Github user galv commented on a diff in the pull request:
https://github.com/apache/spark/pull/20761#discussion_r189449512
--- Diff:
resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/ResourceTypeHelper.scala
---
@@ -0,0 +1,180 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.spark.deploy.yarn
+
+import java.lang.reflect.InvocationTargetException
+
+import scala.util.control.NonFatal
+
+import org.apache.hadoop.yarn.api.records.Resource
+
+import org.apache.spark.internal.Logging
+import org.apache.spark.util.Utils
+
+/**
+ * This helper class uses some of Hadoop 3 methods from the yarn API,
+ * so we need to use reflection to avoid compile error when building
against Hadoop 2.x
+ */
+object ResourceTypeHelper extends Logging {
+ private val AMOUNT_AND_UNIT_REGEX = "([0-9]+)([A-Za-z]*)".r
+ private val resourceTypesNotAvailableErrorMessage =
+ "Ignoring updating resource with resource types because " +
+ "the version of YARN does not support it!"
+
+ def setResourceInfoFromResourceTypes(
+ resourceTypesParam: Map[String, String],
+ resource: Resource): Resource = {
+ if (resource == null) {
+ throw new IllegalArgumentException("Resource parameter should not be
null!")
+ }
+
+ if (!ResourceTypeHelper.isYarnResourceTypesAvailable()) {
+ logWarning(resourceTypesNotAvailableErrorMessage)
+ return resource
+ }
+
+ val resourceTypes = resourceTypesParam.map { case (k, v) => (
+ if (k.equals("memory")) {
+ logWarning("Trying to use 'memory' as a custom resource, converted
it to 'memory-mb'")
+ "memory-mb"
+ } else k, v)
+ }
+
+ logDebug(s"Custom resource types: $resourceTypes")
+ resourceTypes.foreach { rt =>
+ val resourceName: String = rt._1
+ val (amount, unit) = getAmountAndUnit(rt._2)
+ logDebug(s"Registering resource with name: $resourceName, amount:
$amount, unit: $unit")
+
+ try {
+ val resInfoClass = Utils.classForName(
+ "org.apache.hadoop.yarn.api.records.ResourceInformation")
+ val setResourceInformationMethod =
+ resource.getClass.getMethod("setResourceInformation",
classOf[String],
+ resInfoClass)
+
+ val resourceInformation =
+ createResourceInformation(resourceName, amount, unit,
resInfoClass)
+ setResourceInformationMethod.invoke(resource, resourceName,
resourceInformation)
+ } catch {
+ case e: InvocationTargetException =>
+ if (e.getCause != null) {
+ throw e.getCause
+ } else {
+ throw e
+ }
+ case NonFatal(e) =>
+ logWarning(resourceTypesNotAvailableErrorMessage, e)
+ }
+ }
+ resource
+ }
+
+ def getCustomResourcesAsStrings(resource: Resource): String = {
+ if (resource == null) {
+ throw new IllegalArgumentException("Resource parameter should not be
null!")
+ }
+
+ if (!ResourceTypeHelper.isYarnResourceTypesAvailable()) {
+ logWarning(resourceTypesNotAvailableErrorMessage)
+ return ""
+ }
+
+ var res: String = ""
+ try {
+ val resUtilsClass = Utils.classForName(
+ "org.apache.hadoop.yarn.util.resource.ResourceUtils")
+ val getNumberOfResourceTypesMethod =
resUtilsClass.getMethod("getNumberOfKnownResourceTypes")
+ val numberOfResourceTypes: Int =
getNumberOfResourceTypesMethod.invoke(null).asInstanceOf[Int]
+ val resourceClass = Utils.classForName(
+ "org.apache.hadoop.yarn.api.records.Resource")
+
+ // skip memory and vcores (index 0 and 1)
+ for (i <- 2 until numberOfResourceTypes) {
+ val getResourceInfoMethod =
resourceClass.getMethod("getResourceInformation",
+ classOf[Int])
+ res ++= getResourceInfoMethod.invoke(resource,
i.asInstanceOf[AnyRef]).toString()
+ }
+ } catch {
+ case e: InvocationTargetException =>
+ if (e.getCause != null) {
+ throw e.getCause
+ } else {
+ throw e
+ }
+ case NonFatal(e) =>
+ logWarning(resourceTypesNotAvailableErrorMessage, e)
+
+ }
+ res
+ }
+
+ def getAmountAndUnit(s: String): (Long, String) = {
+ try {
+ val AMOUNT_AND_UNIT_REGEX(amount, unit) = s
+ (amount.toLong, convertToYarnResourceFormat(unit))
+ } catch {
+ case _: MatchError => throw new IllegalArgumentException(
+ s"Value of resource type should match pattern
$AMOUNT_AND_UNIT_REGEX, unmatched value: $s")
+ }
+ }
+
+ def convertToYarnResourceFormat(unit: String): String = {
--- End diff --
I actually don't think this is a good idea. When you look at YARN's
documentation for resource types, you see that they explicitly specify "m" and
"p" as corresponding to milli and pico, not mega and peta. See
[here](https://hadoop.apache.org/docs/r3.0.0/hadoop-yarn/hadoop-yarn-site/ResourceModel.html).
You can seethe units defined in code
[here](https://github.com/apache/hadoop/blob/532bbf4602e418276961cfc4f6b5e3f70e9cedc0/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-api/src/main/java/org/apache/hadoop/yarn/util/UnitsConversionUtil.java#L51).
Now, I cannot think of a single situation where someone would want to
request a pico quantity of a resource, but if some user does need this feature
at some point, this code will silently request the wrong quantity by changing
the unit to peta. Which is bad, because fixing it in a backwards-compatible
fashion would be challenging.
I realize that you may have done this because it is valid to request 8
megabytes in spark via "8m". Two options come to mind: Apply this function only
for memory, OR disallow requesting memory and vcores via resource types. It
looks like vanzin already suggests disallowing seting vcores and memory this
way, so that's probably the right thing to do, since you have less code (you
can drop this function).
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
