databricks-migration-deep-dive

Databricks Migration Deep Dive

Overview

Comprehensive migration strategies for moving to Databricks from Hadoop, Snowflake, Redshift, Synapse, or legacy data warehouses. Covers discovery and assessment, schema conversion, data migration with batching and validation, ETL/pipeline conversion, and cutover planning with rollback procedures.

Prerequisites

• Access to source and target systems
• Databricks workspace with Unity Catalog enabled
• Understanding of current data architecture and dependencies
• Stakeholder alignment on migration timeline

Migration Patterns

Source	Pattern	Complexity	Timeline
Hive Metastore (same workspace)	SYNC / CTAS / DEEP CLONE	Low	Days
On-prem Hadoop/HDFS	Lift-and-shift to cloud storage + UC	High	6-12 months
Snowflake	Parallel run + cutover	Medium	3-6 months
AWS Redshift	Unload to S3 + Auto Loader	Medium	3-6 months
Legacy DW (Oracle/Teradata)	Full rebuild with JDBC extraction	High	12-18 months

Instructions

Step 1: Discovery and Assessment

Inventory all source tables with metadata for migration planning.

from pyspark.sql import SparkSession
from dataclasses import dataclass

spark = SparkSession.builder.getOrCreate()

@dataclass
class TableInventory:
    database: str
    table: str
    table_type: str
    format: str
    row_count: int
    size_mb: float
    columns: int
    partitions: list[str]

def assess_hive_metastore() -> list[TableInventory]:
    """Inventory all Hive Metastore tables for migration planning."""
    inventory = []
    databases = [r.databaseName for r in spark.sql("SHOW DATABASES").collect()]

    for db in databases:
        tables = spark.sql(f"SHOW TABLES IN hive_metastore.{db}").collect()
        for t in tables:
            table_name = f"hive_metastore.{db}.{t.tableName}"
            try:
                detail = spark.sql(f"DESCRIBE DETAIL {table_name}").first()
                schema = spark.table(table_name).schema

                inventory.append(TableInventory(
                    database=db,
                    table=t.tableName,
                    table_type=detail.format or "unknown",
                    format=detail.format or "unknown",
                    row_count=spark.table(table_name).count(),
                    size_mb=detail.sizeInBytes / 1048576 if detail.sizeInBytes else 0,
                    columns=len(schema),
                    partitions=detail.partitionColumns or [],
                ))
            except Exception as e:
                print(f"  Skipping {table_name}: {e}")

    return inventory

# Generate migration plan
tables = assess_hive_metastore()
tables.sort(key=lambda t: t.size_mb, reverse=True)

print(f"\nTotal tables: {len(tables)}")
print(f"Total size: {sum(t.size_mb for t in tables):.0f} MB")
print(f"\nTop 10 by size:")
for t in tables[:10]:
    print(f"  {t.database}.{t.table}: {t.size_mb:.0f}MB, {t.row_count:,} rows, {t.format}")

Step 2: Schema Migration

# Schema conversion for common type mismatches
TYPE_MAP = {
    # Hadoop/Hive types → Delta Lake/Spark types
    "CHAR": "STRING",
    "VARCHAR": "STRING",
    "TINYINT": "INT",
    "SMALLINT": "INT",
    "BINARY": "BINARY",
    # Snowflake types
    "NUMBER": "DECIMAL",
    "VARIANT": "STRING",  # Store as JSON string, parse in Silver
    "TIMESTAMP_NTZ": "TIMESTAMP",
    "TIMESTAMP_TZ": "TIMESTAMP",
    # Redshift types
    "SUPER": "STRING",
    "TIMETZ": "TIMESTAMP",
}

def generate_create_table(source_table: str, target_table: str) -> str:
    """Generate CREATE TABLE DDL with type conversions."""
    schema = spark.table(source_table).schema
    cols = []
    for field in schema:
        dtype = TYPE_MAP.get(str(field.dataType).upper(), str(field.dataType))
        cols.append(f"  {field.name} {dtype}")

    return f"""CREATE TABLE IF NOT EXISTS {target_table} (
{',\n'.join(cols)}
) USING DELTA
TBLPROPERTIES (
    'delta.autoOptimize.optimizeWrite' = 'true',
    'delta.autoOptimize.autoCompact' = 'true'
);"""

Step 3: Data Migration with Validation

def migrate_table(
    source_table: str,
    target_table: str,
    method: str = "ctas",
    batch_size_mb: int = 500,
) -> dict:
    """Migrate a table with validation."""
    result = {"source": source_table, "target": target_table, "method": method}

    if method == "sync":
        # In-place metadata migration (fastest, no data copy)
        spark.sql(f"SYNC TABLE {target_table} FROM {source_table}")

    elif method == "deep_clone":
        # Delta-to-Delta with history preservation
        spark.sql(f"CREATE TABLE {target_table} DEEP CLONE {source_table}")

    elif method == "ctas":
        # Full data copy (works with any source format)
        source_size_mb = spark.sql(
            f"DESCRIBE DETAIL {source_table}"
        ).first().sizeInBytes / 1048576

        if source_size_mb > batch_size_mb:
            # Batch large tables by partition or row number
            spark.sql(f"""
                CREATE TABLE {target_table}
                USING DELTA
                AS SELECT * FROM {source_table}
            """)
        else:
            spark.sql(f"CREATE TABLE {target_table} AS SELECT * FROM {source_table}")

    elif method == "jdbc":
        # External database migration
        df = (spark.read
            .format("jdbc")
            .option("url", f"jdbc:postgresql://host:5432/db")
            .option("dbtable", source_table)
            .option("fetchsize", "10000")
            .load())
        df.write.format("delta").saveAsTable(target_table)

