Skip to content
Back to Blog
1 min read

Change Data Capture in Microsoft Fabric

Microsoft Fabric CDC Data Integration ETL Real-Time

I wrote “Change Data Capture in Microsoft Fabric” to share practical, production-minded guidance on this topic.

CDC Fundamentals

CDC captures row-level changes (inserts, updates, deletes) from source systems:

Source Database
      │
      ├── INSERT → CDC Log → Fabric
      ├── UPDATE → CDC Log → Fabric
      └── DELETE → CDC Log → Fabric

CDC Options in Fabric

1. Database Mirroring (Built-in CDC)

The simplest option for supported sources:

# Mirroring handles CDC automatically
# Just enable change tracking on source

# Azure SQL
"""
ALTER DATABASE [YourDB] SET CHANGE_TRACKING = ON;
ALTER TABLE [dbo].[Orders] ENABLE CHANGE_TRACKING;
"""

# The mirroring service polls for changes and applies them

2. Data Factory CDC Connector

For more control over the CDC process:

def create_cdc_copy_activity():
    """Create CDC-aware copy activity in Data Factory."""

    activity = {
        "name": "IncrementalCopy",
        "type": "Copy",
        "inputs": [{
            "referenceName": "SqlChangeTrackingSource",
            "type": "DatasetReference"
        }],
        "outputs": [{
            "referenceName": "LakehouseDelta",
            "type": "DatasetReference"
        }],
        "typeProperties": {
            "source": {
                "type": "SqlServerSource",
                "sqlReaderQuery": """
                    DECLARE @from_lsn binary(10), @to_lsn binary(10);
                    SET @from_lsn = @LastLSN;
                    SET @to_lsn = sys.fn_cdc_get_max_lsn();

                    SELECT
                        CT.__$operation,
                        CT.__$start_lsn,
                        T.*
                    FROM cdc.fn_cdc_get_all_changes_dbo_Orders(
                        @from_lsn, @to_lsn, 'all'
                    ) CT
                    JOIN dbo.Orders T ON CT.OrderID = T.OrderID
                """
            },
            "sink": {
                "type": "DeltaLakeSink",
                "writeBehavior": "upsert",
                "updateCondition": "__$operation IN (2, 4)",  # Update operations
                "deleteCondition": "__$operation = 1"         # Delete operations
            }
        }
    }

    return activity

3. Eventstream with CDC Sources

For streaming CDC:

class EventstreamCDC:
    def create_debezium_source(self):
        """Configure Debezium CDC source for Eventstream."""

        # Debezium connector configuration
        connector_config = {
            "name": "sql-server-connector",
            "connector.class": "io.debezium.connector.sqlserver.SqlServerConnector",
            "database.hostname": "sqlserver.database.windows.net",
            "database.port": "1433",
            "database.user": "cdc_user",
            "database.password": "${secrets:sql-password}",
            "database.dbname": "SalesDB",
            "database.server.name": "salesdb",
            "table.include.list": "dbo.Orders,dbo.Customers",
            "database.history.kafka.bootstrap.servers": "eventhub.servicebus.windows.net:9093",
            "database.history.kafka.topic": "schema-changes"
        }

        return connector_config

    def process_cdc_events(self):
        """Eventstream processing for CDC events."""

        eventstream_config = {
            "name": "CDC-Orders-Stream",
            "source": {
                "type": "Kafka",  # Debezium outputs to Kafka-compatible endpoint
                "topic": "salesdb.dbo.Orders"
            },
            "transformations": [
                {
                    "name": "parse_debezium",
                    "type": "ParseJson",
                    "extract_fields": ["before", "after", "op", "ts_ms"]
                },
                {
                    "name": "transform_changes",
                    "type": "Expression",
                    "expressions": [
                        "CASE op WHEN 'c' THEN 'INSERT' WHEN 'u' THEN 'UPDATE' WHEN 'd' THEN 'DELETE' END AS operation",
                        "COALESCE(after.OrderID, before.OrderID) AS OrderID",
                        "after.* AS current_values"
                    ]
                }
            ],
            "destinations": [
                {
                    "type": "Lakehouse",
                    "table": "orders_cdc_raw",
                    "mode": "append"
                }
            ]
        }

        return eventstream_config

Implementing CDC Processing

Delta Lake Merge Pattern

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

class CDCProcessor:
    def __init__(self, spark, target_path: str):
        self.spark = spark
        self.target_path = target_path

    def apply_changes(self, changes_df, key_columns: list[str]):
        """Apply CDC changes to Delta table."""

