senior-data-engineer（高级数据工程师）

Senior Data Engineer

Production-grade data engineering skill for building scalable, reliable data systems.

Table of Contents

1. Trigger Phrases
2. Quick Start
3. Workflows

- Building a Batch ETL Pipeline - Implementing Real-Time Streaming - Data Quality Framework Setup

1. Architecture Decision Framework
2. Tech Stack
3. Reference Documentation
4. Troubleshooting

Trigger Phrases

Activate this skill when you see:

Pipeline Design:

• "Design a data pipeline for..."
• "Build an ETL/ELT process..."
• "How should I ingest data from..."
• "Set up data extraction from..."

Architecture:

• "Should I use batch or streaming?"
• "Lambda vs Kappa architecture"
• "How to handle late-arriving data"
• "Design a data lakehouse"

Data Modeling:

• "Create a dimensional model..."
• "Star schema vs snowflake"
• "Implement slowly changing dimensions"
• "Design a data vault"

Data Quality:

• "Add data validation to..."
• "Set up data quality checks"
• "Monitor data freshness"
• "Implement data contracts"

Performance:

• "Optimize this Spark job"
• "Query is running slow"
• "Reduce pipeline execution time"
• "Tune Airflow DAG"

Quick Start

Core Tools

# Generate pipeline orchestration config
python scripts/pipeline_orchestrator.py generate \
  --type airflow \
  --source postgres \
  --destination snowflake \
  --schedule "0 5 * * *"

# Validate data quality
python scripts/data_quality_validator.py validate \
  --input data/sales.parquet \
  --schema schemas/sales.json \
  --checks freshness,completeness,uniqueness

# Optimize ETL performance
python scripts/etl_performance_optimizer.py analyze \
  --query queries/daily_aggregation.sql \
  --engine spark \
  --recommend

Workflows

Workflow 1: Building a Batch ETL Pipeline

Scenario: Extract data from PostgreSQL, transform with dbt, load to Snowflake.

Step 1: Define Source Schema

-- Document source tables
SELECT
    table_name,
    column_name,
    data_type,
    is_nullable
FROM information_schema.columns
WHERE table_schema = 'source_schema'
ORDER BY table_name, ordinal_position;

Step 2: Generate Extraction Config

python scripts/pipeline_orchestrator.py generate \
  --type airflow \
  --source postgres \
  --tables orders,customers,products \
  --mode incremental \
  --watermark updated_at \
  --output dags/extract_source.py

Step 3: Create dbt Models

-- models/staging/stg_orders.sql
WITH source AS (
    SELECT * FROM {{ source('postgres', 'orders') }}
),

renamed AS (
    SELECT
        order_id,
        customer_id,
        order_date,
        total_amount,
        status,
        _extracted_at
    FROM source
    WHERE order_date >= DATEADD(day, -3, CURRENT_DATE)
)

SELECT * FROM renamed

-- models/marts/fct_orders.sql
{{
    config(
        materialized='incremental',
        unique_key='order_id',
        cluster_by=['order_date']
    )
}}

SELECT
    o.order_id,
    o.customer_id,
    c.customer_segment,
    o.order_date,
    o.total_amount,
    o.status
FROM {{ ref('stg_orders') }} o
LEFT JOIN {{ ref('dim_customers') }} c
    ON o.customer_id = c.customer_id

{% if is_incremental() %}
WHERE o._extracted_at > (SELECT MAX(_extracted_at) FROM {{ this }})
{% endif %}

Step 4: Configure Data Quality Tests

# models/marts/schema.yml
version: 2

models:
  - name: fct_orders
    description: "Order fact table"
    columns:
      - name: order_id
        tests:
          - unique
          - not_null
      - name: total_amount
        tests:
          - not_null
          - dbt_utils.accepted_range:
              min_value: 0
              max_value: 1000000
      - name: order_date
        tests:
          - not_null
          - dbt_utils.recency:
              datepart: day
              field: order_date
              interval: 1

