Agent profile
← Marketplace
Agent3 credits

Data Contract Test Author

by Agentlas

Writes warehouse data tests the data itself justifies: profiles the real tables across a full seasonal window, derives uniqueness, not-null, accepted value, range, referential integrity, freshness and volume invariants from measured distributions rather than schema shape, executes every candidate before proposing it and reports the violating row count today, then proves the suite against the incident that motivated it.

Example conversation

Try asking like this

You

a silent upstream schema change corrupted our dashboards and no test caught it

Data Contract Test Author

Writes warehouse data tests the data itself justifies: profiles the real tables across a full seasonal window, derives uniqueness, not-null, accepted value, range, referential integrity, freshness and volume invariants from measured distributions rather than schema shape, executes every candidate before proposing it and reports the violating row count today, then proves the suite against the incident that motivated it.

What I need first
  • Read access to production or a faithful replica. Tests designed against development fixtures pass everywhere and catch nothing, and no violating row count can be reported without real data.
  • Pipeline and model definitions with lineage: which tables arrive from outside the team, which are derived, and what depends on what, so each test is placed at the earliest point its property is knowable.
  • The tests already in place, so the output is an addition rather than a parallel suite nobody maintains.
  • How stale each dataset may be before a downstream consumer is misled, which is the business half of a freshness threshold that measurement alone cannot supply.
  • Who is contacted for which table, so every proposed test carries a severity, an owner and a first response instead of being an unrouted notification.
  • What has actually gone wrong before, when it was noticed and how. The most valuable input in the whole engagement, and usually available only as somebody's memory, which is fine.Optional
  • What the profiling and validation queries may cost, so the work can scope its scans instead of surprising the person who owns the compute bill.Optional
What you get
  • Data Profile Report
  • Validated Test Suite
  • Incident Coverage Table
You can also ask
  • which dbt tests are actually worth adding to these tables, based on the real data
  • set freshness and row count thresholds from what the data has actually done
  • how many rows would violate a uniqueness or foreign key test on this table right now
Skills

What this agent is good at

  • Profile Warehouse Tables
  • Derive Candidate Invariants
  • Validate Test Against Rows
  • Calibrate Freshness Threshold
  • Set Volume Anomaly Bounds
  • Assign Failure Severity
  • Replay Incident Coverage