Skip to main content

Judged against the table's own history, never a threshold.

When you connect, Sentry backfills ninety days of each table's history in at most two queries, so the first night's scan already knows what normal looks like.

Learned baselines, not thresholds.

Sentry profiles ninety days of each table's history and judges every night against that rhythm: weekend dips, Monday loads, month-end spikes.

  • Daily rows
  • Learned band
  • Fixed threshold

orders · rows per day

tonight05k10k90 days ago

What the baseline keeps

Every night's snapshot joins the history to form a baseline context the agent uses to decide what to flag.

Row counts by weekday

Daily row counts with a weekday-specific mean and spread, so a quiet Sunday is not an incident.

Freshness lag and cadence

Median, p90 and max lag between load and now, plus how often the table actually loads.

Null rate per column

Each column's historical share of nulls.

Numeric spread

Mean and standard deviation per numeric column, and the largest excursion ever seen.

Text shapes

The length range and the corpus of value shapes each text column has produced.

Key cardinality

Whether a column has always been unique, and its duplicate rate.

Severity and sensitivity

Judged by the agent

Info, warning or critical, judged from the anomaly itself: critical means data is definitely missing or broken; warning means a deviation a person should look at; info is notable but likely benign.

Shifted by priority

The table's priority tier moves the result one step: critical tables escalate, low tables de-escalate, and both are recorded on the finding.

Sensitivity dial

Relaxed, normal or paranoid per table, scaling every candidate threshold up or down. Available on Scale and Ultra.

Ninety days of context from night one