Welcome to the second part of this introduction to the YARN API. Today I will walk through the resources available for a job and how to find your queue.
How Many Total Resources Are Available for My Job?
When configuring a spark driver, each driver and executor cannot be larger than the size of one Yarn container. Additionally, the total resources requested in terms of both CPU and memory cannot be larger than all the resources available on the cluster to the user submitting the job. In other words, we can consider the resource requirements to be governed by the following equations.
Total vCores Available ≥ spark.executor.cores ✕ spark.executor.instances + 1
(Where the 1 is for the one core required by the driver)
Total_MB_Available ≥(spark.executor.memory + executor memory overhead)
+ spark.driver.memory + driver memory overhead
Securing this information particularly in a system with queuing configured and other active users can be a little complicated.
The behavior of a job that is submitted with too many total resources seems to differ depending on the user, cluster, and how excessive the request is. My experience has been that if the resources requested are larger than the size of the cluster, according to the configuration, the job will fail with a meaningful error. If however a Spark Application requests more resources than are available on the cluster at the time of submission, or more resources than are available in this queue (but not in the cluster), but less than the size of the whole cluster, the job may hang in an “unassigned” state until such time (in many cases never) those resources become available. Because my program requires configuring Spark jobs submitted by users of our applications, who often do not themselves have access to the resource manager and thus don’t have a way to learn about the traffic on their cluster, it is important to avoid this case. It is possible that in a different scenario making a large request and waiting for resources to be available might be desirable.
Furthermore, here is what I learned about querying the Yarn Scheduler API to determine cluster burn. I am going to try to keep the explanation simple here, but I found this blog really useful for me when I was trying to understand the yarn scheduler.
The Yarn scheduler API has a “scheduler” object. Inside the scheduler object is a “schedulerInfo” object. The schedulerInfo object has a queue “type.” The “type” is the kind of queue, either “capacityScheduler” or “fairScheduler.” The rest of the structure of the object is different depending on the queue type. The values shown on the official documentation page are for the capcaityScheduler.
Find queue type from the scheduler API API Call: http://<rm http address:port>/ws/v1/cluster/scheduler Path to Json Scheduler
→ (will be either “capacityScheduler” or “fairScheduler” )
How to Find Your Queue
Each cluster has a root queue. Thus, the “schedulerInfo” object has a field called “rootQueue.” ‘rootQueue’ contains a nested object of queues.
Find the root queue from the Scheduler API API Call: http://<rm http address:port>/ws/v1/cluster/scheduler Path to Json scheduler
Note: This is another place where my experience differed from the documentation. I could not access the “queues” object directly, I had to get to the “rootQueue” first.
The queues are in tree structure, so you can search through the queues (I used a breadth first search) to find the one whose name corresponds to the name of the user submitting the application. In some instances, a queue with this name will not exist. In this case, the resources allocated for this job will be specified by the “default” queue, so you will need to do a second search to find the default queue object. In the case of the fair scheduler, the user will not have a queue until they have submitted one job. The first time they submit they will get the resources specified by the default queue. In subsequent submissions, the “queues” object will contain a queue corresponding to their username.
In the next blog, I will walk through the difference between the capacity scheduler and the fair scheduler in addition to how to calculate your memory overhead. Stay tuned for more information!