    # Validate
    src_count = spark.table(source_table).count()
    tgt_count = spark.table(target_table).count()
    result["source_rows"] = src_count
    result["target_rows"] = tgt_count
    result["match"] = src_count == tgt_count
    result["status"] = "OK" if result["match"] else "MISMATCH"

    return result

# Migrate with validation
result = migrate_table(
    "hive_metastore.legacy.customers",
    "analytics.migrated.customers",
    method="ctas",
)
print(f"{result['source']} -> {result['target']}: "
      f"{result['source_rows']:,} rows [{result['status']}]")

Step 4: Snowflake / Redshift Migration

# Snowflake: Use Lakehouse Federation or Unload + Auto Loader
# Option A: Lakehouse Federation (query in place, no copy)
spark.sql("""
    CREATE FOREIGN CATALOG snowflake_catalog
    USING CONNECTION snowflake_conn
    OPTIONS (database 'PROD_DB')
""")
# Query directly: SELECT * FROM snowflake_catalog.schema.table

# Option B: Unload to S3 + ingest
# In Snowflake:
# COPY INTO @my_s3_stage/export/customers/
# FROM PROD_DB.PUBLIC.CUSTOMERS
# FILE_FORMAT = (TYPE = PARQUET);

# In Databricks:
df = spark.read.parquet("s3://migration-bucket/export/customers/")
df.write.format("delta").saveAsTable("analytics.migrated.customers")

# Redshift: Unload to S3 + Auto Loader
# In Redshift:
# UNLOAD ('SELECT * FROM prod.customers')
# TO 's3://migration-bucket/redshift/customers/'
# FORMAT PARQUET;

# In Databricks:
(spark.readStream
    .format("cloudFiles")
    .option("cloudFiles.format", "parquet")
    .option("cloudFiles.schemaLocation", "/checkpoints/migration/schema")
    .load("s3://migration-bucket/redshift/customers/")
    .writeStream
    .format("delta")
    .option("checkpointLocation", "/checkpoints/migration/data")
    .toTable("analytics.migrated.customers"))

Step 5: ETL Pipeline Conversion

# Convert Oozie/Airflow jobs to Databricks Asset Bundles
# Before (Oozie/spark-submit):
#   spark-submit --class com.company.ETL --master yarn app.jar
#   hive -e "INSERT OVERWRITE TABLE target SELECT * FROM staging"

# After (Asset Bundle):
# databricks.yml resources:
"""
resources:
  jobs:
    migrated_etl:
      name: migrated-etl
      tasks:
        - task_key: extract
          notebook_task:
            notebook_path: src/extract.py
        - task_key: transform
          depends_on: [{task_key: extract}]
          notebook_task:
            notebook_path: src/transform.py
"""

# Convert HiveQL to Spark SQL
# Before: INSERT OVERWRITE TABLE target SELECT ...
# After:  (Use MERGE for upserts or write.mode("overwrite").saveAsTable)

Step 6: Cutover Planning

cutover_steps = [
    {"step": 1, "action": "Final validation", "rollback": "No action needed"},
    {"step": 2, "action": "Disable source pipelines", "rollback": "Re-enable source"},
    {"step": 3, "action": "Final data sync", "rollback": "Data already in place"},
    {"step": 4, "action": "Switch apps to Databricks endpoints", "rollback": "Revert app config"},
    {"step": 5, "action": "Enable Databricks pipelines", "rollback": "Disable and restore source"},
    {"step": 6, "action": "Monitor for 24 hours", "rollback": "Full rollback if issues"},
]

# Validation query to run at each step
validation_query = """
SELECT 'source' AS system, COUNT(*) AS rows FROM source_table
UNION ALL
SELECT 'target', COUNT(*) FROM target_table
"""

Output

• Migration assessment with table inventory (sizes, formats, dependencies)
• Schema conversion with type mapping and DDL generation
• Data migration with row-count validation per table
• ETL pipeline conversion from Oozie/Airflow to Asset Bundles
• Cutover plan with step-by-step rollback procedures

Error Handling

Error	Cause	Solution
Schema incompatibility	Unsupported types (VARIANT, SUPER)	Convert to STRING, parse in Silver layer
Row count mismatch	Truncation or filter during migration	Check for NULLs, encoding issues, or WHERE clauses
JDBC timeout	Large table extraction	Use fetchsize, partition reads, or incremental export
SYNC fails	External table storage inaccessible	Verify cloud storage credentials and network access
Pipeline dependency failure	Wrong migration order	Build dependency graph, migrate leaf tables first

Exam

---

Content truncated.