I'm trying to run multiple spark clients on Airflow(ETL scheduler).
I'm running in cluster mode on YARN, therefore ApplicationMaster Executor and Driver are all running on executor in Yarn context.
However, my spark client which sample the process and monitor the state is running in airflow worker.
The problem is that the Spark client take lot's of memory ~500 MB per job. It may sound as not much in terms of executors or drivers but for the role of spark client it sounds crazy.
My question is, how can I configure/manipulate spark client memory/cpu requirements can I limit it's intervals ? can I limit it's memory with flags?
So in spark code it make a distinction if it's running in standalone mode or cluster mode. For standalone it set a default of -Xmx 1G and in cluster mode it doesn't have default but it trying to read java options from environment variable called SPARK_SUBMIT_OPTS.
So if you wanna set any java opts for the client java process only use SPARK_SUBMIT_OPTS
I have a daily pipeline running on Spark Standalone 2.1. Its deployed in and runs on AWS EC2 and uses S3 for its persistence layer. For the most part, the pipeline runs without a hitch, but occasionally the job hangs on a single worker node during a reduceByKey operation. When I work into the worker, I notice that the CPU (as seen via top) is pegged at 100%. My remedy so far is to reboot the worker node so that Spark re-assigns the task and the job proceeds fine from there.
I would like to be able to mitigate this issue. I gather that I can prevent CPU pegging by switching to use YARN as my cluster manager, but I wonder whether I could configure Spark Standalone to prevent CPU pegging by maybe limiting the number of cores that get assigned to the Spark job ? Any suggestions would be greatly appreciated.
The spark docs have the following paragraph that describes the difference between yarn client and yarn cluster:
There are two deploy modes that can be used to launch Spark applications on YARN. In cluster mode, the Spark driver runs inside an application master process which is managed by YARN on the cluster, and the client can go away after initiating the application. In client mode, the driver runs in the client process, and the application master is only used for requesting resources from YARN.
I'm assuming there are two choices for a reason. If so, how do you choose which one to use?
Please use facts to justify your response so that this question and answer(s) meet stackoverflow's requirements.
There are a few similar questions on stackoverflow, however those questions focus on the difference between the two approaches, but don't focus on when one approach is more suitable than the other.
A common deployment strategy is to submit your application from a gateway machine that is physically co-located with your worker machines (e.g. Master node in a standalone EC2 cluster). In this setup, client mode is appropriate. In client mode, the driver is launched directly within the spark-submit process which acts as a client to the cluster. The input and output of the application is attached to the console. Thus, this mode is especially suitable for applications that involve the REPL (e.g. Spark shell).
Alternatively, if your application is submitted from a machine far from the worker machines (e.g. locally on your laptop), it is common to use cluster mode to minimize network latency between the drivers and the executors. Note that cluster mode is currently not supported for Mesos clusters. Currently only YARN supports cluster mode for Python applications." -- Submitting Applications
What I understand from this is that both strategies use the cluster to distribute tasks; the difference is where the "driver program" runs: locally with spark-submit, or, also in the cluster.
When you should use either of them is detailed in the quote above, but I also did another thing: for big jars, I used rsync to copy them to the cluster (or even to master node) with 100 times the network speed, and then submitted from the cluster. This can be better than "cluster mode" for big jars. Note that client mode does not probably transfer the jar to the master. At that point the difference between the 2 is minimal. Probably client mode is better when the driver program is idle most of the time, to make full use of cores on the local machine and perhaps avoid transferring the jar to the master (even on loopback interface a big jar takes quite a bit of seconds). And with client mode you can transfer (rsync) the jar on any cluster node.
On the other hand, if the driver is very intensive, in cpu or I/O, cluster mode may be more appropriate, to better balance the cluster (in client mode, the local machine would run both the driver and as many workers as possible, making it over loaded and making it that local tasks will be slower, making it such that the whole job may end up waiting for a couple of tasks from the local machine).
Conclusion :
To sum up, if I am in the same local network with the cluster, I would
use the client mode and submit it from my laptop. If the cluster is
far away, I would either submit locally with cluster mode, or rsync
the jar to the remote cluster and submit it there, in client or
cluster mode, depending on how heavy the driver program is on
resources.*
AFAIK With the driver program running in the cluster, it is less vulnerable to remote disconnects crashing the driver and the entire spark job.This is especially useful for long running jobs such as stream processing type workloads.
Spark Jobs Running on YARN
When running Spark on YARN, each Spark executor runs as a YARN container. Where MapReduce schedules a container and fires up a JVM for each task, Spark hosts multiple tasks within the same container. This approach enables several orders of magnitude faster task startup time.
Spark supports two modes for running on YARN, “yarn-cluster” mode and “yarn-client” mode. Broadly, yarn-cluster mode makes sense for production jobs, while yarn-client mode makes sense for interactive and debugging uses where you want to see your application’s output immediately.
