON_UDF, Array(Array(0))) {
 
   protected override def newWriterThread(
       env: SparkEnv,

core/src/main/scala/org/apache/spark/internal/config/package.scala

@@ -114,6 +114,10 @@ package object config {
     .checkValue(_ >= 0, "The off-heap memory size must not be negative")
     .createWithDefault(0)
 
+  private[spark] val PYSPARK_EXECUTOR_MEMORY = ConfigBuilder("spark.executor.pyspark.memory")
+    .bytesConf(ByteUnit.MiB)
+    .createOptional
+
   private[spark] val IS_PYTHON_APP = ConfigBuilder("spark.yarn.isPython").internal()
     .booleanConf.createWithDefault(false)
 

docs/configuration.md

@@ -179,6 +179,18 @@ of the most common options to set are:
     (e.g. <code>2g</code>, <code>8g</code>).
   </td>
 </tr>
+<tr>
+ <td><code>spark.executor.pyspark.memory</code></td>
+  <td>Not set</td>
+  <td>
+    The amount of memory to be allocated to PySpark in each executor, in MiB
+    unless otherwise specified.  If set, PySpark memory for an executor will be
+    limited to this amount. If not set, Spark will not limit Python's memory use
+    and it is up to the application to avoid exceeding the overhead memory space
+    shared with other non-JVM processes. When PySpark is run in YARN, this memory
+    is added to executor resource requests.
+  </td>
+</tr>
 <tr>
  <td><code>spark.executor.memoryOverhead</code></td>
   <td>executorMemory * 0.10, with minimum of 384 </td>

python/pyspark/worker.py

@@ -22,6 +22,7 @@
 import os
 import sys
 import time
+import resource
 import socket
 import traceback
 
@@ -263,6 +264,28 @@ def main(infile, outfile):
         isBarrier = read_bool(infile)
         boundPort = read_int(infile)
         secret = UTF8Deserializer().loads(infile)
+
+        # set up memory limits
+        memory_limit_mb = int(os.environ.get('PYSPARK_EXECUTOR_MEMORY_MB', "-1"))
+        total_memory = resource.RLIMIT_AS
+        try:
+            if memory_limit_mb > 0:
+                (soft_limit, hard_limit) = resource.getrlimit(total_memory)
+                msg = "Current mem limits: {0} of max {1}\n".format(soft_limit, hard_limit)
+                print(msg, file=sys.stderr)
+
+                # convert to bytes
+                new_limit = memory_limit_mb * 1024 * 1024
+
+                if soft_limit == resource.RLIM_INFINITY or new_limit < soft_limit:
+                    msg = "Setting mem limits to {0} of max {1}\n".format(new_limit, new_limit)
+                    print(msg, file=sys.stderr)
+                    resource.setrlimit(total_memory, (new_limit, new_limit))
+
+        except (resource.error, OSError, ValueError) as e:
+            # not all systems support resource limits, so warn instead of failing
+            print("WARN: Failed to set memory limit: {0}\n".format(e), file=sys.stderr)
+
         # initialize global state
         taskContext = None
         if isBarrier:

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala

@@ -91,6 +91,13 @@ private[spark] class Client(
   private val executorMemoryOverhead = sparkConf.get(EXECUTOR_MEMORY_OVERHEAD).getOrElse(
     math.max((MEMORY_OVERHEAD_FACTOR * executorMemory).toLong, MEMORY_OVERHEAD_MIN)).toInt
 
+  private val isPython = sparkConf.get(IS_PYTHON_APP)
+  private val pysparkWorkerMemory: Int = if (isPython) {
+    sparkConf.get(PYSPARK_EXECUTOR_MEMORY).map(_.toInt).getOrElse(0)
+  } else {
+    0
+  }
+
   private val distCacheMgr = new ClientDistributedCacheManager()
 
   private val principal = sparkConf.get(PRINCIPAL).orNull
@@ -333,12 +340,12 @@ private[spark] class Client(
     val maxMem = newAppResponse.getMaximumResourceCapability().getMemory()
     logInfo("Verifying our application has not requested more than the maximum " +
       s"memory capability of the cluster ($maxMem MB per container)")
-    val executorMem = executorMemory + executorMemoryOverhead
+    val executorMem = executorMemory + executorMemoryOverhead + pysparkWorkerMemory
     if (executorMem > maxMem) {
-      throw new IllegalArgumentException(s"Required executor memory ($executorMemory" +
-        s"+$executorMemoryOverhead MB) is above the max threshold ($maxMem MB) of this cluster! " +
-        "Please check the values of 'yarn.scheduler.maximum-allocation-mb' and/or " +
-        "'yarn.nodemanager.resource.memory-mb'.")
+      throw new IllegalArgumentException(s"Required executor memory ($executorMemory), overhead " +
+        s"($executorMemoryOverhead MB), and PySpark memory ($pysparkWorkerMemory MB) is above " +
+        s"the max threshold ($maxMem MB) of this cluster! Please check the values of " +
+        s"'yarn.scheduler.maximum-allocation-mb' and/or 'yarn.nodemanager.resource.memory-mb'.")
     }
     val amMem = amMemory + amMemoryOverhead
     if (amMem > maxMem) {

