Two of them, I believe - the master for the cluster as a whole, and, for any individual job, its driver.
I wasn't really meaning to say that a Spark cluster is as robust as a mainframe (though maybe someone's figured out some tricks), more that I could totally see some sales folks conducting the same demo using something like Spark.
If you're using Spark's built-in scheduler then the cluster manager is a SPOF. Hadoop docs say you can get active/standy ResourceManager, not that I've tried it. Spark can also use k8s and nomad to schedule executors, and those have HA modes as well. I assume Mesos does HA.
You're still boned if the driver dies. I am pretty sure that the driver keeps some important state in RAM so if the node hosting it goes down you have to restart from the beginning, even if the cluster manager restarts the driver.
I wasn't really meaning to say that a Spark cluster is as robust as a mainframe (though maybe someone's figured out some tricks), more that I could totally see some sales folks conducting the same demo using something like Spark.