This version is under construction, please use an official release version

The Operator

The operator path is to describe the run in a ConfigMaxRun resource, apply it, watch it, and read the result. Prerequisites are listed on the Running ConfigMax page.

1. Describe the run

Create a file my-run.yaml. This example discovers the subnets-per-tenant ceiling and stops when the cluster shows strain, when a one-hour timeout elapses, or when 10,000 subnets are reached, whichever occurs first:

apiVersion: configmax.kubermatic.io/v1alpha1
kind: ConfigMaxRun
metadata:
  name: my-run
  namespace: configmax
spec:
  image: "quay.io/kubermatic/configmax:latest"   # benchmark image
  profile: custom                                 # custom applies the fields below
  tests:
    - subnetsPerVPC                               # the test or tests to run
  overrides:
    subnetsPerVPC:
      mode: discovery                             # discovery | target | workload-slo
      discovery:
        initialBatch: 100                         # objects in the first batch
        growthFactor: "1.5"                        # batch size multiplied by 1.5 each round
        batchPauseSec: 5                          # settle time between batches
        hardTimeoutSec: 3600                      # wall-clock cap (one hour)
        safetyBound: 10000                        # absolute upper limit
      distressOverrides:
        sampleIntervalSec: 30                     # how often to check for strain
        consecutiveSamples: 3                     # trips required before stopping
  settings:
    cleanupAfterRun: true                         # delete test objects when done
    stopOnFirstError: false
  ttlAfterFinished: 259200                        # keep the result for three days, then GC

The two most significant fields are mode, which determines how the run decides to stop, and discovery, which determines how aggressively the run applies load.

2. Apply and monitor the run

kubectl apply -f my-run.yaml

# high-level progress
kubectl -n configmax get configmaxrun my-run

# detailed status (phase, current test, messages)
kubectl -n configmax describe configmaxrun my-run

The get output shows the profile, the phase (Pending, Provisioning, Running, Completed), and a progress summary.

3. Read the result

Once the phase is Completed, the primary figures are available in .status.results:

kubectl -n configmax get configmaxrun my-run \
  -o jsonpath='{range .status.results[*]}{.test}{": "}{.highestCount}{" (stopped by "}{.stoppedBy}{")\n"}{end}'

Example output:

subnetsPerVPC: 8001 (stopped by safety-bound)

highestCount is the discovered maximum, and stoppedBy states why the run stopped: a distress signal, timeout, or safety-bound (the configured cap). The full per-batch curve and cluster snapshots are available under .status.report.

Key parameters

These are the fields changed most often. All live under spec.

Field	Values	Purpose
`profile`	`smoke` · `medium` · `full` · `custom`	Preset scale. Use `custom` to control the fields below.
`tests`	list of test IDs	Which benchmarks to run (for example `subnetsPerVPC`, `vpcsPerCluster`, `networkPoliciesPerNamespace`, `nicsPerVM`). Honored only with `profile: custom`.
`overrides.<test>.mode`	`discovery` · `target` · `workload-slo`	How the test decides to stop.
`discovery.initialBatch`	integer (default 10)	Objects created in the first batch.
`discovery.growthFactor`	string float (default `"2.0"`)	Batch-size multiplier each round. A lower value produces a more gradual ramp.
`discovery.batchPauseSec`	integer (default 5)	Settle time between batches; raise it to let the controller catch up.
`discovery.hardTimeoutSec`	integer (default 14400)	Wall-clock cap on the whole run.
`discovery.safetyBound`	integer (default 100000)	Absolute object-count cap; stops the run even with no distress.
`discovery.targetCount`	integer	For `mode: target` only: the count to reach for a pass.
`distressOverrides.*`	integers	Tune the strain thresholds and sampling cadence.
`settings.cleanupAfterRun`	bool (default true)	Delete benchmark objects when the run finishes.
`settings.stopOnFirstError`	bool (default false)	Stop a test at the first create error instead of probing for the breaking point.

Profiles at a glance

Profile	Scale	Typical duration	Use for
`smoke`	minimal counts, every test	about 15 to 30 minutes	a post-install sanity check or a CI gate
`medium`	moderate limits	about 2 to 4 hours	a balanced capacity snapshot
`full`	large-scale, true breaking points	8 to 12 hours or more	a complete configuration-maximum sweep
`custom`	whatever you specify	varies	targeting one test or one parameter

Begin with profile: smoke to confirm that the operator and harness function end to end, then switch to a custom run that targets the specific capability of interest. Full-cluster sweeps are best scheduled to run overnight.