Fournos (φούρνος) = "oven" in Greek.
Fournos is a Kubernetes operator that schedules benchmark jobs via Kueue and executes them as Tekton PipelineRuns on remote clusters through a pluggable execution engine.
Jobs are submitted as FournosJob custom resources. Every job first
passes through a mandatory Resolving phase where a resolve Job (driven
by the configured execution engine) populates GPU requirements and secret
references directly on the FournosJob spec. The operator then creates a
Kueue Workload for quota management, waits for admission, and launches the
corresponding Tekton PipelineRun.
The following operators must be installed in the cluster before deploying Fournos:
- Red Hat OpenShift Pipelines (
1.21) - Red Hat build of Kueue (
1.3) - Builds for Red Hat OpenShift Operator (
1.7) - Red Hat OpenShift GitOps (
1.20)- only for the GitOps deployment of Fournos
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
pre-commit installCreate a FournosJob resource. Use generateName for automatic unique naming
and displayName for a human-readable label:
apiVersion: fournos.dev/v1
kind: FournosJob
metadata:
generateName: sample-run-benchamark-
spec:
owner: perf-team
displayName: sample-run-benchmark
cluster: cluster-1
hardware:
gpuType: a100
gpuCount: 2
pipeline: forge-full
executionEngine:
forge:
project: llmd
args:
- cks
configOverrides:
batch_size: 64
env:
OCPCI_SUITE: regression
OCPCI_VARIANT: nightlyFOURNOS_NAMESPACE=fournos-$USER-dev
oc create -f config/forge/samples/job-full.yaml -n $FOURNOS_NAMESPACE # returns the generated name, e.g. forge-full-sample-x7k2m
oc get FournosJobs -n $FOURNOS_NAMESPACE -w # watch status transitions
oc patch FournosJob <name> -n $FOURNOS_NAMESPACE --type merge -p '{"spec":{"shutdown":"Stop"}}' # graceful stop (runs finally tasks)
oc patch FournosJob <name> -n $FOURNOS_NAMESPACE --type merge -p '{"spec":{"shutdown":"Terminate"}}' # immediate terminate (skips finally tasks)
oc delete FournosJob -n $FOURNOS_NAMESPACE <name> # cleanup| Field | Required | Description |
|---|---|---|
spec.executionEngine |
yes | Execution engine configuration. The single top-level key is the engine name (e.g. forge); its value is opaque engine-specific config passed through as-is. |
spec.env |
no | Environment variables available to the execution engine (read from the FournosJob spec via K8s API) |
spec.cluster |
* | Pin to a specific cluster (Kueue ResourceFlavor). Since exclusive defaults to true, this also locks the cluster — set exclusive: false for shared access. |
spec.hardware.gpuType |
* | Short GPU model name — e.g. a100, h200. The operator prepends the FOURNOS_GPU_RESOURCE_PREFIX (default fournos/gpu-) automatically, so do not include the full resource path. |
spec.hardware.gpuCount |
with gpuType | Number of GPUs (minimum 1) |
spec.owner |
no | Team or individual that owns this job |
spec.displayName |
no | Human-readable job name (defaults to metadata.name) |
spec.pipeline |
no | Tekton Pipeline name (default: fournos-full). The Pipeline must carry a fournos.dev/resolve-image annotation with the full image reference for the resolve Job. |
spec.priority |
no | Kueue WorkloadPriorityClass name |
spec.secretRefs |
no | Vault-synced K8s Secret names (prefixed with vault-) to mount into the pipeline. Populated by the execution engine during the Resolving phase. The operator validates each name in FOURNOS_SECRETS_NAMESPACE, copies the secrets into the operator namespace, and mounts them as a projected volume at /var/run/secrets/fournos/<entry-name>/. |
spec.exclusive |
no (default true) |
If true, locks the target cluster so no other FournosJob can run there. Requires spec.cluster. Hardware is optional — when omitted the Workload only requests cluster-slot resources for locking. |
spec.shutdown |
no | Shutdown action: Stop cancels gracefully (Tekton CancelledRunFinally — runs finally tasks); Terminate cancels immediately (Tekton Cancelled — skips finally tasks). Both wait for the PipelineRun to finish before releasing Kueue quota. |
* spec.hardware is required unless the job uses exclusive cluster locking
(exclusive: true + cluster), in which case it may be omitted — the
Workload only needs cluster-slot resources. Every job passes through the
Resolving phase where the execution engine populates spec.hardware (if
not already set) and spec.secretRefs directly on the FournosJob. Since exclusive defaults
to true, any job with spec.cluster locks the cluster exclusively —
including jobs that also specify spec.hardware. Set exclusive: false for
shared access (hardware is then required). Jobs without spec.cluster must
set exclusive: false.
