Networking Minerva User Guide

Notifications

NVIDIA Minerva sends a notification when a view on the Anomalies, Incidents, or Trends page returns new events. Notifications go to the endpoints you choose, such as a Slack channel, an email list, or ServiceNow. The notification framework runs in the Platform Stack, as part of the Presentation layer described in Architecture.

How Notifications Work

  • Views: each view decides which events trigger a notification, which endpoints receive it, and how duplicate notifications are grouped.

  • Endpoints: the destinations, each with a unique name and an endpoint type.

  • Endpoint types: settings shared by every endpoint of one type, the rate limiter and the templates.

  • Templates: define the content of the notification for an endpoint type.

For setup steps, see Configure Notifications.

Events

A view can notify on raw events or on events that NVIDIA Minerva generates.

Raw Events

Raw events include syslog messages, JSON logs, and SNMP traps. For example, a view can notify on one SNMP trap type, such as coldStart, or on one log label, such as "Parity Error".

Generated Events

  • ML anomalies on metrics or logs: for example, "GPU Temperature Anomalies" or "Rare Syslogs".

  • Finite state machine (FSM) events: BGP and interface state changes, such as "Link Flaps".

  • Auto-correlation events: events that NVIDIA Minerva groups with ML.

  • Threshold events on metrics: notifications when a metric crosses a threshold you set. NVIDIA recommends ML-based metric anomalies instead, because they produce fewer false positives.

Deduplication and Rate Limiting

Deduplication dimensions group a view's notifications. Notifications with the same values for those dimensions share a dedupe key. The rate limiter of the endpoint type then limits how many notifications with the same dedupe key each endpoint receives within a time window. Notifications over the limit are dropped.

For example, a view that notifies on interface down events uses Device Name and Interface Name as its deduplication dimensions. If the endpoint type allows one notification per 24 hours, a link that keeps flapping sends one notification a day to each endpoint instead of one for every flap.

The rate limiter of the endpoint type applies only to views with deduplication dimensions.

Supported Endpoints

Endpoint type

Custom templates

Notes

Email

Yes

A subject and a plain-text body.

Kafka

No


OpenTelemetry (alpha)

Yes

Maps event fields to an OTLP log record.

ServiceNow

Yes

Creates and updates tickets. See ServiceNow Endpoints.

Slack

Yes


Syslog

No


Webhook

Yes


Last updated: