Skip to content

A run is a one-shot Job

What you notice: you install nothing in the cluster before your first run. dokimi submit creates a Kubernetes Job, streams its log, and the Job ends when the run does. Pressing Ctrl+C stops the stream, not the run, and dokimi attach picks it up again, from the same machine or another one.

Why

A test should not need a permanent resident. A controller or an operator would have to be installed, upgraded and given standing permissions in every cluster you test, before it has measured anything. A Job exists only while it runs, with only the permissions its service account grants.

The run is self-contained. Your test definition travels with the run in a ConfigMap, and the images it runs are chosen when you submit and written into the Job. The Job needs nothing from the machine that submitted it, which is why attach works from anywhere and why a laptop going to sleep never interrupts a test.

It bounds itself. Every Job and every test pod carries a deadline derived from the test's own duration, so a run that is abandoned or crashes stops on its own instead of sending traffic until someone notices.

Its exit code is the verdict. A Job's exit code is something every CI system already reads. dokimi's codes are a fixed table, and the Job never retries: a failed test that ran again would replace a finding with a second measurement.

What it costs

  • What a run leaves behind is deliberate. When it ends, the run removes the node-exporter it deployed. A run that passed removes its test pods too. One that didn't keeps its owner, its definition and the pods of tests that were measured or never started, for inspection, until dokimi cleanup. Its Prometheus stays, so its history can be queried later with dokimi prometheus <run-id>. Removing it is one kubectl delete the run prints for you, or dokimi cleanup <run-id> --prometheus.
  • Reports come back through the log. A Job has no file server, so the run writes its reports into its log at the end, and submit and attach extract them.
  • One process per run. There is nothing to share between runs, which is also why a cluster runs one at a time.