Data Anomalies – Use Cases¶

When to Reach for This Module¶

digna Data Anomalies is the right module when you cannot describe in advance what "wrong" looks like.

It learns the natural patterns in your data and alerts you whenever something looks implausible — with no thresholds and no rules to maintain. That makes it the default starting point for any table you have just onboarded, and the only practical option for environments where hand-written checks would run into the hundreds.

The same engine covers two purposes that are usually split across separate tools:

• Data quality — is this delivery complete, well-formed, and consistent with every delivery before it?
• Business and operational KPIs — has a number that the business steers by moved in a way nobody expected?

Both are computed from the same column metrics, in-database, so adding KPI monitoring on top of quality monitoring costs no extra profiling run.

Banking and Financial Services¶

Incomplete overnight loads before risk reporting¶

A core banking table normally receives around 500 000 transaction records per night. One morning the load completes without error, but only 50 000 rows arrived — an upstream extract silently truncated.

Balances and exposure drifting out of range¶

Average account balance, total exposure per segment, or the sum of a position column shifts gradually rather than breaking. Because digna tracks sum, avg, min, and max per column over time, the drift is visible as an anomaly long before it is visible in a quarterly report.

Swapped or misaligned columns after an ETL change¶

A pipeline change reverses first_name and last_name. No job fails, no null appears, and row counts are perfect. What changes is the average string length and value distribution of each column — which is exactly what digna profiles, so the swap surfaces as a paired anomaly on both columns.

Unexpected categorical values¶

A city column that has only ever contained Austrian cities suddenly contains Zurich. digna classifies the column as categorical, learns its value distribution, and flags the new member as unexpected.

Healthcare¶

Clinical and claims data completeness¶

Patient, encounter, and claims tables are assembled from many feeder systems, each with its own release cycle. A missing-value ratio that jumps from 0.2 % to 11 % on a diagnosis code column is not a schema break and not a rule violation — it is a change in behaviour, and it is what this module is built to catch.

Reference data that quietly changes meaning¶

Code lists (diagnosis, procedure, department) are updated by external bodies. When the distribution across a coded column reshapes after such an update, digna reports it, giving analytics teams a chance to check whether historical comparisons are still valid.

Telecommunications¶

Telecommunications teams move high volumes of customer and operational data through distributed pipelines, so a small change at one source can create widespread downstream errors.

• CDR and usage volumes — call, message, and data-usage counts have strong daily and weekly seasonality. digna learns that shape, so a genuine drop on a Tuesday is flagged while the normal Sunday dip is not.
• Per-region deliveries — with metrics defined on filtered subsets (Datasets), each region, product line, or network element is monitored on its own baseline instead of disappearing into a national total.
• Churn and activation KPIs — the same profiling run that guards data quality also watches the operational numbers built from it.

Retail and E-Commerce¶

• Sales and order volumes per store, channel, or SKU group, each with its own learned pattern.
• Rising NULL rates after a migration — a classic post-cutover failure where the job succeeds and the content degrades.
• Duplicate batch loads — an ingestion replay doubles a day's revenue; the record count and the sum of the amount column both move together, which reads unmistakably as a duplicated load rather than a good day.

Public Sector¶

Public sector agencies typically need traceability across systems rather than speed alone. digna Data Anomalies contributes the audit trail: every flagged anomaly carries the metric, the observed value, the learned expected range, a timestamp, and a confidence score, so a data issue raised months later can be reconstructed rather than argued about.

Register and case-management data also tends to be irregular by nature — filing peaks at deadlines, quiet periods between them. Learned seasonality handles this without an analyst maintaining a calendar of exceptions.

Choosing What to Monitor¶

Broad coverage is the point of this module, but not every column deserves attention. digna lets you disable metrics per column, per table, or per project, which keeps profiling fast and the alert list credible.

A practical starting point:

Where This Module Fits¶

Related Pages¶

• Data Anomalies – Introduction
• Data Anomalies – How It Works
• digna Modules – Technical Overview
• How to add a datasource to a project