Two complementary tools keep the Ntuition Engine honest: maintenance windows (for planned disruption) and tuning rules (for individual signals that need custom handling).

Maintenance windows

Configured on Settings → Ntuition Engine. A window does two things at once:

  1. Training hygiene — datapoints inside the window are dropped before models train, so a patch-night CPU spike is never learned as "normal Tuesday behavior."
  2. Detection suppression — anomalies from matching hosts are suppressed while the window is active, so patch night doesn't light up the Analysis page either.

Two kinds:

  • One-off — an explicit start and end timestamp. Use for migrations, load tests, failover drills.
  • Weekly — days of the week, a start time (HH:MM), and a duration. Use for recurring backup and patch schedules.

Each window takes a host pattern (substring match; blank = all hosts), so "every Saturday 02:00–06:00 on backup-" is one rule.

Create windows before the disruption. The training-hygiene effect is retroactive only in the sense that excluded intervals stay excluded at every future retrain — data from chaos that ran without a window will influence baselines until it ages out of the training lookback.

Tuning rules

Managed inline from the Ntuition (Analysis) page — every anomaly row has a Tune action, and the Tuning Rules table lists what's in force. A rule matches a metric (or *) and host (or *) and applies one of:

  • suppress — never flag this combination. For the test box that's supposed to spike.
  • threshold — custom bounds: a min value, max value, and/or max acceptable deviation (σ) that override the learned baseline for this combination.

Each rule takes a free-text note — future-you will want to know why the rule exists.

Which tool when?

Situation Tool
Planned, time-bounded disruption Maintenance window
Recurring scheduled jobs Weekly maintenance window
One metric/host pair that's forever weird Tuning rule (suppress)
A signal where the learned baseline is wrong Tuning rule (threshold)
Everything is too noisy platform-wide Neither — raise global sensitivity (see Tuning Ntuition Engine Sensitivity)