Data Autocleaning

Data Autocleaning Skill

Automated data profiling, quality assessment, and transformation for data sourced from BigQuery or Google Cloud Storage (GCS).

When to Use

[!IMPORTANT]

You MUST use this skill for ANY task where the source is BigQuery or GCS — including seemingly simple operations like "move data" or "copy table".

• Apply to all operations on new and existing sources: copying, moving, appending, ingesting, or extracting data.
• Apply to the source node specifically, not to subsequent pipeline steps.
• Never skip Dataplex profiling (Steps 1 and 3). Always use Dataplex — not ad-hoc BigQuery profiling.

Task Execution Workflow

Step 1: Preliminary Checks (Before Implementation Planning)

Perform these checks before generating the implementation_plan.md.

1. Check Eligibility — You MUST confirm the source is a BigQuery table or GCS source.

2. Gather Data Profile via Dataplex:

   • GCS sources: For GCS sources, you MUST create an external table first before running the dataplex scan.

   • Wait for results: You MUST NOT proceed until the Dataplex profile is available, unless user scan approval was denied.

   • Use the profile as input for cleansing and schema mapping decisions. The transformations MUST NOT be finalized before profile information is available (unless scan was denied).

   • Commands:

     1. Obtain user approval: Present the scripts/dataplex_scanner.py scan command to the user and obtain explicit approval before executing it. Use the following template to present the command:

        • Command: python3 scripts/dataplex_scanner.py ... (Fetch full arguments from step 6 below)
        • Summary: The script automates Dataplex data profiling. It checks table sizes, applies dynamic sampling for large tables (>1M rows) to reduce costs, skips empty tables, executes concurrent scans for multiple tables, and polls for results automatically.
        • Value Add: Enables deep data analysis (null rates, distinct values, distributions) allowing data-driven cleansing decisions. It helps identify hidden anomalies (garbage values, format variance) to guide accurate transformations and verifies that the cleaning logic resolves them without introducing regressions.
        • Scope: The approval obtained here covers all executions of this scanner script for this task (including verification steps).

     2. Run the scripts/dataplex_scanner.py script located in the same directory as this SKILL.md file. This script handles concurrent scan creation, dynamic sampling for large tables, and polling for results. Use --help to learn more.

     3. The script will save the full results as JSON files in the specified output directory.

     4. [!IMPORTANT] The location MUST be a specific Google Cloud region like us-central1; multi-regions like us are not supported in Dataplex scan.

     5. If there are multiple tables to scan, provide them all in the --tables argument to run them concurrently.

     6. Use the following command template:

        python3 scripts/dataplex_scanner.py \
          --tables <project.dataset.table> <project.catalog.namespace.table> \
          --location <location> \
          --output-dir <output_dir>
        

        Note: The script accepts table IDs in the format project.dataset.table for BigQuery tables and project.catalog.namespace.table for BigLake Iceberg tables.

3. Fetch Schema & Samples — Use bq commands to fetch schema and sample data for both source and destination tables.

Step 1.5: Implementation Plan Requirements

1. Your implementation_plan.md MUST include a Profiling Evidence section. Note: If scan execution was denied by the user, document the denial reason here instead of Job IDs.

## Profiling Evidence
- [ ] Dataplex Data Profile Job ID: <JOB_ID>
- [ ] Profile Result Summary: <Brief summary of key findings, e.g., % nulls, distinct values>

1. Your implementation_plan.md MUST include a step to generate cleansing SQL transformations based on the profile output and instructions in Step 2: Generate Transformations.
2. Your implementation_plan.md MUST also reference Step 3 (Quality Review) under its Verification Plan section.

[!CAUTION]

Do not proceed to implementation until both sections are completed. You MUST ensure that the verification phase only validates that your transformations successfully addressed the anomalies found in Step 1.

Step 2: Generate Transformations

Schema Alignment

• Match the destination table schema (types and names) if provided.
• Do not perform column splits, merges, or other schema operations when no destination table is specified.

Data Cleaning Rules

• Garbage Values: Drop or convert to NULL only for malformed data (e.g., unparseable dates, zero-length strings for non-nullable integers).
• Unit Normalization: Standardize measurable units (e.g., 'C' → 'F') to the most common unit. If units are too varied (e.g., mg, liter), leave as-is.
• Type Conversion: Use COALESCE with SAFE.PARSE_* functions for multiple date/time/datetime/timestamp formats. Fetch diverse samples when source data shows high variance.

