Start runs without a Polyaxonfile

Start a Python version check, try a package update, or open a sandbox directly from your terminal. Since Polyaxon 2.18, polyaxon run accepts the image, command, environment variables and resources inline.

The examples use a configured, authenticated CLI and a Polyaxon deployment supporting 2.18. Replace acme/demo with an existing project you can run workloads in.

Try an image

Check which Python version an image provides:

polyaxon run -p acme/demo \
  --image python:3.12-slim \
  --cmd 'python -V' \
  --logs

Polyaxon schedules a job, streams its logs, and records its status. The job exits after printing the version. Change the image to python:3.13-slim and submit again to check that environment; each submission creates a separate run.

Try a dependency or environment change

Install a package at startup and print the resulting package versions:

polyaxon run -p acme/demo \
  --image python:3.12-slim \
  --env PACKAGE_SPEC=rich \
  --cmd 'python -m pip install "$PACKAGE_SPEC"' \
  --cmd 'python -m pip freeze' \
  --logs

Change PACKAGE_SPEC to a version constraint, such as rich==13.9.4, to try a particular release. You can repeat --env for other environment variables. The single quotes around the command leave $PACKAGE_SPEC for the remote shell to expand.

Repeated --cmd options run in one /bin/sh script with set -e: if installation fails, the package listing does not run. The image needs that shell and access to the package index. The install changes this run's container; it does not rebuild the image.

Try changes to an existing job

If you already have a job in train.yaml, pass it explicitly and override just the settings you want to try. Replace the example registry address with your candidate training image, containing the dependencies that job needs:

polyaxon run -p acme/demo -f train.yaml \
  --image registry.example.com/team/training:candidate \
  --env LOG_LEVEL=debug \
  --cpu 1..2 \
  --memory 2Gi..4Gi \
  --logs

Inline options apply after the files. This submits a new run with the overrides and leaves train.yaml unchanged.

Keep -f when overriding a file. With --image and no explicit source, Polyaxon defines the run inline and skips the directory's default Polyaxonfile. A bare polyaxon run --image … therefore starts a new inline workload, even inside an initialized project containing a Polyaxonfile.

The same command can select resources and a compute queue:

Quick start a sandbox

For an environment you can keep using between commands, start a sandbox service:

polyaxon run -p acme/demo --sandbox --image python:3.12-slim

Your deployment must support sandbox access. Wait for the run to reach running, then copy its UUID into your terminal:

RUN_UUID=PASTE_RUN_UUID_HERE

polyaxon sandbox ping -p acme/demo -uid "$RUN_UUID"
polyaxon sandbox exec -p acme/demo -uid "$RUN_UUID" -- python -V

ping checks the endpoint; it does not wait for startup. Now try a package inside that same running container:

polyaxon sandbox exec -p acme/demo -uid "$RUN_UUID" \
  --stream -- python -m pip install rich

Or open an interactive shell:

polyaxon sandbox shell -p acme/demo -uid "$RUN_UUID"

These commands operate inside the existing service. Package changes remain local to that container; a replacement starts from its image and declared setup. To try a different base image, submit another polyaxon run --sandbox --image ….

Exiting the shell leaves the service running. Stop it when you finish:

polyaxon ops stop -p acme/demo -uid "$RUN_UUID"

When an experiment becomes something you want to repeat or share, put the image and startup commands into a short Polyaxonfile or your team's preset. The run CLI reference lists the inline options; the sandbox CLI reference covers file transfers, background commands and terminal access.