The operator writes status to .status:
| Field | Description |
|---|---|
phase |
Resolving → Pending → Admitted → Running → Succeeded / Failed / Stopping → Stopped |
cluster |
Cluster assigned by Kueue |
pipelineRun |
Name of the Tekton PipelineRun |
dashboardURL |
Tekton Dashboard link (if configured) |
message |
Error details on failure |
Prerequisites: Podman,
kind, and kubectl.
make dev-setup # creates a kind cluster, installs Tekton + Kueue + CRD, applies mock resources
make dev-run # starts the operator locally (connects to the kind cluster)Both targets default to the fournos-local-dev namespace. Override with
FOURNOS_NAMESPACE=<YOUR_NAMESPACE> make dev-setup dev-run.
In another terminal:
FOURNOS_NAMESPACE=fournos-local-dev make test # run the integration test suitemake dev-teardown # deletes the kind clusterdev-setup installs real Tekton Pipelines and Kueue controllers into the kind
cluster, but substitutes lightweight mock Tasks (echo + sleep) in place of the
real execution engine runner. The dev environment uses its own Kueue config
(dev/mock-kueue-config.yaml) with four mock clusters and synthetic GPU quotas,
plus matching kubeconfig Secrets (kubeconfig-cluster-{1..4}) in the dedicated
secrets namespace (psap-secrets).
make lint # lint (fournos/ + tests/)
make test # integration tests (operator must be running)Execution engine on the hub: config/forge/ is the real OpenShift configuration for this repo — ImageStreams, Builds, Tekton Tasks and Pipelines, and sample jobs you apply to a cluster. It is not the same as the lightweight stand-ins under dev/mock-pipelines/, which make dev-setup installs on kind for local testing only.
Prepare the namespaces
FOURNOS_NAMESPACE=fournos-$USER-dev
FOURNOS_SECRETS_NAMESPACE=psap-secrets
oc create ns $FOURNOS_NAMESPACE
oc label ns/$FOURNOS_NAMESPACE fournos.dev/queue-access=true
oc create ns $FOURNOS_SECRETS_NAMESPACEDeploy the operator:
oc apply -n $FOURNOS_NAMESPACE -f manifests/crd.yaml
for rbac_file in manifests/rbac/*.yaml; do
cat $rbac_file | NAMESPACE=$FOURNOS_NAMESPACE envsubst | oc apply -f- -n $FOURNOS_NAMESPACE
done
cat manifests/secrets-ns-rbac.yaml \
| NAMESPACE=$FOURNOS_NAMESPACE SECRETS_NAMESPACE=$FOURNOS_SECRETS_NAMESPACE envsubst \
| oc apply -f-
oc apply -n $FOURNOS_NAMESPACE -f manifests/deployment.yamlThree things are needed to make a target cluster available to Fournos:
- Create a kubeconfig Secret in the dedicated secrets namespace:
FOURNOS_SECRETS_NAMESPACE=psap-secrets
CLUSTER_NAME=<name>
oc create secret generic kubeconfig-${CLUSTER_NAME} \
--from-file=kubeconfig=/path/to/auth/kubeconfig \
-n $FOURNOS_SECRETS_NAMESPACEThe secret name must match the FOURNOS_KUBECONFIG_SECRET_PATTERN (default
kubeconfig-{cluster}). Secrets are stored in the dedicated namespace
(FOURNOS_SECRETS_NAMESPACE, default psap-secrets).
