dispatch-admission¶

dispatch-admission is the Kueue AdmissionCheck controller for dispatch's readmission gate. It runs in the cluster, beside the ClusterQueue configuration that names its AdmissionCheck. Install the admission controller deploys it; the start gate and the readmission gate says what it is for.

dispatch-admission [flags]

Go's flag package parses the flags, so -name and --name both work, and booleans take -name or -name=false.

Flags¶

The adapter's kueue.admissionLease setting must name the same Lease as -lease-namespace and -lease-name. Its default is the same pair as the flags' defaults.

What it does¶

1. Leader election on a Lease, using client-go's leaderelection. Replicas that do not hold it wait as candidates. The Lease is not released on shutdown.
2. A full pass, on becoming leader and then once per -resync:
3. finds every AdmissionCheck whose spec.controllerName is dispatch.phpboyscout.uk/admission, whatever it is called, and sets its Active condition True (reason ControllerStarted) if it is not. It never sets False;
4. lists every Workload and reconciles it.
5. A watch on Workloads for the rest of the interval, reconciling each Workload as it changes. DELETED events are ignored. A failed watch waits out the interval before the next pass.

The checks the controller serves are those it found at its last full pass that completed the first step. An AdmissionCheck created later is served from the next pass, up to -resync later. If the first step fails, for example because the controller may not update admissionchecks/status, it logs full pass at error level and skips that pass's list of Workloads; the watch still runs, for the checks found before.

To reconcile a Workload, the controller acts only when the Workload has quota reserved (QuotaReserved=True) and has a check of ours whose state is Pending. Then:

A TrainJob is dispatch's when its name matches dj-<ULID>-s<N> or it has a runtime patch whose manager is dispatch.phpboyscout.uk/adapter. A Workload's TrainJob is the one named by its owner reference, whose UID must match.

Every write to the Workload is preconditioned on the Workload as read, and the TrainJob update on the TrainJob as read. A conflict is logged at info level and the Workload is reconciled again when it next changes.

Logging and exit¶

• Logs go to standard error as log/slog text lines. leading and stopped leading are logged with the identity; raised the admission count with the TrainJob and the new A; failures at error level.
• On SIGINT or SIGTERM it stops and exits 0.
• On losing the Lease, or if its reconciliation loop ends, it exits 1 with dispatch-admission: lost the leader-election Lease, so that its Deployment restarts it as a candidate.
• A flag error, or a conflicting -in-cluster, exits non-zero with a message on standard error.

Verbs it uses¶

The ClusterRole in the install guide, which the test cluster uses, grants a little more than this.