spark on kubernetes: Expected HTTP 101 response but was '403 Forbidden'

问题

I am using spark 2.4.4. Having trouble running spark-submit on my ubuntu machine with the following config and system environment:

cpchung:small$ minikube start --kubernetes-version=1.17.0
😄  minikube v1.6.2 on Ubuntu 19.10
✨  Selecting 'kvm2' driver from user configuration (alternates: [virtualbox none])
💡  Tip: Use 'minikube start -p <name>' to create a new cluster, or 'minikube delete' to delete this one.
🏃  Using the running kvm2 "minikube" VM ...
⌛  Waiting for the host to be provisioned ...
🐳  Preparing Kubernetes v1.17.0 on Docker '18.09.9' ...
💾  Downloading kubelet v1.17.0
💾  Downloading kubeadm v1.17.0
🚜  Pulling images ...
🚀  Launching Kubernetes ... 
🏄  Done! kubectl is now configured to use "minikube"
cpchung:small$ minikube docker-env
export DOCKER_TLS_VERIFY="1"
export DOCKER_HOST="tcp://192.168.39.246:2376"
export DOCKER_CERT_PATH="/home/cpchung/.minikube/certs"
# Run this command to configure your shell:
# eval $(minikube docker-env)
cpchung:small$ eval $(minikube docker-env)
cpchung:small$ ./run.sh

inside my run.sh is this:

spark-submit --class SimpleApp \
--master "k8s://https://192.168.39.246:8443" \
--deploy-mode cluster \
--conf spark.kubernetes.container.image=somename:latest \
--conf spark.kubernetes.image.pullPolicy=IfNotPresent \
--conf spark.executor.instances=1 \
--conf spark.kubernetes.authenticate.driver.serviceAccountName=spark \
local:///small.jar

Then I got this:

simpleapp-1580791262398-driver   0/1     Error    0          21s
cpchung:small$ kubectl logs simpleapp-1580791262398-driver

20/02/04 04:41:10 INFO Utils: Successfully started service 'SparkUI' on port 4040.
20/02/04 04:41:10 INFO SparkUI: Bound SparkUI to 0.0.0.0, and started at http://simpleapp-1580791262398-driver-svc.default.svc:4040
20/02/04 04:41:10 INFO SparkContext: Added JAR file:///small.jar at spark://simpleapp-1580791262398-driver-svc.default.svc:7078/jars/small.jar with timestamp 1580791270603
20/02/04 04:41:14 INFO ExecutorPodsAllocator: Going to request 1 executors from Kubernetes.
20/02/04 04:41:14 WARN WatchConnectionManager: Exec Failure: HTTP 403, Status: 403 - 
java.net.ProtocolException: Expected HTTP 101 response but was '403 Forbidden'

回答1:

Spark up to v2.4.4 uses fabric8 Kubernetes client v4.1.2 which works only with Kubernetes APIs 1.9.0 - 1.12.0, please refer compatibility matrix and pom.xml.

Spark 2.4.5 will upgrade Kubernetes client to v4.6.1 and will support Kubernetes APIs up to 1.15.2.

So you have the following options:

Downgrade Kubernetes cluster to 1.12
Wait until Spark 2.4.5/3.0.0 is released and downgrade Kubernetes cluster to 1.15.2
Upgrade Spark fabric8 Kubernetes client dependency and do custom build of Spark and its Docker image

Hope it helps.

Update 1

In order to run Spark on Kubernetes cluster with API version 1.17.0+ you need to patch Spark code based on tag v2.4.5 with the following changes:

diff --git a/resource-managers/kubernetes/core/pom.xml b/resource-managers/kubernetes/core/pom.xml
index c9850ca512..41468e1363 100644
--- a/resource-managers/kubernetes/core/pom.xml
+++ b/resource-managers/kubernetes/core/pom.xml
@@ -29,7 +29,7 @@
   <name>Spark Project Kubernetes</name>
   <properties>
     <sbt.project.name>kubernetes</sbt.project.name>
-    <kubernetes.client.version>4.6.1</kubernetes.client.version>
+    <kubernetes.client.version>4.7.1</kubernetes.client.version>
   </properties>

