Aiperf Kube Triage

Triaging an AIPerf Kubernetes Run

Classify first, then follow the branch. Do not read logs before you know the phase — the phase determines which logs matter.

Related skills: aiperf-kube-run (lifecycle and exit codes), aiperf-kube-setup (cluster/operator install problems).

The machine-parseable playbook with full JSON schemas and jq/python snippets is references/debugging.md (bundled with this skill). This skill is the routing layer.

Step 1: classify

kubectl get aiperfjob <NAME> -n <NS> -o json | jq '{
  phase: .status.phase, subPhase: .status.subPhase,
  workers: .status.workers, error: .status.error,
  conditions: .status.conditions }'

aiperf kube debug -j <NAME> -n <NS> --verbose

aiperf kube debug reports pod states, recent events, node resources, and a slice of logs from any pod with problems. -A inspects all namespaces. For a sweep child, pass --variation <idx> (long form only — -v is --verbose here) and -t <trial>.

Step 2: branch on status.phase

request_error_rate.avg in status.liveMetrics.metrics is the only error signal published while a job runs — error_request_count is ERROR_ONLY and is filtered out of live metrics. Do not conclude "no errors" from its absence.

Step 3: failure patterns

Scheduling messages map directly to causes: Insufficient cpu/memory -> lower --total-workers; nvidia.com/gpu -> no free GPU nodes; didn't match Pod's node affinity/selector -> wrong spec.podTemplate.nodeSelector; had untolerated taint -> missing spec.podTemplate.tolerations; quota -> namespace ResourceQuota exhausted.

None of those are patchable on the live CR. Nineteen spec fields are immutable after creation — image, imagePullPolicy, benchmark, resourceMode, connectionsPerWorker, ttlSecondsAfterFinished, resultsTtlDays, keepFailedPods, podTemplate, scheduling, schemaVersion, sweep, multiRun, plot, variables, randomSeed, noSweepTable, skipEndpointCheck, failurePolicy. Only timeoutSeconds and cancel accept an update; everything else rejects the patch with spec.<field> is immutable after creation. So every scheduling, placement, resource, priority, or workload fix means delete and resubmit, not kubectl edit.

Rules that bite

• A benchmark that looks hung at high concurrency is often CPU starvation of a single service, not a deadlock. The records-manager and system-controller default to 75m CPU (AIPERF_K8S_RECORDS_MANAGER_CPU, AIPERF_K8S_SYSTEM_CONTROLLER_CPU). At very high request rates a pegged core starves the event loop, heartbeats expire, and the CR freezes mid-phase. Raise the CPU before hunting for a logic bug.
• The default resourceMode is burstable, so pods have requests and no limits. spec.resourceMode defaults to burstable and emits requests only. That means a container cannot be cgroup-OOM-killed for exceeding its memory budget and cannot be CFS-throttled for exceeding its CPU budget — the 75m/150m numbers are scheduling hints, not ceilings. Two consequences: an OOMKilled container is evidence of node memory pressure or kubelet eviction, not of a too-small AIPERF_K8S_*_MEMORY; and a starving service is losing a fair-share race against co-tenants, which raising the request fixes only indirectly (by moving the pod or reserving more share). Set resourceMode: guaranteed if you actually want the env budgets enforced as limits, and expect real OOM kills once you do.
• Heartbeat expiry is a symptom, not a cause. AIPERF_K8S_CONTROLLER_HEARTBEAT_EXPIRY_SECONDS defaults to 30 s and must be at least twice the interval; widening it hides the starvation instead of fixing it.
• Workers pass two startup probes, and they fail differently. The PUB/SUB connection probe fails closed: the pod exits on purpose so Kubernetes restarts it with a fresh ZMQ context (AIPERF_K8S_JOBSET_WORKER_CONNECTION_PROBE_TIMEOUT, backoff limit 20). A few early restarts on first deploy are expected, not a bug. The credit-return PUSH probe fails open: credit dispatch rides the DEALER but returns ride a separate PUSH/PULL fan-in, and on budget expiry (AIPERF_WORKER_RETURN_PROBE_BUDGET, default 30 s; AIPERF_WORKER_RETURN_PROBE_RETRY_DELAY, default 0.1 s) the worker announces dispatchability anyway and logs a warning. So a dead return path produces a stalled run with zero restarts -- grep worker logs for Credit-return channel still has no peer before blaming the endpoint.
• Operator-mode results are harvested only on terminal phase. "results missing" mid-run is correct behavior; use --from-pods to read the live controller instead.
• On a successful run the pods do not last 300 s — they are deleted at once. AIPERF_K8S_JOBSET_TTL_SECONDS_AFTER_FINISHED (300 s) is only the fallback: once the operator has harvested results it calls its post-success cleanup and deletes the JobSet immediately. The TTL window is real only for failed runs and for partial harvests that are being retried. Never plan to "grab the logs after it finishes" — stream them with --follow during the run, or accept that a clean success leaves you only the PVC artifacts. Direct (operator-less) mode is the opposite: AIPERF_K8S_JOBSET_DIRECT_MODE_TTL_SECONDS defaults to 28800 s (8 h) so pods stay around for manual retrieval.
• debug, cancel, delete, list exit 0 on a missing target by design. Never gate a script on their exit status; use attach or logs.
• First deploy attempts flake more than steady state (ConfigMap propagation, ZMQ probe). One clean retry before deep investigation is a legitimate step.

Diagnosis thresholds

aiperf kube debug findings are tunable via AIPERF_K8S_DIAGNOSIS_*: STALLED_PENDING_THRESHOLD_SECONDS (60), STALLED_RUNNING_THRESHOLD_SECONDS (30), HIGH_ERROR_RATE_THRESHOLD (0.05), FAIL_ABOVE_ERROR_RATE (1.0), HIGH_LATENCY_P99_MULTIPLIER (10). Full env-var table in references/debugging.md.

Log collection

aiperf kube logs <ID> -n <NS> -o ./triage --ignore-not-found   # per-pod files
aiperf kube logs <ID> --container control-plane --follow

Containers worth naming: control-plane (SystemController), api, event-bus-proxy, worker containers. Save logs to a directory before the TTL reaps the pods.