data-quality-monitor-designer

Data Quality Monitor Designer

Purpose

Systems monitor their services and stay blind to their data: the pipeline runs green while a third of yesterday's rows never arrived, a join silently fans out duplicates, and a upstream format change poisons every downstream report. This skill designs the data-quality monitoring layer — which checks, on which datasets, at which pipeline stage, with what severity, owner, and failure action — so data problems surface as typed, routed incidents instead of being discovered by the user who trusted the number. It produces the monitor SPEC and the quarantine policy; it does not wire alerts (that is observability-operator's implementation surface) and it never designs checks that silently "fix" data.

Use When

• Use when: bad, stale, duplicated, or missing data keeps reaching users, reports, or models, and detection is currently "someone notices".
• Use when: a pipeline, warehouse, or critical dataset needs quality gates and nobody can enumerate what is checked today.
• Use when: defining per-dataset quality SLAs (freshness by 06:00, ≥99.5% rows with valid foreign keys) and what happens when they are missed.
• Use when: a data incident post-mortem asks "what monitor would have caught this?" and the answer needs designing, not guessing.
• Do NOT use when: the data in question is TEST data — deterministic fixtures, factories, and seed hygiene are test-data-architect; this skill watches production data.
• Do NOT use when: the signals are system telemetry — latency, errors, saturation, health checks are observability-operator (implementation) and slo-reliability-architect (targets); this skill owns checks on the DATA CONTENT itself.
• Do NOT use when: validating an LLM's response shape against a schema — that is structured-output-validator.
• Do NOT use when: the failing thing is a migration's correctness during a planned data move — one-time verification gates belong to data-migration-runbook-author's runbook; this skill owns the STANDING monitors that outlive the move.

Inputs to Inspect

1. The dataset inventory in scope: tables/topics/files, their producers, their downstream consumers (reports, features, models, exports), and which consumers make decisions on them.
2. The pipeline topology: ingest points, transform stages, serving surfaces — where a check can physically run and how early a defect is catchable.
3. Incident history: past data incidents (silent loss, duplication, drift, late arrival) — each one is a candidate check with a known severity.
4. Existing checks and their gaps: constraint definitions in the store, ad-hoc assertion scripts, pipeline-native tests already present.
5. Contract sources for validity rules: schema definitions, enum registries, business rules docs — what "valid" even means per field.
6. The tenant dimension: whether quality must be judged per tenant (one tenant's feed failing while the aggregate looks fine).

Workflow

1. Rank datasets by blast radius. For each dataset: who consumes it, what decision or feature breaks when it is wrong, and how long a defect goes unnoticed today. Monitor design effort follows this ranking — full six-dimension coverage for the critical few, freshness+volume floors for the long tail.
2. Design checks per dimension, per dataset. For each critical dataset, work the six dimensions and state each check concretely:
   • Freshness: latest event/partition age vs an expected-by deadline.
   • Completeness/volume: row/event counts vs a seasonal-aware expectation band, plus required-field null rates.
   • Uniqueness: duplicate rate on declared keys (joins and replays are the usual injectors).
   • Validity: type/format/enum/range conformance per field, sourced from the contract inputs, not invented.
   • Consistency: referential integrity across stores, aggregate reconciliation (sum of parts vs reported total), cross-system counts.
   • Drift: distribution shift on fields where meaning matters (categorical mix, numeric distribution) — flag, never auto-correct.
3. Place each check at the earliest viable stage. Ingest-time checks stop poison at the door (schema/validity), transform-time checks catch what only exists post-join (uniqueness, consistency), serving-time checks guard what users actually read (freshness, reconciliation). State the placement and why.
4. Assign severity and failure action per check. The action vocabulary is exactly three: block (pipeline stops; defect cannot pass), quarantine (offending rows diverted to a reviewable location with provenance; healthy rows proceed; quarantine has an owner and a drain SLA — mirroring DLQ discipline), alert-and-pass (data flows, humans notified). Silent correction — coercing, defaulting, dropping without a record — is not in the vocabulary.
5. Define per-dataset quality SLAs. The consumer-facing promise: freshness deadline, completeness floor, validity floor — each derived from what the top consumer can tolerate, stated with its measurement window. Where a journey-level SLO exists (slo-reliability-architect), the data SLA feeds it rather than duplicating it.
6. Route incidents. Per dataset: who owns a red check (a named role, not a channel), what the first diagnostic step is (the check's own evidence: failing rows sample, expected-vs-actual counts), and where quarantined data is reviewed. Alert wiring, dedup, and dashboards are handed to observability-operator as an implementation package.
7. Design the monitor-as-code shape. Checks live versioned next to the pipeline they guard (declarative where the stack allows), with the convention stated: adding a dataset of class X requires its floor checks in the same change — a review-time rule, not tooling magic.
8. Deliver the spec in the Output Format with coverage honestly stated: which datasets got full coverage, which got floors, which got nothing yet — an explicit residual list, not an implied "all".