Understanding the difference requires an understanding of YARN’s Application Master concept. In YARN, each application instance has an Application Master process, which is the first container started for that application. The application is responsible for requesting resources from the ResourceManager, and, when allocated them, telling NodeManagers to start containers on its behalf. Application Masters obviate the need for an active client — the process starting the application can go away and coordination continues from a process managed by YARN running on the cluster.
In yarn-cluster mode, the driver runs in the Application Master. This means that the same process is responsible for both driving the application and requesting resources from YARN, and this process runs inside a YARN container. The client that starts the app doesn’t need to stick around for its entire lifetime.
yarn-cluster mode
The yarn-cluster mode is not well suited to using Spark interactively, but the yarn-client mode is. Spark applications that require user input, like spark-shell and PySpark, need the Spark driver to run inside the client process that initiates the Spark application. In yarn-client mode, the Application Master is merely present to request executor containers from YARN. The client communicates with those containers to schedule work after they start:
yarn-client mode
This table offers a concise list of differences between these modes:
Reference: https://blog.cloudera.com/blog/2014/05/apache-spark-resource-management-and-yarn-app-models/ - Apache Spark Resource Management and YARN App Models (web.archive.com mirror)
In yarn-cluster mode, the driver program will run on the node where application master is running where as in yarn-client mode the driver program will run on the node on which job is submitted on centralized gateway node.
Both answers by Ram and Chavati/Wai Lee are useful and helpful, but here I just want to added a couple of other points:
Much has been written about the proximity of driver to the executors, and that is a big factor. The other factors are:
Will the driver process be around until execution of job is finished?
How's returned data being handled?
#1. In spark client mode, the driver process must be up and running the whole time when the spark job is in execution. So if you have a truly long job that say take many hours to run, you need to make sure the driver process is still up and running, and that the driver session is not auto-logout.
On the other hand, after submitting a job to run in cluster mode, the process can go away. The cluster mode will keep running. So this is typically how a production job will run: the job can be triggered by a timer, or by an external event and then the job will run to its completion without worrying about the lifetime of the process submitting the spark job.
#2. In client mode, you can call sc.collect() to gather all the data back from all executors, and then write/save the returned data to a local Linux file on local disk. Now this may not work for cluster mode, as the 'driver' typically run in a different remote host. The data written up therefore need to be persisted in a common mounted volume (such as GPFS, NFS) or in distributed file system like HDFS. If the job is running under Hadoop/YARN, the more common way for cluster mode is simply ask each executor to persist the data to HDFS, and not to run collect( ) at all. Collect() actually have scalability issue when there are a large number of executors returning large amount of data - it can overwhelm the driver process.
I'm running a Spark Streaming process on a 16 CPU's 64 GB RAM host with Mesos.
When I'm running it using Mesos as a cluster manager (by setting --master mesos://leader.mesos:5050) it's running much slower than when it is run in local mode (--master local[4]).
I can't find the reason for that and I have no clue. One of the things I've noticed is that there is one specific task that is taking significantly more time on Mesos than in Local.
The weird thing (maybe that should be the questions' title) is that the task itself takes 6s and its stage (it has only one stage) takes less than a second. See attached pictures (Mesos (1) and (2)). How come? Isn't a job equal to the sum of its parts?
Local:
Mesos:
(1)
(2)
Another note: I did manage to run this exact same Spark Streaming process on another Mesos cluster, and it runs in a sensible amount of time, pretty much like in the local mode described above. The only difference that I can think of is that this cluster has more than one host, and that Spark is running with 2 executors rather than 1. (I couldn't find a way to run more than 1 executor on the same host on Mesos). Is this may be the reason?
Any clues would be much appreciated.
Spark can run over Mesos in two modes: coarse-grained (default) and fine-grained (see documentation).
In coarse-grained mode Spark launches exactly one executor on each machine it was assigned to by Mesos. Inside this task Spark launches other mini-tasks. It has the benefit of lower startup overhead (in your case you don't want to change this mode).
Could you be more specific about your streaming job? Is it CPU, disk, or network bounded? You can easily compare performance if you run some of Spark examples.
If your task is CPU intensive you might consider setting spark.mesos.extra.cores. By default Spark tries to acquire all cores that are being offered by Mesos. So, if there's no other task running on that cluster it shouldn't be a problem.
I have a spark application that finishes without error, but once it's done and saved all of its outputs and the process terminates, the Spark standalone cluster master process becomes a CPU hog, using 16 CPU's full time for hours, and the web UI becomes unresponsive. I have no idea what it could be doing, is there some complicated clean up step?
Some more details:
I've got a Spark standalone cluster (27 workers/nodes) that I've been successfully submitting jobs to for a while. I recently scaled up the size of my applications, the largest now takes 3.5 hours using 100 cores over 27 workers, and each worker has ~dozens of GB of shuffle read/write over the course of the job. Otherwise, the application is no different than the smaller jobs that have run successfully before.
This is a known issue with Spark's standalone cluster, and is caused by the massive event log created by large applications. You can read more at the issue tracking link below.
https://issues.apache.org/jira/browse/SPARK-12299
At the current time, the best work-around is to disable event logging for large jobs.