Data Lake Management Skill

Purpose

Design, build, and maintain production-grade data lakes on AWS using medallion architecture, ensuring data quality, governance, and cost optimization.

When to Use This Skill

Auto-activates when working with:

• Data lake architecture design
• Medallion architecture (Bronze/Silver/Gold)
• AWS S3 data organization
• AWS Glue Data Catalog
• Data partitioning strategies
• Schema evolution
• Data quality and governance
• Cost optimization
• Query performance tuning

Core Concepts

Medallion Architecture

Bronze Zone (Raw)
├── Immutable source data
├── Original format
├── Minimal validation
└── Append-only

Silver Zone (Cleaned)
├── Validated and cleaned
├── Deduplicated
├── Standardized schema
└── Business rules applied

Gold Zone (Curated)
├── Aggregated
├── Business-ready
├── Optimized for queries
└── Subject-oriented

S3 Data Organization

s3://data-lake-{env}/
├── bronze/                      # Raw zone
│   ├── source_system_1/
│   │   ├── table_name/
│   │   │   ├── year=2025/
│   │   │   │   ├── month=11/
│   │   │   │   │   ├── day=04/
│   │   │   │   │   │   └── data.parquet
│   ├── _metadata/               # Metadata and manifests
│   └── _dlq/                    # Dead letter queue
├── silver/                      # Processed zone
│   ├── domain_1/
│   │   ├── entity_name/
│   │   │   ├── date=2025-11-04/
│   │   │   │   └── part-*.parquet
├── gold/                        # Curated zone
│   ├── analytics/
│   │   ├── daily_aggregates/
│   │   ├── monthly_reports/
└── temp/                        # Temporary processing

Quick Start Examples

1. Bronze Zone Ingestion

import boto3
import pyarrow as pa
import pyarrow.parquet as pq
from datetime import datetime
import json

class BronzeIngestion:
    """Ingest raw data to bronze zone."""

    def __init__(self, bucket: str, source_system: str):
        self.s3 = boto3.client('s3')
        self.bucket = bucket
        self.source_system = source_system

    def ingest(self, data: list, table_name: str):
        """Ingest data with partitioning by ingestion date."""
        now = datetime.now()
        partition_path = f"year={now.year}/month={now.month:02d}/day={now.day:02d}"

        # Convert to Arrow Table
        table = pa.Table.from_pylist(data)

        # Add metadata
        metadata = {
            b'ingestion_timestamp': now.isoformat().encode(),
            b'source_system': self.source_system.encode(),
            b'record_count': str(len(data)).encode()
        }
        table = table.replace_schema_metadata(metadata)

        # Write to S3
        s3_path = f"s3://{self.bucket}/bronze/{self.source_system}/{table_name}/{partition_path}/data.parquet"

        pq.write_to_dataset(
            table,
            root_path=s3_path,
            filesystem=pa.fs.S3FileSystem(),
            compression='snappy',
            use_dictionary=True,
            write_statistics=True
        )

        # Update Glue Data Catalog
        self.update_catalog(table_name, partition_path)

    def update_catalog(self, table_name: str, partition: str):
        """Update Glue Data Catalog with new partition."""
        glue = boto3.client('glue')

        try:
            glue.create_partition(
                DatabaseName='bronze',
                TableName=f"{self.source_system}_{table_name}",
                PartitionInput={
                    'Values': partition.split('/'),
                    'StorageDescriptor': {
                        'Location': f"s3://{self.bucket}/bronze/{self.source_system}/{table_name}/{partition}/",
                        'InputFormat': 'org.apache.hadoop.hive.ql.io.parquet.MapredParquetInputFormat',
                        'OutputFormat': 'org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat',
                        'SerdeInfo': {
                            'SerializationLibrary': 'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'
                        }
                    }
                }
            )
        except glue.exceptions.AlreadyExistsException:
            pass  # Partition already exists

2. Silver Zone Processing (PySpark)

from pyspark.sql import SparkSession
from pyspark.sql.functions import *
from delta.tables import DeltaTable

class SilverZoneProcessor:
    """Process bronze to silver with Delta Lake."""

    def __init__(self):
        self.spark = (SparkSession.builder
            .config("spark.sql.extensions", "io.delta.sql.DeltaSparkSessionExtension")
            .config("spark.sql.catalog.spark_catalog", "org.apache.spark.sql.delta.catalog.DeltaCatalog")
            .getOrCreate())

    def process_incremental(self, bronze_table: str, silver_path: str, primary_key: str):
        """Incremental processing with deduplication and SCD Type 2."""

        # Read bronze (new records only)
        df_bronze = (self.spark.read
            .format("delta")
            .load(bronze_table)
            .filter(col("processed") == False))

        # Clean and transform
        df_cleaned = (df_bronze
            # Remove corrupt records
            .filter(col("_corrupt_record").isNull())

            # Standardize schema
            .select(
                col("id").cast("string").alias("id"),
                col("timestamp").cast("timestamp"),
                col("value").cast("double"),
                to_json(col("metadata")).alias("metadata_json")
            )

            # Add processing metadata
            .withColumn("processed_at", current_timestamp())
            .withColumn("is_active", lit(True))
            .withColumn("valid_from", col("timestamp"))
            .withColumn("valid_to", lit(None).cast("timestamp"))

            # Deduplicate (keep latest per key)
            .withColumn("row_num",
                row_number().over(
                    Window.partitionBy(primary_key)
                    .orderBy(col("timestamp").desc())
                )
            )
            .filter(col("row_num") == 1)
            .drop("row_num")
        )

        # Merge into silver (SCD Type 2)
        if DeltaTable.isDeltaTable(self.spark, silver_path):
            delta_table = DeltaTable.forPath(self.spark, silver_path)

