spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From jianl miao <>
Subject Integration about submitting and monitoring spark tasks
Date Tue, 24 Mar 2020 06:23:57 GMT
I now need to integrate spark into our own platform built with spring to
reflect the ability of task submission and task monitoring. Spark tasks run
on yarn and are in cluster mode. And our current service may submit tasks
to different yarn clusters.
According to the current method provided by spark, there are two classes,
sparkLauncher and InProcessLauncher. The former calls spark-submit through
child processes, and the latter calls the sparksubmit class in the current
process. But both methods are done through the launcher, and all will
always monitor the current task by themselves, which will cause yarn app
status queries to be very frequent as the number of tasks increases. In
essence, we can launch the submit program after the task submission is
completed, after getting the applicationId of the task and after the
submission is completed, and then uniformly monitor the status of all my
tasks. But now I can't stop the submit program, even if I configure
waitAppCompletion to false, because launcherBackend.isConnected is true.
This is very resource intensive.

View raw message