   <dependencies>
diff --git a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
index 575bc54ffe..c180ea3a7b 100644
--- a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
+++ b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
@@ -67,7 +67,8 @@ private[spark] class BasicDriverFeatureStep(
       .withAmount(driverCpuCores)
       .build()
     val driverMemoryQuantity = new QuantityBuilder(false)
-      .withAmount(s"${driverMemoryWithOverheadMiB}Mi")
+      .withAmount(driverMemoryWithOverheadMiB.toString)
+      .withFormat("Mi")
       .build()
     val maybeCpuLimitQuantity = driverLimitCores.map { limitCores =>
       ("cpu", new QuantityBuilder(false).withAmount(limitCores).build())
diff --git a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
index d89995ba5e..9c589ace92 100644
--- a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
+++ b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
@@ -86,7 +86,8 @@ private[spark] class BasicExecutorFeatureStep(
     // executorId
     val hostname = name.substring(Math.max(0, name.length - 63))
     val executorMemoryQuantity = new QuantityBuilder(false)
-      .withAmount(s"${executorMemoryTotal}Mi")
+      .withAmount(executorMemoryTotal.toString)
+      .withFormat("Mi")
       .build()
     val executorCpuQuantity = new QuantityBuilder(false)
       .withAmount(executorCoresRequest)

QuantityBuilder has changed the way of parsing the input with this PR which is available since fabric8 Kubernetes client v4.7.0.

Update 2

Spark images are based on openjdk:8-jdk-slim which runs Java 8u252 which in order has a bug related to OkHttp. To fix it we require fabric8 Kubernetes client v4.9.2, please refer its release notes for more details.

Also the patch above can be simplified:

diff --git a/resource-managers/kubernetes/core/pom.xml b/resource-managers/kubernetes/core/pom.xml
index c9850ca512..595aaba8dd 100644
--- a/resource-managers/kubernetes/core/pom.xml
+++ b/resource-managers/kubernetes/core/pom.xml
@@ -29,7 +29,7 @@
   <name>Spark Project Kubernetes</name>
   <properties>
     <sbt.project.name>kubernetes</sbt.project.name>
-    <kubernetes.client.version>4.6.1</kubernetes.client.version>
+    <kubernetes.client.version>4.9.2</kubernetes.client.version>
   </properties>

   <dependencies>
diff --git a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
index 575bc54ffe..a1559e07a4 100644
--- a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
+++ b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicDriverFeatureStep.scala
@@ -66,9 +66,7 @@ private[spark] class BasicDriverFeatureStep(
     val driverCpuQuantity = new QuantityBuilder(false)
       .withAmount(driverCpuCores)
       .build()
-    val driverMemoryQuantity = new QuantityBuilder(false)
-      .withAmount(s"${driverMemoryWithOverheadMiB}Mi")
-      .build()
+    val driverMemoryQuantity = new Quantity(s"${driverMemoryWithOverheadMiB}Mi")
     val maybeCpuLimitQuantity = driverLimitCores.map { limitCores =>
       ("cpu", new QuantityBuilder(false).withAmount(limitCores).build())
     }
diff --git a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
index d89995ba5e..b439ebf837 100644
--- a/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
+++ b/resource-managers/kubernetes/core/src/main/scala/org/apache/spark/deploy/k8s/features/BasicExecutorFeatureStep.scala
@@ -85,9 +85,7 @@ private[spark] class BasicExecutorFeatureStep(
     // name as the hostname.  This preserves uniqueness since the end of name contains
     // executorId
     val hostname = name.substring(Math.max(0, name.length - 63))
-    val executorMemoryQuantity = new QuantityBuilder(false)
-      .withAmount(s"${executorMemoryTotal}Mi")
-      .build()
+    val executorMemoryQuantity = new Quantity(s"${executorMemoryTotal}Mi")
     val executorCpuQuantity = new QuantityBuilder(false)
       .withAmount(executorCoresRequest)
       .build()

来源：https://stackoverflow.com/questions/60050782/spark-on-kubernetes-expected-http-101-response-but-was-403-forbidden

标签

Docker

apache-spark

Kubernetes

Minikube