Apache Spark với Python: Xây dựng Data Pipeline từng bước

# spark_setup.py
from pyspark.sql import SparkSession

spark = (
    SparkSession.builder
    .appName("ETLPipeline")
    .config("spark.sql.adaptive.enabled", "true")       # AQE for runtime optimization
    .config("spark.sql.shuffle.partitions", "200")       # Tune based on cluster size
    .config("spark.serializer", "org.apache.spark.serializer.KryoSerializer")
    .getOrCreate()
)

# Verify Spark version
print(spark.version)  # 4.0.0

# read_sources.py
from pyspark.sql.types import StructType, StructField, StringType, DoubleType, TimestampType

# Define schema explicitly — avoids slow schema inference on large files
order_schema = StructType([
    StructField("order_id", StringType(), nullable=False),
    StructField("customer_id", StringType(), nullable=False),
    StructField("product_code", StringType(), nullable=True),
    StructField("amount", DoubleType(), nullable=True),
    StructField("order_date", TimestampType(), nullable=True),
    StructField("region", StringType(), nullable=True),
])

# Read CSV with explicit schema
raw_orders = (
    spark.read
    .schema(order_schema)                    # Skip inference, enforce types
    .option("header", "true")                # First row contains column names
    .option("mode", "DROPMALFORMED")         # Skip rows that don't match schema
    .csv("s3a://data-lake/raw/orders/")
)

# Read Parquet — schema is embedded in the file format
products = spark.read.parquet("s3a://data-lake/raw/products/")

raw_orders.printSchema()
raw_orders.show(5, truncate=False)

# clean_orders.py
from pyspark.sql import functions as F

cleaned_orders = (
    raw_orders
    .filter(F.col("order_id").isNotNull())                     # Remove null primary keys
    .filter(F.col("amount") > 0)                               # Filter invalid amounts
    .withColumn("customer_id", F.trim(F.upper(F.col("customer_id"))))  # Normalize IDs
    .withColumn("order_date", F.to_date(F.col("order_date")))  # Cast to date type
    .withColumn("year_month",                                   # Extract partition key
        F.date_format(F.col("order_date"), "yyyy-MM"))
    .dropDuplicates(["order_id"])                               # Deduplicate by order ID
)

print(f"Raw: {raw_orders.count()} rows -> Cleaned: {cleaned_orders.count()} rows")

# enrich_orders.py
# Join orders with product catalog
enriched = (
    cleaned_orders
    .join(products, on="product_code", how="left")      # Keep all orders, even unmatched
    .select(
        "order_id",
        "customer_id",
        "product_code",
        F.col("name").alias("product_name"),              # Rename for clarity
        "amount",
        F.col("category"),                                # From products table
        "order_date",
        "year_month",
    )
)

# Aggregate: monthly revenue per product category
monthly_revenue = (
    enriched
    .groupBy("year_month", "category")
    .agg(
        F.sum("amount").alias("total_revenue"),            # Sum all order amounts
        F.countDistinct("order_id").alias("order_count"),  # Unique orders
        F.avg("amount").alias("avg_order_value"),          # Average basket size
    )
    .orderBy(F.desc("total_revenue"))
)

monthly_revenue.show(10)

# write_output.py
# Write enriched data partitioned by year_month
(
    enriched
    .repartition("year_month")                             # Align partitions with output
    .write
    .mode("overwrite")                                     # Replace existing partition data
    .partitionBy("year_month")                             # Physical directory partitioning
    .parquet("s3a://data-lake/curated/enriched_orders/")
)

# Write aggregated metrics as a single compact file
(
    monthly_revenue
    .coalesce(1)                                           # Single output file for small results
    .write
    .mode("overwrite")
    .parquet("s3a://data-lake/curated/monthly_revenue/")
)

print("Pipeline complete — output written to curated layer")

# custom_source.py
from pyspark.sql.datasource import DataSource, DataSourceReader
from pyspark.sql.types import StructType, StructField, StringType, IntegerType

class APIDataSource(DataSource):
    """Custom data source reading from an internal REST API."""

    @classmethod
    def name(cls) -> str:
        return "internal_api"                              # Registered source name

    def schema(self) -> StructType:
        return StructType([                                # Define output schema
            StructField("id", IntegerType()),
            StructField("name", StringType()),
            StructField("status", StringType()),
        ])

    def reader(self, schema: StructType) -> "APIDataSourceReader":
        return APIDataSourceReader(self.options)           # Pass options to reader

class APIDataSourceReader(DataSourceReader):
    def __init__(self, options):
        self.endpoint = options.get("endpoint", "")        # API endpoint URL

    def read(self, partition):
        import requests                                    # Import inside read for serialization
        response = requests.get(self.endpoint)
        for record in response.json():
            yield (record["id"], record["name"], record["status"])

# Register and use the custom source
spark.dataSource.register(APIDataSource)
api_df = spark.read.format("internal_api").option("endpoint", "https://api.internal/users").load()
api_df.show()