data-quality-monitoring

Data Quality Monitoring

Trust your data, or fix it before anyone else sees it.

Quality Dimensions

Dimension	Definition	Measurement
Completeness	All expected data present	Row count, null ratio
Consistency	Data follows expected patterns	Schema match, value ranges
Timeliness	Data arrives on schedule	Freshness, latency
Accuracy	Data reflects reality	Comparison to source
Uniqueness	No unexpected duplicates	Distinct key ratio

Anomaly Detection

Row Count Monitoring (Z-Score)

def check_row_count(source: str, current_count: int,
                    expected: float, stddev: float,
                    threshold_pct: float = 0.25) -> dict:
    """Detect row count anomalies using rolling baseline."""
    z_score = abs(current_count - expected) / max(stddev, 1)
    deviation_pct = abs(current_count - expected) / max(expected, 1)

    return {
        "source": source,
        "current": current_count,
        "expected": round(expected),
        "deviation_pct": round(deviation_pct, 4),
        "z_score": round(z_score, 2),
        "is_anomaly": deviation_pct > threshold_pct,
        "severity": _severity(deviation_pct),
    }

def _severity(deviation: float) -> str:
    if deviation > 0.5: return "CRITICAL"
    if deviation > 0.25: return "HIGH"
    if deviation > 0.1: return "MEDIUM"
    return "LOW"

Severity Actions

Severity	Deviation	Action
CRITICAL	> 50%	BLOCK — Manual review required
HIGH	> 25%	WARN — Review before proceeding
MEDIUM	> 10%	LOG — Monitor trend
LOW	≤ 10%	PASS — Within normal range

Schema Drift Detection

def detect_schema_drift(actual_cols: dict, expected_cols: dict,
                        strict: bool = False) -> dict:
    """Compare actual vs expected schema."""
    missing = set(expected_cols) - set(actual_cols)
    added = set(actual_cols) - set(expected_cols)
    type_changes = {
        col: (expected_cols[col], actual_cols[col])
        for col in set(actual_cols) & set(expected_cols)
        if expected_cols[col] != actual_cols[col]
    }

    return {
        "has_drift": bool(missing or (strict and added) or type_changes),
        "missing_columns": list(missing),
        "added_columns": list(added),
        "type_changes": type_changes,
        "severity": "HIGH" if missing or type_changes else "LOW",
    }

Null Ratio Monitoring

def check_null_ratios(columns: dict, total_rows: int,
                      thresholds: dict = None) -> list:
    """Flag columns exceeding null thresholds."""
    default_threshold = 0.1
    thresholds = thresholds or {}
    issues = []

    for col, null_count in columns.items():
        ratio = null_count / max(total_rows, 1)
        threshold = thresholds.get(col, default_threshold)
        if ratio > threshold:
            issues.append({
                "column": col,
                "null_count": null_count,
                "null_ratio": round(ratio, 4),
                "threshold": threshold,
                "severity": "HIGH" if ratio > threshold * 2 else "MEDIUM",
            })

    return issues

Freshness Check

from datetime import datetime

def check_freshness(latest_timestamp: datetime,
                    max_age_hours: int = 24) -> dict:
    """Check data freshness against maximum allowed age."""
    age_hours = (datetime.now() - latest_timestamp).total_seconds() / 3600

    return {
        "fresh": age_hours <= max_age_hours,
        "latest_record": latest_timestamp.isoformat(),
        "age_hours": round(age_hours, 2),
        "max_age_hours": max_age_hours,
        "status": "FRESH" if age_hours <= max_age_hours else "STALE",
    }

Data Protection Patterns

Safe Write (Prevent Empty Overwrites)

def safe_write(current_count: int, existing_count: int,
               min_rows: int = 100, max_drop_pct: float = 0.5) -> list:
    """Validate before overwriting production data."""
    failures = []

    if current_count == 0 and existing_count > 0:
        failures.append(f"NO_EMPTY_OVERWRITE: cannot replace {existing_count} rows with 0")

    if existing_count > 0 and current_count < existing_count * (1 - max_drop_pct):
        drop = (1 - current_count / existing_count) * 100
        failures.append(f"MAX_DROP: {drop:.1f}% exceeds {max_drop_pct * 100}% threshold")

    if current_count < min_rows:
        failures.append(f"MIN_ROWS: {current_count} below minimum {min_rows}")

    return failures  # Empty list = safe to write

Protection Decision Matrix

Scenario	Action
New data = 0 rows, existing > 0	BLOCK — empty overwrite
Drop > 50%	BLOCK — dramatic data loss
Drop > 25%	WARN — review before proceeding
Below minimum row count	BLOCK — incomplete pipeline
All checks pass	WRITE

Quality Report

def generate_quality_report(source: str, row_count: int,
                            checks: dict) -> dict:
    """Combine all checks into a single report."""
    report = {
        "source": source,
        "timestamp": datetime.now().isoformat(),
        "row_count": row_count,
        "checks": checks,
        "overall_status": "PASS" if all(
            not c.get("is_anomaly") and c.get("severity", "LOW") != "HIGH"
            for c in checks.values() if isinstance(c, dict)
        ) else "FAIL",
    }
    return report

Baseline Management

Rolling Baseline Update

def update_baseline(history: list, new_value: int,
                    window_size: int = 30) -> dict:
    """Update rolling baseline with new observation."""
    history = history[-window_size:] + [new_value]
    avg = sum(history) / len(history)
    variance = sum((x - avg) ** 2 for x in history) / len(history)

    return {
        "avg": round(avg, 2),
        "stddev": round(variance ** 0.5, 2),
        "samples": len(history),
        "last_value": new_value,
        "history": history,
    }

Implementation Checklist

• Row count anomaly detection on every pipeline run
• Schema drift check against registered schema
• Null ratio thresholds per column documented
• Freshness alerts for time-sensitive data
• Safe-write guards on production tables
• Rolling baselines updated after successful runs
• Quality reports stored for trend analysis
• Alerting integrated (email, Slack, PagerDuty)