- Add a ResourceFlavor and quota in
config/kueue-config.yaml. Add a newResourceFlavorwith a matchingfournos.dev/clusternodeLabel, and list it under thefournos-queueClusterQueue with the appropriate GPU/CPU quotas. Then apply:
oc apply -f config/kueue-config.yaml- Verify connectivity by submitting a lightweight validate-only
job. Edit
cluster(and optionallyhardware) inconfig/fournos-validation/samples/test-connectivity-job.yamlto match the new target, then:
FOURNOS_NAMESPACE=fournos-$USER-dev
oc create -f config/fournos-validation/samples/test-connectivity-job.yaml -n $FOURNOS_NAMESPACE
oc get fournosjobs -n $FOURNOS_NAMESPACE -w # should reach SucceededThis runs the fournos-validate-only pipeline, which only checks oc cluster-info against the target — no benchmark workload is launched. If the job
reaches Succeeded, the kubeconfig secret and Kueue quota are correctly
configured. If it fails, check the operator logs and the PipelineRun status for
details.
Apply the production execution engine assets from config/forge/ (not the kind mocks in dev/mock-pipelines/). Deploy the cluster configuration (Builds + Tekton):
oc apply -n $FOURNOS_NAMESPACE -f config/forge/images/is_forge.yaml
cat config/forge/images/build_forge-main.yaml \
| sed 's/psap-automation/'$FOURNOS_NAMESPACE'/g' \
| oc apply -n $FOURNOS_NAMESPACE
oc create -n $FOURNOS_NAMESPACE -f config/forge/images/buildrun_forge-main.yaml
for wf_file in config/forge/workflows/*.yaml; do
cat "$wf_file" | NAMESPACE=$FOURNOS_NAMESPACE envsubst '$NAMESPACE' | oc apply -f- -n $FOURNOS_NAMESPACE
donePipeline jobs can reference Kubernetes Secrets via spec.secretRefs. These
secrets originate in a HashiCorp Vault instance. Because there is no permanent
programmatic access to the vault, secrets are synchronized manually on demand —
whenever the vault content changes.
The sync script reads vault entries and creates one Opaque Secret per entry
in the dedicated secrets namespace (FOURNOS_SECRETS_NAMESPACE, default
psap-secrets), using a vault- prefix followed by the vault entry name as
the K8s Secret name (e.g. vault entry my-creds becomes Secret
vault-my-creds). Entries whose names are not valid DNS-1123 subdomain
names are skipped with an error. Individual keys within an entry that are
not valid K8s Secret data keys (allowed: alphanumeric, -, _, .) are
also skipped. Existing secrets are updated in-place.
# 1. Set the required environment variables
export VAULT_ADDR="https://vault.example.com" # Vault server URL
export VAULT_TOKEN="s.xxxxx" # your short-lived token
export VAULT_SECRET_PATH="path/to/secrets" # directory path within the KV engine
# 2. Sync all vault entries under the configured path
python hacks/sync_vault_secrets.py -n psap-secrets
# 3. Preview without touching the cluster
python hacks/sync_vault_secrets.py -n psap-secrets --dry-runMakefile shortcuts (VAULT_ADDR, VAULT_TOKEN, and VAULT_SECRET_PATH must be set):
make sync-vault-secrets # syncs all entries
make sync-vault-secrets-dry-run # preview onlyThe synced secrets are labelled fournos.dev/vault-entry=true and
app.kubernetes.io/managed-by=fournos-vault-sync for easy identification.
Secret references are populated by the execution engine during the Resolving phase directly
on the FournosJob spec.secretRefs field. The operator validates each
referenced Secret exists in the secrets namespace and carries the vault
label during the Resolving phase, then copies them into the operator
namespace during the Admitted phase and mounts them as a projected volume
into the PipelineRun pods. Each secret's keys are placed under a
subdirectory matching the original name:
/var/run/secrets/fournos/
vault-my-creds/
username
password
vault-other-creds/
token
Copied secrets are named <fjob-name>-<secret-name> and carry
ownerReferences back to the FournosJob, so Kubernetes garbage-collects
them automatically when the job is deleted.
