vmanomaly-preset

Dynamic anomaly presets

Deliver a small, reproducible monitoring bundle. Start with the customer's authorized datasource/tenant, service scope, detection intent and deployment preference. Infer metric properties from data; ask only for business constraints or destinations that cannot be inferred. This workflow does not authorize deployment, remote dashboard updates or alert delivery.

For MCP/Copilot delivery, follow the versioned preset completion guide, the canonical source of the compact guide distributed by MCP. It defines the bounded autotune and holdout quality check, file-generation completion, missing-destination handling and workspace-preview boundaries.

Discover and choose

1. Check runtime tools and schemas. vmanomaly 1.31.0 sizing and MCP 0.5.0 are established capabilities; discovery and artifact generation require the newer endpoints described in the API contract. On 404, verify the server version and configured base URL, including any path prefix; only a confirmed missing endpoint means unsupported capability. Never interpret it as an empty inventory. If unavailable, use an already-configured VictoriaMetrics query tool within the same scope or deliver a clearly limited manual plan.
2. Set discovery budgets before calls: ordinarily at most 10 metadata requests, 100 items per call and one day of metadata, then profile at most five candidate queries. Broaden only with evidence and a stated cost. Track cumulative calls, returned items and incomplete responses; stop when the budget is reached. These are workflow defaults, not claims about TSDB scan cost.
3. Discover metric names through __name__ values, then relevant label names/values. Metric prefixes are clues, not proof of semantics. Do not enumerate tenants. Read recipes for the selected domain; check every referenced metric and label before constructing a query.
4. Rank by user relevance, coverage, actionable interpretation, cost and existing alerts. Classify static, temporal, joint, peer or context-only. Keep counters as rates, normalize units, and avoid unsupported histogram assumptions. Missing metadata is unknown. Sample across entity groups rather than selecting only the first series.
5. For chosen exact queries, follow the available vmanomaly-config workflow or perform its equivalent: characteristics, live model schema, causal candidate evaluation and full-config validation at the same step/timezone. Prefer online models; preserve complete multivariate groups and comparable peer pools. Default inference to minute-scale cadence, ordinarily 5m, unless intent requires otherwise. Choose query step separately to retain relevant events and preserve existing query resolution. Choose explicit refit cadence as described in the completion guide; periodic fit_every otherwise inherits inference cadence.
6. Use existing deployment sizing for the actual topology and cardinality evidence. Sample counts are lower bounds, not exact fleet size. Keep unsupported configurations unsupported; do not substitute fewer peers, shorter history or a slower cadence to obtain a feasible estimate.