Check catalogs per dimension, expectation-band guidance, and the quarantine-policy template: references/quality-dimension-catalog.md.

Output Format

DATA QUALITY MONITOR SPEC — <scope>
Datasets ranked: <N, by blast radius; coverage tier per dataset>
Per-dataset card:
  <dataset>: consumers=<top consumers/decisions>
  checks: <dimension → check, placement (ingest|transform|serving),
           severity, action (block|quarantine|alert-and-pass)>
  SLA: freshness=<deadline> completeness=<floor> validity=<floor> (window=<w>)
  incident: owner=<role> first-diagnostic=<evidence the check emits>
Quarantine policy: <location, provenance fields, owner, drain SLA>
Tenant dimension: <per-tenant checks where aggregate masks single-tenant failure | n/a + why>
Handoffs: alert/dashboard wiring → observability-operator;
          journey SLO targets → slo-reliability-architect (consumed, not set here)
Residual: <datasets/dimensions deliberately not covered yet>

Validation Checklist

• Every check names its dataset, dimension, placement, severity, owner, and one of the three failure actions — no check "just logs".
• No check silently mutates data: quarantine preserves offending rows with provenance; nothing coerces, defaults, or drops unrecorded.
• Validity rules cite a contract source (schema, enum registry, business rule) — none invented from sample data alone.
• Volume/drift checks use expectation bands aware of seasonality, not naive fixed thresholds guaranteed to false-positive.
• Quarantine has an owner and a drain SLA — it is not a write-only graveyard.
• Per-tenant blindness considered: aggregate-level checks that a single tenant's failure cannot trip are named as such.
• Coverage residual is explicit — datasets left unmonitored are listed, not implied covered.
• The spec hands wiring to observability-operator rather than embedding alert-platform configuration.

Gotchas

• Green pipelines, rotten data: orchestration success says the job ran, not that the data is right. Quality checks assert on CONTENT — never accept "the DAG succeeded" as a quality signal.
• Fixed thresholds rot: "alert if rows < 10,000" false-positives every holiday and misses a 30% drop on the biggest day. Bands come from history with seasonality, and get reviewed like code.
• Silent coercion is data loss with better manners: a transform that defaults bad values to zero destroys the evidence AND the number. Quarantine exists so correction is a reviewed decision.
• Aggregate checks mask per-tenant failure: total volume looks fine while one integration's feed died. The tenant dimension check exists for the datasets where that matters.
• Duplicates arrive at joins and replays, not ingest: uniqueness checked only at the door misses fan-out introduced mid-pipeline. Placement follows where the defect is created.
• Drift checks on the wrong fields alarm-fatigue everyone: monitor distribution only where shift changes meaning (model features, pricing inputs), not every column.
• Reconciliation needs a lag budget: comparing system A to system B before B's sync window closes manufactures false mismatches — state the comparison point, not just the comparison.
• Quality SLAs without consumers are decoration: a freshness deadline nobody depends on will be ignored the first week it is red. Every SLA cites the consumer that needs it.

Stop Conditions

• Asked to design checks that auto-repair failing data (coerce, default, drop, "clean up") as the failure action → refuse silent correction; offer quarantine + reviewed repair, and record any human decision to auto-correct a specific field as an explicit, logged transform, not a monitor behavior.
• Validity rules

---

Content truncated.