plotter/k8s-job-dispatcher repository overview

⁠K8s job dispatcher

This is a minimal API used to dispatch prepared kubernetes jobs. It's an experiment in using jq⁠ as a configuration language.

The exposed API transforms requests into kubernetes API requests for creating and retrieving jobs, and the transformation is executed using jq filters and the jaq⁠ library. The jq filters can be configured by the user, giving them freedom to interpret the requests and assemble the job manifests.

You can probably achieve the same behaviour using OpenResty's ngx_http_lua_module⁠, for example using rewrite_by_lua_block, and putting some effort into implementing the kubernetes authentication and service account discovery (this repo skips all of that thanks to the kube⁠ library).

⁠Synopsis

Job-dispatching interface acting as a thin wrapper over K8s API

Usage: k8s-job-dispatcher [OPTIONS] [FILTER]

Arguments:
  [FILTER]  Filter converting requests to K8s manifests

Options:
  -f, --from-file <FROM_FILE>  Read filter from a file [env: FROM_FILE=]
  -p, --port <PORT>            TCP port to listen on [env: PORT=] [default: 8000]
      --log-level <LOG_LEVEL>  Log level [env: LOG_LEVEL=] [default: INFO]
  -h, --help                   Print help
  -V, --version                Print version

⁠Additional motivation

Combined with a proper job queue such as Kueue⁠ and/or a cluster autoscaler⁠, this could be a viable strategy to schedule demand-dependent amounts of jobs, forming the basis of an asynchronous execution mesh.

⁠Concurrency control using the scheduling queue

One simple way of controlling maximum concurrency is to use the ResourceQuota⁠ object with pod limits, and dedicate a whole namespace to jobs. Alternatively you can configure pod priority classes and set the resource quota with suitable scope selectors.

When using resource quotas, a newly created job won't generate a pod that could exceed the limit imposed. Only after enough pods have terminated will the job be able to schedule its pod.