Working with the runtime

The runtime is the collection of omegaml workers that wait for models, jobs or scripts to be run. The runtime works fully asynchronously by receiving, processing and returning messages.


Technically the omegaml runtime is a Celery cluster made up of 1 to many worker nodes. Every worker node runs a pool of 1 to many Celery processes. Think of a worker as a pre-started and pre-configured Python environment waiting for work.

$ om help runtime

  om runtime model <name> <model-action> [<X>] [<Y>] [--result=<output-name>] [--param=<kw=value>]... [--async] [options]
  om runtime script <name> [<script-action>] [<kw=value>...] [--async] [options]
  om runtime job <name> [<job-action>] [<args...>] [--async] [options]
  om runtime result <taskid> [options]
  om runtime ping [options]
  om runtime env <action> [<package>] [--file <requirements.txt>] [--every] [options]
  om runtime log [-f] [options]
  om runtime status [workers|labels|stats] [options]
  om runtime restart app <name> [options]
  om runtime [control|inspect|celery] [<celery-command>...] [--worker=<worker>] [--queue=<queue>] [--celery-help] [--flags <celery-flags>...] [options]

  --async           don't wait for results, will print taskid
  -f                tail log
  --require=VALUE   worker label
  --flags=VALUE     celery flags, list as "--flag VALUE"
  --worker=VALUE    celery worker
  --queue=VALUE     celery queue
  --celery-help     show celery help
  --file=VALUE      path/to/requirements.txt
  --local           if specified the task will run locally. Use this for testing
  --every           if specified runs task on all workers

  model, job and script commands

  <model-action> can be any valid model action like fit, predict, score,
  transform, decision_function etc.

  <script-action> defaults to run
  <job-action> defaults to run

    om runtime model <name> fit <X> <Y>
    om runtime model <name> predict <X>
    om runtime job <name>
    om runtime script <name>
    om runtime script <name> run myparam="value"

  running asynchronously

  model, job, script commands accept the --async paramter. This will submit
  the a task and return the task id. To wait for and get the result run use
  the result command

        om runtime model <name> fit <X> <Y> --async
        => <task id>
        om runtime result <task id>
        => result of the task

  restart app

  This will restart the app on omegaml apphub. Requires a login to omegaml cloud.


  Prints workers, labels, list of active tasks per worker, count of tasks

    om runtime status             # defaults to workers
    om runtime status workers
    om runtime status labels
    om runtime status stats

  celery commands

  This is the same as calling celery -A omegaml.celeryapp <commands>. Command
  commands include:

  inspect active         show currently running tasks
  inspect active_queues  show active queues for each worker
  inspect stats          show stats of each worker, including pool size (processes)
  inspect ping           confirm that worker is connected

  control pool_shrink N  shrink worker pool by N, specify 99 to remove all
  control pool_grow N    grow worker poool by N
  control shutdown       stop and restart the worker

        om runtime celery inspect active
        om runtime celery control pool_grow N

  env commands

  This talks to an omegaml worker's pip environment

  a) install a specific package

     env install <package>    install the specified package, use name==version pip syntax for specific versions
     env uninstall <package>  uninstall the specified package

     <package> is in pip install syntax, e.g.

     env install "six==1.0.0"
     env install "git+"

  b) use a requirements file

     env install --file requirements.txt
     env uninstall --file requirements.txt

  c) list currently installed packages

     env freeze
     env list

  d) install on all or a specific worker

     env install --require gpu package
     env install --every package

     By default the installation runs on the default worker only. If there are multiple nodes where you
     want to install the package(s) worker nodes, be sure to specify --every

        om runtime env install pandas
        om runtime env uninstall pandas
        om runtime env install --file requirements.txt
        om runtime env install --file gpu-requirements.txt --require gpu
        om runtime env install --file requirements.txt --every

Inspect runtime status

$ om runtime status
$ om runtime status labels
$ om runtime status stats

Monitor live events

Launch the Celery Events monitor as follows

$ om runtime celery events
$ om runtime celery flower

See active work items

$ om runtime inspect active
  -> celery@worker-worker-omdemo: OK
      - empty -

Ping the worker

$ om runtime ping
{'message': 'ping return message', 'time': '2021-02-22T13:27:49.723657', 'args': (), 'kwargs': {}, 'worker': 'celery@worker-worker-omdemo'}

Restart the worker

$ om runtime control shutdown

Install new packages

$ om runtime env install "git+" --require default

Note it is best to restart the worker to make sure the newest package versions are loaded. For example to upgrade omegaml itself:

$ om runtime env install omegaml
$ om runtime control shutdown

To upgrade a package on every worker, specify --every. This will send the installation command to all workers.

$ om runtime env install omegaml --every