[jira] [Created] (FLINK-8886) Job isolation via scheduling in shared cluster

["Elias Levy (JIRA)" <ji...@apache.org>]

Elias Levy created FLINK-8886:
---------------------------------

             Summary: Job isolation via scheduling in shared cluster
                 Key: FLINK-8886
                 URL: https://issues.apache.org/jira/browse/FLINK-8886
             Project: Flink
          Issue Type: Improvement
          Components: Scheduler
    Affects Versions: 1.5.0
            Reporter: Elias Levy


Flink's TaskManager executes tasks from different jobs within the same JMV as threads.  We prefer to isolate different jobs on their on JVM.  Thus, we must use different TMs for different jobs.  As currently the scheduler will allocate task slots within a TM to tasks from different jobs, that means we must stand up one cluster per job.  This is wasteful, as it requires at least two JobManagers per cluster for high-availability, and the JMs have low utilization.

Additionally, different jobs may require different resources.  Some jobs are compute heavy.  Some are IO heavy (lots of state in RocksDB).  At the moment the scheduler threats all TMs are equivalent, except possibly in their number of available task slots.  Thus, one is required to stand up multiple cluster if there is a need for different types of TMs.

 

It would be useful if one could specify requirements on job, such that they are only scheduled on a subset of TMs.  Properly configured, that would permit isolation of jobs in a shared cluster and scheduling of jobs with specific resource needs.

 

One possible implementation is to specify a set of tags on the TM config file which the TMs used when registering with the JM, and another set of tags configured within the job or supplied when submitting the job.  The scheduler could then match the tags in the job with the tags in the TMs.  In a restrictive mode the scheduler would assign a job task to a TM only if all tags match.  In a relaxed mode the scheduler could assign a job task to a TM if there is a partial match, while giving preference to a more accurate match.

 

 

 

 



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)