OnCallVal taxonomy
Generated from src/oncallval/faults.py. Tier = f(weight):
medium 2–4, hard 5–7, severe 8+.
weight = Σ fault base weights + 3 × (extra faults) + 1 × decoys.
Fault families
| Family | Base weight | Signature | Minimal fix | Health predicate |
|---|---|---|---|---|
bad_deploy |
3 | error spike + panic logs aligned with the newest deploy | rollback |
current version ≠ buggy version |
memory_leak |
4 | mem_pct climbs since deploy, OOMKilled events, restart churn | rollback |
current version ≠ buggy version |
config_drift |
3 | config ≠ runbook recommended; family-specific error logs | set_config to recommended |
config[key] == recommended |
dependency_outage |
4 | caller logs connection refused to the dead endpoint; external probes down |
fail caller's endpoint key to the backup | endpoint ≠ dead endpoint |
capacity_shortfall |
3 | cpu pegged, 503 saturated logs, replicas < ceil(rps / runbook rps_per_replica) | scale to required |
replicas ≥ required |
secret_expiry |
3 | x509 expired logs; cert expires_at_min negative |
rotate_secret |
rotated_at is set |
poison_message |
4 | worker crash loop naming one message id; queue shows huge redelivery count | purge_message |
poison message absent |
host_failure |
4 | one host unhealthy; every service with replicas there degrades; per-host error 55% | drain_host |
primary service has 0 replicas on the host |
Decoys (weight 1 each)
| Decoy | Effect |
|---|---|
irrelevant_deploy |
an innocent deploy on a healthy service, timestamp-correlated with the incident |
known_issue |
periodic WARN logs marked as a tracked known issue |
cosmetic_slow |
p99 elevated but within SLO on an unrelated service |
Actions
Reads (0.2 sim-minutes, 1 budget step): list_services, get_service,
get_metrics, get_deploys, get_runbook, query_logs, peek_queue.
Mutations: restart (+2m), rollback (+5m), set_config (+1m),
scale (+3m), rotate_secret (+2m), purge_message (+1m),
drain_host (+6m), submit_report (ends the episode).
Budgets: medium 26, hard 32, severe 40 actions.
Blast radius
host_failure degrades every managed service with replicas on the dead
host, not only the paged one. dependency_outage pages the caller, one hop
away from the true fault. Severe episodes combine two disjoint faults, both of
which must be inactive for mitigated.