        # Check if target exists
        if DeltaTable.isDeltaTable(self.spark, self.target_path):
            target = DeltaTable.forPath(self.spark, self.target_path)

            # Build merge condition
            merge_condition = " AND ".join(
                f"target.{col} = changes.{col}"
                for col in key_columns
            )

            # Apply changes based on operation type
            target.alias("target").merge(
                changes_df.alias("changes"),
                merge_condition
            ).whenMatchedUpdate(
                condition="changes.__$operation = 4",  # Update
                set=self._build_update_set(changes_df.columns, key_columns)
            ).whenMatchedDelete(
                condition="changes.__$operation = 1"   # Delete
            ).whenNotMatchedInsert(
                condition="changes.__$operation IN (2, 4)",  # Insert or Update (new row)
                values=self._build_insert_values(changes_df.columns)
            ).execute()

        else:
            # Initial load
            initial_data = changes_df.filter(
                col("__$operation").isin([2, 4])  # Inserts and Updates
            ).drop("__$operation", "__$start_lsn")

            initial_data.write.format("delta").save(self.target_path)

    def _build_update_set(self, columns: list, key_columns: list) -> dict:
        """Build update set excluding keys and CDC columns."""
        update_cols = [
            c for c in columns
            if c not in key_columns and not c.startswith("__$")
        ]
        return {col: f"changes.{col}" for col in update_cols}

    def _build_insert_values(self, columns: list) -> dict:
        """Build insert values excluding CDC columns."""
        insert_cols = [c for c in columns if not c.startswith("__$")]
        return {col: f"changes.{col}" for col in insert_cols}

# Usage
processor = CDCProcessor(spark, "Tables/orders")

# Read CDC changes
changes = spark.read.format("delta").load("Tables/orders_cdc_raw") \
    .filter(col("processed") == False)

# Apply changes
processor.apply_changes(changes, key_columns=["OrderID"])

Handling Late-Arriving Data

def handle_late_arrivals(self, changes_df, watermark_column: str, max_delay_hours: int = 24):
    """Handle out-of-order CDC events."""

    # Add processing timestamp
    changes_with_ts = changes_df.withColumn(
        "processing_time",
        current_timestamp()
    )

    # Filter out events that are too old
    valid_changes = changes_with_ts.filter(
        col(watermark_column) >= (
            current_timestamp() - expr(f"INTERVAL {max_delay_hours} HOURS")
        )
    )

    # Log dropped events
    dropped = changes_with_ts.exceptAll(valid_changes)
    if dropped.count() > 0:
        dropped.write.format("delta").mode("append").save("Tables/cdc_dropped_events")
        print(f"Dropped {dropped.count()} late events")

    return valid_changes

CDC State Management

class CDCStateManager:
    """Track CDC processing state for incremental loads."""

    def __init__(self, state_path: str):
        self.state_path = state_path

    def get_last_lsn(self, table_name: str) -> bytes:
        """Get last processed LSN for a table."""
        try:
            state = spark.read.format("delta").load(self.state_path) \
                .filter(col("table_name") == table_name) \
                .orderBy(col("updated_at").desc()) \
                .first()

            return state["last_lsn"] if state else None
        except:
            return None

    def update_lsn(self, table_name: str, lsn: bytes, rows_processed: int):
        """Update last processed LSN."""
        from pyspark.sql import Row

        state_row = Row(
            table_name=table_name,
            last_lsn=lsn,
            rows_processed=rows_processed,
            updated_at=datetime.utcnow()
        )

        spark.createDataFrame([state_row]).write \
            .format("delta") \
            .mode("append") \
            .save(self.state_path)

    def get_processing_stats(self, table_name: str, days: int = 7) -> dict:
        """Get CDC processing statistics."""
        stats = spark.read.format("delta").load(self.state_path) \
            .filter(col("table_name") == table_name) \
            .filter(col("updated_at") >= date_sub(current_date(), days)) \
            .agg(
                sum("rows_processed").alias("total_rows"),
                count("*").alias("sync_count"),
                min("updated_at").alias("first_sync"),
                max("updated_at").alias("last_sync")
            ).first()

        return stats.asDict() if stats else {}

Best Practices

  1. Enable CDC early: Easier to set up before data volume grows
  2. Monitor LSN/watermarks: Track your position in the change stream
  3. Handle deletes carefully: Soft deletes often preferred for analytics
  4. Test recovery scenarios: Know how to replay from a specific point
  5. Consider retention: CDC logs can grow large

Conclusion

CDC in Fabric enables efficient incremental data processing. Whether through built-in mirroring, Data Factory, or custom implementations, understanding CDC patterns is essential for modern data platforms.

Choose the approach that matches your latency requirements and operational capabilities.

Michael John Peña

Michael John Peña

Senior Data Engineer based in Sydney. Writing about data, cloud, and technology.