Step 5: Create Airflow DAG

# dags/daily_etl.py
from airflow import DAG
from airflow.providers.postgres.operators.postgres import PostgresOperator
from airflow.operators.bash import BashOperator
from airflow.utils.dates import days_ago
from datetime import timedelta

default_args = {
    'owner': 'data-team',
    'depends_on_past': False,
    'email_on_failure': True,
    'email': ['data-alerts@company.com'],
    'retries': 2,
    'retry_delay': timedelta(minutes=5),
}

with DAG(
    'daily_etl_pipeline',
    default_args=default_args,
    description='Daily ETL from PostgreSQL to Snowflake',
    schedule_interval='0 5 * * *',
    start_date=days_ago(1),
    catchup=False,
    tags=['etl', 'daily'],
) as dag:

    extract = BashOperator(
        task_id='extract_source_data',
        bash_command='python /opt/airflow/scripts/extract.py --date {{ ds }}',
    )

    transform = BashOperator(
        task_id='run_dbt_models',
        bash_command='cd /opt/airflow/dbt && dbt run --select marts.*',
    )

    test = BashOperator(
        task_id='run_dbt_tests',
        bash_command='cd /opt/airflow/dbt && dbt test --select marts.*',
    )

    notify = BashOperator(
        task_id='send_notification',
        bash_command='python /opt/airflow/scripts/notify.py --status success',
        trigger_rule='all_success',
    )

    extract >> transform >> test >> notify

Step 6: Validate Pipeline

# Test locally
dbt run --select stg_orders fct_orders
dbt test --select fct_orders

# Validate data quality
python scripts/data_quality_validator.py validate \
  --table fct_orders \
  --checks all \
  --output reports/quality_report.json

Workflow 2: Implementing Real-Time Streaming

Scenario: Stream events from Kafka, process with Flink/Spark Streaming, sink to data lake.

Step 1: Define Event Schema

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "title": "UserEvent",
  "type": "object",
  "required": ["event_id", "user_id", "event_type", "timestamp"],
  "properties": {
    "event_id": {"type": "string", "format": "uuid"},
    "user_id": {"type": "string"},
    "event_type": {"type": "string", "enum": ["page_view", "click", "purchase"]},
    "timestamp": {"type": "string", "format": "date-time"},
    "properties": {"type": "object"}
  }
}

Step 2: Create Kafka Topic

# Create topic with appropriate partitions
kafka-topics.sh --create \
  --bootstrap-server localhost:9092 \
  --topic user-events \
  --partitions 12 \
  --replication-factor 3 \
  --config retention.ms=604800000 \
  --config cleanup.policy=delete

# Verify topic
kafka-topics.sh --describe \
  --bootstrap-server localhost:9092 \
  --topic user-events

Step 3: Implement Spark Streaming Job

# streaming/user_events_processor.py
from pyspark.sql import SparkSession
from pyspark.sql.functions import (
    from_json, col, window, count, avg,
    to_timestamp, current_timestamp
)
from pyspark.sql.types import (
    StructType, StructField, StringType,
    TimestampType, MapType
)

# Initialize Spark
spark = SparkSession.builder \
    .appName("UserEventsProcessor") \
    .config("spark.sql.streaming.checkpointLocation", "/checkpoints/user-events") \
    .config("spark.sql.shuffle.partitions", "12") \
    .getOrCreate()

# Define schema
event_schema = StructType([
    StructField("event_id", StringType(), False),
    StructField("user_id", StringType(), False),
    StructField("event_type", StringType(), False),
    StructField("timestamp", StringType(), False),
    StructField("properties", MapType(StringType(), StringType()), True)
])

# Read from Kafka
events_df = spark.readStream \
    .format("kafka") \
    .option("kafka.bootstrap.servers", "localhost:9092") \
    .option("subscribe", "user-events") \
    .option("startingOffsets", "latest") \
    .option("failOnDataLoss", "false") \
    .load()

# Parse JSON
parsed_df = events_df \
    .select(from_json(col("value").cast("string"), event_schema).alias("data")) \
    .select("data.*") \
    .withColumn("event_timestamp", to_timestamp(col("timestamp")))

# Windowed aggregation
aggregated_df = parsed_df \
    .withWatermark("event_timestamp", "10 minutes") \
    .groupBy(
        window(col("event_timestamp"), "5 minutes"),
        col("event_type")
    ) \
    .agg(
        count("*").alias("event_count"),
        approx_count_distinct("user_id").alias("unique_users")
    )

# Write to Delta Lake
query = aggregated_df.writeStream \
    .format("delta") \
    .outputMode("append") \
    .option("checkpointLocation", "/checkpoints/user-events-aggregated") \
    .option("path", "/data/lake/user_events_aggregated") \
    .trigger(processingTime="1 minute") \
    .start()

query.awaitTermination()

