Re: Spark on k8s cluster mode, from outside of the cluster. [SOLVED]
I read some of the notes here. I have been away from Zeppelin for a while
but have extensive experience with Spark on Kubernetes (k8s).
First of all I assume that the zeppelin server is just the client that you
are running to submit a job to Spark on k8s.
If the Spark on k8s is offered as a service (say Google GKE) etc, then that
GKE has a defined name (has been created) that you can address by using
something like
KUBERNETES_MASTER_IP=$(gcloud container clusters list --filter
name= --format='value(MASTER_IP)')
so this way you get the master ip address. in your spark-submit job on your
client (where you have Spark binary installed (same version as the docker
image is used inside k8s (say spark version 3.1.1)), then you can initiate
(this is with Python example)
spark-submit --verbose \
--properties-file ${property_file} \
--master k8s://https://$KUBERNETES_MASTER_IP:443 \
--deploy-mode cluster \
--name sparkBQ \
--conf
spark.yarn.appMasterEnv.PYSPARK_PYTHON=./pyspark_venv/bin/python \
--py-files $CODE_DIRECTORY/DSBQ.zip \
--conf spark.kubernetes.namespace=$NAMESPACE \
--conf spark.network.timeout=300 \
--conf spark.executor.instances=$NEXEC \
--conf spark.kubernetes.driver.limit.cores=1 \
Note that $NEXEC is the number of executors requested. Current spark model
works on the basis of the "one-container-per-Pod" model
Re: Spark on k8s cluster mode, from outside of the cluster. [SOLVED]
Thanks for the sharing, it would be nice if you can write a blog to share it with more wide zeppelin users. Fabrizio Fab 于2021年10月28日周四 下午4:29写道： > > Yeah ! Thank you very much Philipp: tonight I explored carefully the > source code and discovered the 2 thrift servers stuff. > > Therefore I solved my problem: here the solution adopted, which can be > useful for other people. > > CONTEXT > I have my Zeppelin Server installation located into a LAN, where a K8s > Cluster is available, and I want to submit notes in cluster mode over the > k8s Cluster. > > SOLUTION > - the driver pod must have its address exposed on the LAN network, > otherwise the Zeppelin server cannot connect to the Interpreter Thrift > server: I suppose that there are several ways of doing this, but I am not a > k8s expert so I simply created a basic driver-pod.template.yaml with a > "hostNetwork" spec and referenced it by the > "spark.kubernetes.driver.podTemplateFile" interpreter setting. > > At this point, the 2 servers can talk each other. > > NOTE > 1) do not set the zeppelin run mode = k8s. It must be "local" (or the > default "auto") > 2) a NFS share (or other shared persistent volume) is required in order to > upload the required JARS and easily access the driver logs when the driver > shuts down: > > spark.kubernetes.driver.volumes.nfs..options.server= server> > spark.kubernetes.driver.volumes.nfs..options.path= path> > spark.kubernetes.driver.volumes.nfs..mount.path= path> > > > > > > > > > > > > > > > > On 2021/10/28 06:48:54, Philipp Dallig wrote: > > Hi Fabrizio, > > > > We have two connections. First, the Zeppelin interpreter opens a > > connection to the Zeppelin server to register and to send back the > > interpreter output. The Zeppelin server is the CALLBACK_HOST and the > > PORT indicates where the Zeppelin server opened the Thrift service for > > the Zeppelin interpreter. > > > > An important part of the registration is that the Zeppelin interpreter > > tells the Zeppelin server where the interpreter pod has an open Thrifts > > server port. This information can be found in the Zeppelin server log > > output. Be on the lookout for this message. > > > https://github.com/apache/zeppelin/blob/master/zeppelin-plugins/launcher/k8s-standard/src/main/java/org/apache/zeppelin/interpreter/launcher/K8sRemoteInterpreterProcess.java#L483 > > Also note the function ZEPPELIN_K8S_PORTFORWARD, which should help your > > Zeppelin server to reach the Zeppelin interpreter in K8s. > > > > > the 1st "spark-submit" in "cluster mode" is started from the client > > (in the zeppelin host, in our case), then the 2nd "spark-submit" in > > "client mode" is started by the "/opt/entrypoint.sh" script inside the > > standard spark docker image. > > > > Are you sure you are using the K8s launcher? As you can see in this part > > of the code > > ( > https://github.com/apache/zeppelin/blob/2f55fe8ed277b28d71f858633f9c9d76fd18f0c3/zeppelin-plugins/launcher/k8s-standard/src/main/java/org/apache/zeppelin/interpreter/launcher/K8sRemoteInterpreterProcess.java#L411), > > > Zeppelin always uses client mode. > > > > The architecture is quite simple: > > > > Zeppelin-Server -> Zeppelin-Interpreter (with Spark in client mode) on > > K8s -> x-Spark-executors (based on your config) > > > > Best Regards > > Philipp > > > > > > Am 27.10.21 um 15:19 schrieb Fabrizio Fab: > > > > > Hi Philipp, okay, I realized just now of my HUGE misunderstanding ! > > > > > > The "double-spark-submit" patter is just the standard spark-on-k8s way > of running spark applications in cluster mode: > > > the 1st "spark-submit" in "cluster mode" is started from the client > (in the zeppelin host, in our case), then the 2nd "spark-submit" in "client > mode" is started by the "/opt/entrypoint.sh" script inside the standard > spark docker image. > > > > > > At this point I can make a more precise question: > > > > > > I see that the interpreter.sh starts the RemoteInterpreterServer with, > in particular the following paramters: CALLBACK_HOST / PORT > > > They refers to the Zeppelin host and RPC port > > > > > > Moreover, when the interpreter starts, it runs a Thrift server on some > random port. > > > > > > So, I ask: which communications are supposed to happen, in order to > correctly set-up my firewall/routing rules ? > > > > > > -1 Must the Zeppelin server connect to the Interpreter Thrift server ? > > > -2 Must the Interpreter Thrift server connect to the Zeppelin server? > > > -3 Both ? > > > > > > - Which ports must the Zeppelin server/ The thrift server find open > on the other server ? > > > > > > Thank you everybody! > > > > > > Fabrizio > > > > > > > > > > > > > > > On 2021/10/26 11:40:24, Philipp Dallig > wrote: > > >> Hi Fabrizio, > > >> > > >> At the moment I think zeppelin does not support running spark jobs in > > >> cluster mode. But in fact K8s mode simulates cluster mode. Because the > > >> Zeppelin interpreter is already started as a pod in K8s, as a manual > > >>
