Github user szyszy commented on a diff in the pull request:
https://github.com/apache/spark/pull/20761#discussion_r190024350
--- 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 --
Yes, with g/G there is no issue.
I see Spark uses 'm' only for memory nowadays.
When you define a custom resource type with YARN, you can indeed specify
its unit in resource-types.xml, meaning this would be the default unit.
So back to your example, if a resource has M as defined default unit with
YARN's config, it does also mean that YARN would accept 'm' as a unit and it
will convert the amount of 'm' to 'M' (millis to mega).
More precisely, default units are for cases when you request a resource but
don't specify the units, e.g. request 5 and YARN will understand it as request
5M.
So in this sense, requesting any amount of 'm' of a custom resource is not
an exceptional case in my opinion.
On the other hand I understand your concern and I see that this conversion
is hacky and not clean nor future proof.
I can come up with the following:
- 'g' (giga) will auto-convert to 'G'
- 'm' / 'M' (millis / mega) will not be auto-converted as it ambiguous
(milli vs mega), in this case Spark documentation should be clear that if 'm'
is specified then millis will be requested, if 'M' is specified then mega will
be requested from YARN.
- 'p' / 'P' (pico / peta): same as millis/mega
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