            # Close out old records
            delta_table.alias("target").merge(
                df_cleaned.alias("source"),
                f"target.{primary_key} = source.{primary_key} AND target.is_active = true"
            ).whenMatchedUpdate(
                condition="source.timestamp > target.valid_from",
                set={
                    "is_active": "false",
                    "valid_to": "source.timestamp"
                }
            ).execute()

            # Insert new records
            delta_table.alias("target").merge(
                df_cleaned.alias("source"),
                f"target.{primary_key} = source.{primary_key} AND target.is_active = true"
            ).whenNotMatchedInsert(
                values={
                    col: f"source.{col}" for col in df_cleaned.columns
                }
            ).execute()
        else:
            # Initial load
            df_cleaned.write.format("delta").save(silver_path)

        # Mark bronze records as processed
        (df_bronze
            .withColumn("processed", lit(True))
            .write.format("delta").mode("overwrite").save(bronze_table))

        # Optimize delta table
        delta_table.optimize().executeCompaction()

3. Gold Zone Aggregation

class GoldZoneAggregator:
    """Create business-ready aggregated tables."""

    def create_daily_summary(self, silver_path: str, gold_path: str, date: str):
        """Daily aggregation for analytics."""

        df_silver = spark.read.format("delta").load(silver_path)

        df_gold = (df_silver
            .filter(col("date") == date)
            .filter(col("is_active") == True)

            # Business aggregations
            .groupBy("date", "category", "region")
            .agg(
                count("*").alias("total_count"),
                sum("value").alias("total_value"),
                avg("value").alias("avg_value"),
                min("value").alias("min_value"),
                max("value").alias("max_value"),
                stddev("value").alias("stddev_value"),
                approx_count_distinct("user_id").alias("unique_users")
            )

            # Add derived metrics
            .withColumn("value_per_user", col("total_value") / col("unique_users"))

            # Add metadata
            .withColumn("created_at", current_timestamp())
            .withColumn("data_version", lit("1.0"))
        )

        # Write with overwrite for specific partition
        (df_gold.write
            .format("delta")
            .mode("overwrite")
            .option("replaceWhere", f"date = '{date}'")
            .partitionBy("date")
            .save(gold_path))

        # Create/update Athena table
        self.create_athena_table(gold_path, "daily_summary")

Data Quality Framework

from great_expectations.dataset import SparkDFDataset
import great_expectations as ge

class DataQualityValidator:
    """Data quality validation with Great Expectations."""

    def validate_silver_data(self, df):
        """Validate silver zone data quality."""

        ge_df = SparkDFDataset(df)

        # Schema validation
        ge_df.expect_column_to_exist("id")
        ge_df.expect_column_to_exist("timestamp")
        ge_df.expect_column_to_exist("value")

        # Data validation
        ge_df.expect_column_values_to_not_be_null("id")
        ge_df.expect_column_values_to_be_unique("id")
        ge_df.expect_column_values_to_be_between("value", min_value=0, max_value=1000000)

        # Business rules
        ge_df.expect_column_values_to_match_regex("email", r"^[a-zA-Z0-9_.+-]+@[a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+$")

        # Get validation results
        results = ge_df.validate()

        if not results.success:
            failed_expectations = [exp for exp in results.results if not exp.success]
            raise DataQualityError(f"Validation failed: {failed_expectations}")

        return results

Resource Files

resources/partitioning-strategies.md

• Partition key selection
• Multi-level partitioning
• Partition pruning optimization
• Partition size management

resources/schema-evolution.md

• Schema versioning
• Backward compatibility
• Column addition/removal
• Data type changes

resources/cost-optimization.md

• S3 storage classes
• Lifecycle policies
• Compression formats
• Query optimization

resources/data-governance.md

• Access control (Lake Formation)
• Data lineage tracking
• Compliance (GDPR, HIPAA)
• Data retention policies

resources/query-optimization.md

• Athena query tuning
• Partition pruning
• Columnar format benefits
• Caching strategies

Best Practices

• Partition by date for time-series data
• Use Parquet with Snappy compression
• Implement schema evolution carefully
• Monitor data quality continuously
• Use Delta Lake for ACID transactions
• Implement data lineage tracking
• Set up lifecycle policies for cost control
• Use Glue Data Catalog for discoverability
• Implement incremental processing
• Monitor query costs and optimize
• Use proper access controls
• Document data schemas
• Implement data retention policies

Common Patterns

Incremental Processing with Bookmarks

# Read with bookmark
bookmark = get_bookmark("my_job")
df_new = spark.read.parquet(source).filter(col("timestamp") > bookmark)

# Process
df_processed = transform(df_new)

# Write and update bookmark
df_processed.write.parquet(target)
update_bookmark("my_job", df_new.agg(max("timestamp")).collect()[0][0])

Data Catalog Registration

glue.create_table(
    DatabaseName='silver',
    TableInput={
        'Name': 'events',
        'StorageDescriptor': {
            'Columns': schema_columns,
            'Location': 's3://data-lake/silver/events/',
            'InputFormat': 'org.apache.hadoop.hive.ql.io.parquet.MapredParquetInputFormat',
            'OutputFormat': 'org.apache.hadoop.hive.ql.io.parquet.MapredParquetOutputFormat',
            'SerdeInfo': {
                'SerializationLibrary': 'org.apache.hadoop.hive.ql.io.parquet.serde.ParquetHiveSerDe'
            }
        },
        'PartitionKeys': [
            {'Name': 'date', 'Type': 'string'}
        ]
    }
)

Status: Production-Ready Last Updated: 2025-11-04 Architecture: Medallion (Bronze → Silver → Gold)