Core features

Auto-apply & verification

A recommendation is only useful if it is closed out once it is fixed. OptiHouse tracks every finding through its lifecycle and can automatically detect when you have already applied a fix — so the list reflects reality without manual bookkeeping.

The recommendation lifecycle

A finding moves through a small set of states. open means it is live and unaddressed. applied means the fix is in place — set either by you or automatically. resolved means the underlying signal is gone and confirmed. dismissed means you decided not to act on it.

The gap auto-apply closes

Most fixes get made but never marked done — someone runs the ALTER, the recommendation just lingers as open. Auto-apply transitions those automatically so the open list stays trustworthy.

Phase 1 — absence rule

If an open recommendation stops appearing in new scans, the fix was probably made. To avoid false positives from a transient threshold miss, a finding is only auto-applied once it has been absent for several consecutive scans and is at least 7 days old. Anything younger that briefly disappears is treated as noise, not a fix.

Phase 2 — evidence verifiers

Disappearance alone is a weak signal. For nine recommendation types, OptiHouse runs a dedicated verifier that looks for positive proof in system.* — a codec actually changed, a skipping index now exists, a column was dropped — and also correlates DDL events from query_log.

Verifier checks	Examples
Direct system-table evidence	New codec on a column, a `data_skipping_indices` row, a projection that now exists.
DDL correlation	An `ALTER ... MODIFY COLUMN` / `ADD INDEX` / `ADD PROJECTION` seen in `query_log` for the right table.
Fragmentation recovery	Tiny-part ratio dropping back under a healthy threshold.

Confidence threshold

A verifier returns a confidence score. Only evidence at 0.85 or above transitions a recommendation to applied. Weaker signals are kept as hints and never auto-apply on their own.

How a fix is attributed

Every applied recommendation records how it got there, so you can always tell a human action from an automated one:

manual — you clicked Mark applied.
auto-rule — Phase 1: absent for several scans and old enough.
auto-evidence — Phase 2: a verifier found positive proof above the confidence threshold.
webhook — closed by an external integration.

Undoing an auto-apply

If a recommendation was auto-applied but you did not actually fix it, the detail page shows a "Wait — I didn't actually fix this" action that reopens it. Manual applies use the normal Reopen button. Either way the finding returns to open and the next scan re-evaluates it.

Real-time alerts

High-risk and critical findings can fan out to Slack, PagerDuty, Opsgenie, Telegram or a generic webhook the moment they are auto-applied. Alerting is opt-in per connection and is gated on severity, so routine low-severity churn never pages anyone.