sla-plugin.md

Service Level Agreement (SLA) Plugin

Introduction

When users apply jobs to volcano, they may need adding some particular constraints to job, for example, longest Pending time aiming to prevent job from starving. And these constraints can be regarded as Service Level Agreement (SLA) which are agreed between volcano and user. So sla plugin is provided to receive and realize SLA settings for both individual job and whole cluster.

Solution

1. In sla plugin, arguments sla-waiting-time is provided to realize job resource reservation: sla-waiting-time is maximum time one job should stay Pending or inqueue status and not be allocated. When sla-waiting-time is over, sla plugin sets the job to be inqueue in enqueue action immediately. Then sla plugin locks idle resources pre-allocated to pods of this job in allocate action, even if the job has not been Ready yet. In this way, sla plugin realizes large job election and resource reservation, thus replaces elect & reserve action in v1.1.0.

2. Arguments sla-waiting-time can be set for one job, and for all jobs in cluster.

   1. For one job, user can set them in job annotations in following format:

      apiVersion: batch.volcano.sh/v1alpha1
      kind: Job
      metadata:
        annotations:
          sla-waiting-time: 1h2m3s

   2. For all jobs, user can set sla-waiting-time field in sla plugin arguments via volcano-scheduler-configmap in following format:

        actions: "enqueue, allocate, backfill"
        tiers:
        - plugins:
          - name: priority
          - name: gang
          - name: sla
            arguments:
              sla-waiting-time: 1h2m3s

3. sla plugin return 3 callback functions: JobEnqueueableFn, JobPipelinedFn, and JobOrderFn:

   1. JobEnqueueableFn returns Permit when job waiting time in Pending status is longer than sla-waiting-time, and job will go through enqueue action and be inqueue instantly, regardless of other plugins returning Reject or Abstain to reject this job from being inqueue.

   2. JobPipelinedFn returns Permit when job waiting time in inqueue status is longer than sla-waiting-time, and job will be Pipelined status instantly, regardless of other plugins returning Reject or Abstain to reject this job from being Pipelined. In this way allocate action reserves resources for pods of the job even if the job is not Ready yet.

   3. JobOrderFn adjusts the order of this job in waiting queues of enqueue & allocate action. The more close to sla-waiting-time that job waiting time is, the higher scored of this job in JobOrderFn of sla plugin, so that job would have larger probability to be front int priority queue, which means that it can touch more idle resources and have higher priority to be inqueue and allocated.

4. the execution flow chart of sla plugin is shown as below:

Feature Interaction

1. By now we only need 1 argument sla-waiting-time, so I add it into annotations for simplicity and invocation, but when sla plugin is extended with more arguments, a better way to invoke this plugin may be job plugin like svc and ssh.