| name | hyperpod-incident-triage |
| description | Correlation and skip rules for SageMaker HyperPod incident triage. Keeps distinct fault types on the same instance group as separate investigations (the default correlator merges them), and prevents periodic-audit re-investigation of an unchanged cluster. Applies at the Incident Triage stage before any investigation runs. |
| metadata | {"version":"0.7.0","agent_types":["INCIDENT_TRIAGE"]} |
HyperPod incident triage
Use these rules when deciding whether a HyperPod incident should be linked
to an existing investigation, skipped, or investigated on its own. They
refine the default correlation for HyperPod's fault model. When a rule below
does not clearly apply, fall back to your normal triage judgment.
Every HyperPod incident carries a description built by the webhook bridge. Two
kinds arrive:
- Cluster events (a real HyperPod fault) — the description contains the
instance group, a
Description:, and often a FailureMessage:.
- Periodic audits (a scheduled health sweep) — the description begins with
either "Periodic audit for HyperPod cluster '' detected N issue(s)"
(issues found) or "Daily heartbeat audit for HyperPod cluster" (the once-daily
all-clear liveness signal). Both ask the RCA skill to run in audit mode.
Linking rules for HyperPod fault events
Link a new HyperPod fault only to an existing HyperPod investigation that
describes the same fault on the same component. Treat the instance group
plus the fault text (Description: together with any FailureMessage:) as the
identity of the fault.
- Do link when both the instance group and the fault text match an open
investigation — it is the same problem arriving again.
- Do NOT link when the instance group is the same but the fault text
differs. Different fault types on the same instance group have different root
causes and need separate investigations. For example, a lifecycle-script
failure on
worker2 is not the same incident as a GPU/Xid fault on
worker2 — keep them separate even though they share the instance group.
- Do NOT link the same fault type across different instance groups (e.g.
an Xid fault on
worker2 and an Xid fault on worker3 are two distinct
hardware faults).
- Do NOT link capacity errors for different instance types (e.g.
ml.g5.8xlarge vs ml.p5.48xlarge) — the FailureMessage differs, so they
are distinct problems.
When in doubt, prefer a separate investigation over an incorrect link:
over-linking hides distinct problems, which is the failure this skill exists to
prevent.
Rules for periodic audits
A periodic audit is a scheduled sweep, not a specific fault. Decide based on
whether the cluster's condition has changed since the last audit:
- Skip the daily heartbeat / all-clear audit (description begins "Daily
heartbeat audit for HyperPod cluster", or otherwise reports no open issues) —
it carries no fault to investigate; the task's presence in the console is
itself the liveness signal.
- Skip the audit if another periodic-audit investigation for the same
cluster is currently in progress — let it finish rather than starting or
re-activating a parallel one.
- Link the audit to the most recent completed periodic-audit
investigation only if the cluster's open problems are unchanged — the
same fault chains, the same CrashLoopBackOff pods, and the same NotReady
nodes as that prior audit found. There is nothing new to report.
- Proceed (investigate) if the cluster shows any new or changed problem
since the last audit — a new fault event, a newly CrashLooping pod, or a
newly NotReady node. A newly appeared problem must never be absorbed into a
stale audit; it always gets a fresh investigation.
- Proceed if there is no prior audit to compare against.
Never link a periodic audit to an investigation that is still PENDING_TRIAGE,
PENDING_START, or IN_PROGRESS — linking re-activates it and can keep it
running indefinitely. Only skip (concurrent) or link to a completed one.
Notes
- These are correlation preferences for the triage stage; the full diagnosis
(timeline, verdict, recommended actions) is produced later by the
hyperpod-incident-rca skill on investigations that proceed.
- Routine
Info-level activity and scale-in-progress churn (events mentioning
"lost orchestration-ready status") are already filtered upstream and should
not, on their own, drive a new investigation.