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocator.scala

@@ -133,10 +133,17 @@ private[yarn] class YarnAllocator(
   // Additional memory overhead.
   protected val memoryOverhead: Int = sparkConf.get(EXECUTOR_MEMORY_OVERHEAD).getOrElse(
     math.max((MEMORY_OVERHEAD_FACTOR * executorMemory).toInt, MEMORY_OVERHEAD_MIN)).toInt
+  protected val pysparkWorkerMemory: Int = if (sparkConf.get(IS_PYTHON_APP)) {
+    sparkConf.get(PYSPARK_EXECUTOR_MEMORY).map(_.toInt).getOrElse(0)
+  } else {
+    0
+  }
   // Number of cores per executor.
   protected val executorCores = sparkConf.get(EXECUTOR_CORES)
   // Resource capability requested for each executors
-  private[yarn] val resource = Resource.newInstance(executorMemory + memoryOverhead, executorCores)
+  private[yarn] val resource = Resource.newInstance(
+    executorMemory + memoryOverhead + pysparkWorkerMemory,
+    executorCores)
 
   private val launcherPool = ThreadUtils.newDaemonCachedThreadPool(
     "ContainerLauncher", sparkConf.get(CONTAINER_LAUNCH_MAX_THREADS))

resource-managers/yarn/src/test/scala/org/apache/spark/deploy/yarn/BaseYarnClusterSuite.scala

@@ -133,7 +133,8 @@ abstract class BaseYarnClusterSuite
       extraClassPath: Seq[String] = Nil,
       extraJars: Seq[String] = Nil,
       extraConf: Map[String, String] = Map(),
-      extraEnv: Map[String, String] = Map()): SparkAppHandle.State = {
+      extraEnv: Map[String, String] = Map(),
+      outFile: Option[File] = None): SparkAppHandle.State = {
     val deployMode = if (clientMode) "client" else "cluster"
     val propsFile = createConfFile(extraClassPath = extraClassPath, extraConf = extraConf)
     val env = Map("YARN_CONF_DIR" -> hadoopConfDir.getAbsolutePath()) ++ extraEnv
@@ -161,6 +162,11 @@ abstract class BaseYarnClusterSuite
     }
     extraJars.foreach(launcher.addJar)
 
+    if (outFile.isDefined) {
+      launcher.redirectOutput(outFile.get)
+      launcher.redirectError()
+    }
+
     val handle = launcher.startApplication()
     try {
       eventually(timeout(2 minutes), interval(1 second)) {
@@ -179,17 +185,22 @@ abstract class BaseYarnClusterSuite
    * the tests enforce that something is written to a file after everything is ok to indicate
    * that the job succeeded.
    */
-  protected def checkResult(finalState: SparkAppHandle.State, result: File): Unit = {
-    checkResult(finalState, result, "success")
-  }
-
   protected def checkResult(
       finalState: SparkAppHandle.State,
       result: File,
-      expected: String): Unit = {
-    finalState should be (SparkAppHandle.State.FINISHED)
+      expected: String = "success",
+      outFile: Option[File] = None): Unit = {
+    // the context message is passed to assert as Any instead of a function. to lazily load the
+    // output from the file, this passes an anonymous object that loads it in toString when building
+    // an error message
+    val output = new Object() {
+      override def toString: String = outFile
+          .map(Files.toString(_, StandardCharsets.UTF_8))
+          .getOrElse("(stdout/stderr was not captured)")
+    }
+    assert(finalState === SparkAppHandle.State.FINISHED, output)
     val resultString = Files.toString(result, StandardCharsets.UTF_8)
-    resultString should be (expected)
+    assert(resultString === expected, output)
   }
 
   protected def mainClassName(klass: Class[_]): String = {

resource-managers/yarn/src/test/scala/org/apache/spark/deploy/yarn/YarnClusterSuite.scala

@@ -282,13 +282,15 @@ class YarnClusterSuite extends BaseYarnClusterSuite {
     val mod2Archive = TestUtils.createJarWithFiles(Map("mod2.py" -> TEST_PYMODULE), moduleDir)
     val pyFiles = Seq(pyModule.getAbsolutePath(), mod2Archive.getPath()).mkString(",")
     val result = File.createTempFile("result", null, tempDir)
+    val outFile = Some(File.createTempFile("stdout", null, tempDir))
 
     val finalState = runSpark(clientMode, primaryPyFile.getAbsolutePath(),
       sparkArgs = Seq("--py-files" -> pyFiles),
       appArgs = Seq(result.getAbsolutePath()),
       extraEnv = extraEnvVars,
-      extraConf = extraConf)
-    checkResult(finalState, result)
+      extraConf = extraConf,
+      outFile = outFile)
+    checkResult(finalState, result, outFile = outFile)
   }
 
   private def testUseClassPathFirst(clientMode: Boolean): Unit = {

sql/core/src/main/scala/org/apache/spark/sql/execution/python/AggregateInPandasExec.scala

@@ -79,8 +79,6 @@ case class AggregateInPandasExec(
   override protected def doExecute(): RDD[InternalRow] = {
     val inputRDD = child.execute()
 
-    val bufferSize = inputRDD.conf.getInt("spark.buffer.size", 65536)
-    val reuseWorker = inputRDD.conf.getBoolean("spark.python.worker.reuse", defaultValue = true)
     val sessionLocalTimeZone = conf.sessionLocalTimeZone
     val pythonRunnerConf = ArrowUtils.getPythonRunnerConfMap(conf)
 
@@ -137,8 +135,6 @@ case class AggregateInPandasExec(
 
       val columnarBatchIter = new ArrowPythonRunner(
         pyFuncs,
-        bufferSize,
-        reuseWorker,
         PythonEvalType.SQL_GROUPED_AGG_PANDAS_UDF,
         argOffsets,
         aggInputSchema,