JSON Data Handling

• Parsing: Use SAFE.PARSE_JSON to cast JSON strings to JSON type. Never use deprecated JSON_EXTRACT_*.
• Extraction: Flatten or extract fields only if a destination schema requires it.
• Accessors: Use JSON_VALUE, JSON_QUERY, JSON_QUERY_ARRAY, JSON_VALUE_ARRAY without SAFE. prefix (they are safe by default).
• Schema mapping: When a destination schema is provided, extract JSON fields to match target column names and types.
• NULL handling: If SAFE.PARSE_JSON returns NULL, keep the original string and note the invalid JSON in the cleaning summary.

Array Data Handling

• Unnesting: Unnest array fields only if a destination schema explicitly requires it.
• Type Casting: Attempt to cast elements to the most appropriate common type.
• CRITICAL: Filter out NULL elements after SAFE_CAST (e.g., using ARRAY) as BigQuery arrays cannot contain NULLs.
• Normalization: Only trim whitespace on string fields. Make sure to PRESERVE CASE. DO NOT perform case conversions (e.g., LOWER(), UPPER()) unless explicitly required.
• Filtering: Filter out NULL values using ARRAY_FILTER(array_column, e -> e IS NOT NULL).
• Deduplication: Use ARRAY(SELECT DISTINCT x FROM UNNEST(array_column)) for case-sensitive deduplication.
• Transformations: Use ARRAY_TRANSFORM or UNNEST/ARRAY_AGG for element-wise changes (e.g., date parsing).
• Restructuring: Use UNNEST to expand to rows, or ARRAY_AGG to group rows into an array, as required by the destination schema.

STRUCT/Record Data Handling

• Extraction: Extract fields to top-level columns only if the destination schema requires it.
• Type Casting: Cast each field using SAFE_CAST based on the destination schema or inferred profile.
• Normalization: Only trim whitespace on string fields. Make sure to PRESERVE CASE. DO NOT perform case conversions (e.g., LOWER(), UPPER()) unless explicitly required.
• Field Mapping: Map directly if structures align; use dot notation (e.g., struct.field) to extract; or use STRUCT() constructor to group columns.
• Schema Alignment: Populate missing fields with NULL and drop fields not present in the destination schema.

Step 3: Quality Review & Profiling

[!IMPORTANT]

You MUST verify transformations strictly using the protocol below before completing the task. Never skip this step. Use Dataplex profiling only (unless scan was denied by the user) — not ad-hoc SQL queries.

Quality review protocol:

1. Extract the SELECT query containing all generated transformations (autocleaning, schema mapping, JSON extractions).

2. Create a temporary sample output table (max 1M rows, 1-hour TTL) by running the transformation query.

3. Fix any runtime errors and re-run until the query succeeds.

4. Profile the temporary sample output table using Dataplex:

   • Verify approval: If scan execution was approved in Step 1, proceed. If approval was DENIED in Step 1, DO NOT run the scanner script and DO NOT ask the user for approval again. Proceed with manual verification using bq sample queries to ensure transformations were successful.
   • Run the scripts/dataplex_scanner.py script on the temporary table.
   • The script will automatically wait for the profile job to finish and save the results as JSON.

5. Compare profiles (Skip if scans were denied) — Check the new profile against the Step 1 profile for every transformed column:

   | Anomaly Type | Threshold |
   | --- | --- |
   | **NULL increase** | >1% increase compared to source (unless expected) |
   | **Value range shift** | Unexpected ranges or formats |
   

6. Iterate on anomalies — For each anomaly:

   1. Identify: Query samples where source is NOT NULL but transformed value IS NULL.
   2. Fix: Update the transformation logic.
   3. Repeat: Re-run Step 3 until the anomaly is resolved.

Step 3.5: Quality Review Evidence Requirements

Your walkthrough.md MUST include a Quality Review Profiling Evidence section. Note: If scan execution was denied by the user, document the denial reason here instead of Job IDs.

## Quality Review Profiling Evidence
- [ ] Post-Transformation Dataplex Profile Job ID: <JOB_ID>
- [ ] Profile Comparison Summary: <Detailed comparison between initial and final profiles per column>

[!CAUTION]

Do not conclude the task or ask for user review until this section is filled and the profile comparison is documented.

Step 4: Documentation

Your walkthrough.md must contain a table for each transformation in the following format:

| Field | Description |
| --- | --- |
| **Destination schema considered** | The target column/type being matched |
| **Issue Detected** | What data quality problem was found |
| **Transformation Applied** | The SQL logic used to fix it |
| **Benefit** | Why this transformation improves the data |

Include a summary of all quality review steps and profiling evidence.

Definition of Done

• All source data quality issues are identified and addressed via SQL transformations.
• Verification MUST be completed using the Quality review protocol and documented with evidence.
• The verification step MUST only test the changes and should succeed when the sql is executed.
• Transformations align with the target schema if provided.
• Cleaning summary is provided with clear justification for each transformation.
• If skipped, the reason is clearly stated (e.g., ineligible source).