Re: Spark on k8s cluster mode, from outside of the cluster. [SOLVED]
Yeah ! Thank you very much Philipp: tonight I explored carefully the source code and discovered the 2 thrift servers stuff. Therefore I solved my problem: here the solution adopted, which can be useful for other people. CONTEXT I have my Zeppelin Server installation located into a LAN, where a K8s Cluster is available, and I want to submit notes in cluster mode over the k8s Cluster. SOLUTION - the driver pod must have its address exposed on the LAN network, otherwise the Zeppelin server cannot connect to the Interpreter Thrift server: I suppose that there are several ways of doing this, but I am not a k8s expert so I simply created a basic driver-pod.template.yaml with a "hostNetwork" spec and referenced it by the "spark.kubernetes.driver.podTemplateFile" interpreter setting. At this point, the 2 servers can talk each other. NOTE 1) do not set the zeppelin run mode = k8s. It must be "local" (or the default "auto") 2) a NFS share (or other shared persistent volume) is required in order to upload the required JARS and easily access the driver logs when the driver shuts down: spark.kubernetes.driver.volumes.nfs..options.server= spark.kubernetes.driver.volumes.nfs..options.path= spark.kubernetes.driver.volumes.nfs..mount.path= On 2021/10/28 06:48:54, Philipp Dallig wrote: > Hi Fabrizio, > > We have two connections. First, the Zeppelin interpreter opens a > connection to the Zeppelin server to register and to send back the > interpreter output. The Zeppelin server is the CALLBACK_HOST and the > PORT indicates where the Zeppelin server opened the Thrift service for > the Zeppelin interpreter. > > An important part of the registration is that the Zeppelin interpreter > tells the Zeppelin server where the interpreter pod has an open Thrifts > server port. This information can be found in the Zeppelin server log > output. Be on the lookout for this message. > https://github.com/apache/zeppelin/blob/master/zeppelin-plugins/launcher/k8s-standard/src/main/java/org/apache/zeppelin/interpreter/launcher/K8sRemoteInterpreterProcess.java#L483 > Also note the function ZEPPELIN_K8S_PORTFORWARD, which should help your > Zeppelin server to reach the Zeppelin interpreter in K8s. > > > the 1st "spark-submit" in "cluster mode" is started from the client > (in the zeppelin host, in our case), then the 2nd "spark-submit" in > "client mode" is started by the "/opt/entrypoint.sh" script inside the > standard spark docker image. > > Are you sure you are using the K8s launcher? As you can see in this part > of the code > (https://github.com/apache/zeppelin/blob/2f55fe8ed277b28d71f858633f9c9d76fd18f0c3/zeppelin-plugins/launcher/k8s-standard/src/main/java/org/apache/zeppelin/interpreter/launcher/K8sRemoteInterpreterProcess.java#L411), > > Zeppelin always uses client mode. > > The architecture is quite simple: > > Zeppelin-Server -> Zeppelin-Interpreter (with Spark in client mode) on > K8s -> x-Spark-executors (based on your config) > > Best Regards > Philipp > > > Am 27.10.21 um 15:19 schrieb Fabrizio Fab: > > > Hi Philipp, okay, I realized just now of my HUGE misunderstanding ! > > > > The "double-spark-submit" patter is just the standard spark-on-k8s way of > > running spark applications in cluster mode: > > the 1st "spark-submit" in "cluster mode" is started from the client (in the > > zeppelin host, in our case), then the 2nd "spark-submit" in "client mode" > > is started by the "/opt/entrypoint.sh" script inside the standard spark > > docker image. > > > > At this point I can make a more precise question: > > > > I see that the interpreter.sh starts the RemoteInterpreterServer with, in > > particular the following paramters: CALLBACK_HOST / PORT > > They refers to the Zeppelin host and RPC port > > > > Moreover, when the interpreter starts, it runs a Thrift server on some > > random port. > > > > So, I ask: which communications are supposed to happen, in order to > > correctly set-up my firewall/routing rules ? > > > > -1 Must the Zeppelin server connect to the Interpreter Thrift server ? > > -2 Must the Interpreter Thrift server connect to the Zeppelin server? > > -3 Both ? > > > > - Which ports must the Zeppelin server/ The thrift server find open on the > > other server ? > > > > Thank you everybody! > > > > Fabrizio > > > > > > > > > > On 2021/10/26 11:40:24, Philipp Dallig wrote: > >> Hi Fabrizio, > >> > >> At the moment I think zeppelin does not support running spark jobs in > >> cluster mode. But in fact K8s mode simulates cluster mode. Because the > >> Zeppelin interpreter is already started as a pod in K8s, as a manual > >> Spark submit execution would do in cluster mode. > >> > >> Spark-submit is called only once during the start of the Zeppelin > >> interpreter. You will find the call in these lines: > >> https://github.com/apache/zeppelin/blob/2f55fe8ed277b28d71f858633f9c9d76fd18f0c3/bin/interpreter.sh#L303-L305 > >> > >> Best Regards > >>