All settings are read from environment variables with the FOURNOS_ prefix:
| Variable | Default | Description |
|---|---|---|
FOURNOS_NAMESPACE |
required | Kubernetes namespace |
FOURNOS_SECRETS_NAMESPACE |
psap-secrets |
Namespace where kubeconfig and vault-synced secrets are stored |
FOURNOS_TEKTON_DASHBOARD_URL |
Tekton Dashboard base URL | |
FOURNOS_KUBECONFIG_SECRET_PATTERN |
kubeconfig-{cluster} |
Pattern for resolving cluster names to Secret names |
FOURNOS_VAULT_SECRET_PATTERN |
vault-{entry} |
Pattern for naming vault-synced Secrets |
FOURNOS_KUEUE_LOCAL_QUEUE_NAME |
fournos-queue |
Kueue LocalQueue name |
FOURNOS_GPU_RESOURCE_PREFIX |
fournos/gpu- |
Resource name prefix for GPU types |
FOURNOS_LOG_LEVEL |
INFO |
Logging level |
FOURNOS_GC_INTERVAL_SEC |
300 |
Resource GC interval (seconds) |
FOURNOS_RESOLVE_DEADLINE_SEC |
300 |
Deadline for the resolve Job (seconds) |
FOURNOS_RESOLVE_JOB_TEMPLATE |
config/forge/resolve_job.yaml |
Path (relative to project root) to the Job YAML template for the resolve step. Override with dev/mock-resolve/resolve_job.yaml for local dev/CI. |
FOURNOS_ARTIFACT_PVC_SIZE |
1Gi |
Size of the per-PipelineRun PVC used for shared artifact storage across pipeline tasks |
FournosJob CR ──→ Operator ──→ Resolve Job (e.g. FORGE, patches FournosJob spec) ──→ Kueue Workload ──→ (admission) ──→ Tekton PipelineRun ──→ Execution Engine (e.g. FORGE) ──→ target cluster
The operator runs as a single-replica Deployment using
kopf. On each FournosJob, it:
- Resolves job requirements by launching a resolve K8s Job (using the configured execution engine image) that populates the FournosJob spec with GPU type/count and secret references
- Creates a Kueue Workload with the resolved GPU resources (owned by the FournosJob via
ownerReferences) - Polls (5 s timer) for Kueue admission and assigned cluster
- Copies referenced Vault secrets from the secrets namespace into the operator namespace (per-job copies with
ownerReferencesfor automatic cleanup) and launches a Tekton PipelineRun withFJOB_NAME+FOURNOS_NAMESPACE(so the execution engine can look up the full FournosJob spec), the secrets mounted as a projected volume at/var/run/secrets/fournos/(owned by the FournosJob viaownerReferences), and a sharedartifactsworkspace backed by avolumeClaimTemplatePVC for cross-task artifact storage (managed by Tekton) - Watches the PipelineRun until completion
- Deletes the Workload to release Kueue quota
Setting spec.shutdown on a FournosJob triggers cancellation of the
PipelineRun and transitions to phase=Stopping. Stop uses Tekton's
CancelledRunFinally (runs finally cleanup tasks); Terminate uses
Cancelled (skips finally tasks). In both cases the operator keeps
the Kueue Workload alive until the PipelineRun finishes, ensuring the
cluster slot is not released prematurely. Once done, the Workload is
deleted and the job moves to phase=Stopped.
Deleting a FournosJob automatically cascade-deletes its Workload and PipelineRun through Kubernetes owner references.
Target clusters need nothing installed — the execution engine runs on the hub
cluster inside Tekton Task pods and communicates with targets via
oc/kubectl through kubeconfig Secrets.
For a detailed breakdown of the CRD, scheduling, operator internals, and key design decisions, see the Design Document.