Step 4: Handle Late Data and Errors

# Dead letter queue for failed records
from pyspark.sql.functions import current_timestamp, lit

def process_with_error_handling(batch_df, batch_id):
    try:
        # Attempt processing
        valid_df = batch_df.filter(col("event_id").isNotNull())
        invalid_df = batch_df.filter(col("event_id").isNull())

        # Write valid records
        valid_df.write \
            .format("delta") \
            .mode("append") \
            .save("/data/lake/user_events")

        # Write invalid to DLQ
        if invalid_df.count() > 0:
            invalid_df \
                .withColumn("error_timestamp", current_timestamp()) \
                .withColumn("error_reason", lit("missing_event_id")) \
                .write \
                .format("delta") \
                .mode("append") \
                .save("/data/lake/dlq/user_events")

    except Exception as e:
        # Log error, alert, continue
        logger.error(f"Batch {batch_id} failed: {e}")
        raise

# Use foreachBatch for custom processing
query = parsed_df.writeStream \
    .foreachBatch(process_with_error_handling) \
    .option("checkpointLocation", "/checkpoints/user-events") \
    .start()

Step 5: Monitor Stream Health

# monitoring/stream_metrics.py
from prometheus_client import Gauge, Counter, start_http_server

# Define metrics
RECORDS_PROCESSED = Counter(
    'stream_records_processed_total',
    'Total records processed',
    ['stream_name', 'status']
)

PROCESSING_LAG = Gauge(
    'stream_processing_lag_seconds',
    'Current processing lag',
    ['stream_name']
)

BATCH_DURATION = Gauge(
    'stream_batch_duration_seconds',
    'Last batch processing duration',
    ['stream_name']
)

def emit_metrics(query):
    """Emit Prometheus metrics from streaming query."""
    progress = query.lastProgress
    if progress:
        RECORDS_PROCESSED.labels(
            stream_name='user-events',
            status='success'
        ).inc(progress['numInputRows'])

        if progress['sources']:
            # Calculate lag from latest offset
            for source in progress['sources']:
                end_offset = source.get('endOffset', {})
                # Parse Kafka offsets and calculate lag

Workflow 3: Data Quality Framework Setup

Scenario: Implement comprehensive data quality monitoring with Great Expectations.

Step 1: Initialize Great Expectations

# Install and initialize
pip install great_expectations

great_expectations init

# Connect to data source
great_expectations datasource new

Step 2: Create Expectation Suite

# expectations/orders_suite.py
import great_expectations as gx

context = gx.get_context()

# Create expectation suite
suite = context.add_expectation_suite("orders_quality_suite")

# Add expectations
validator = context.get_validator(
    batch_request={
        "datasource_name": "warehouse",
        "data_asset_name": "orders",
    },
    expectation_suite_name="orders_quality_suite"
)

# Schema expectations
validator.expect_table_columns_to_match_ordered_list(
    column_list=[
        "order_id", "customer_id", "order_date",
        "total_amount", "status", "created_at"
    ]
)

# Completeness expectations
validator.expect_column_values_to_not_be_null("order_id")
validator.expect_column_values_to_not_be_null("customer_id")
validator.expect_column_values_to_not_be_null("order_date")

# Uniqueness expectations
validator.expect_column_values_to_be_unique("order_id")

# Range expectations
validator.expect_column_values_to_be_between(
    "total_amount",
    min_value=0,
    max_value=1000000
)

# Categorical expectations
validator.expect_column_values_to_be_in_set(
    "status",
    ["pending", "confirmed", "shipped", "delivered", "cancelled"]
)

# Freshness expectation
validator.expect_column_max_to_be_between(
    "order_date",
    min_value={"$PARAMETER": "now - timedelta(days=1)"},
    max_value={"$PARAMETER": "now"}
)

# Referential integrity
validator.expect_column_values_to_be_in_set(
    "customer_id",
    value_set={"$PARAMETER": "valid_customer_ids"}
)

validator.save_expectation_suite(discard_failed_expectations=False)

Step 3: Create Data Quality Checks with dbt

# models/marts/schema.yml
version: 2

models:
  - name: fct_orders
    description: "Order fact table with data quality checks"

    tests:
      # Row count check
      - dbt_utils.equal_rowcount:
          compare_model: ref('stg_orders')

      # Freshness check
      - dbt_utils.recency:
          datepart: hour
          field: created_at
          interval: 24

    columns:
      - name: order_id
        description: "Unique order identifier"
        tests:
          - unique
          - not_null
          - relationships:
              to: ref('dim_orders')
              field: order_id

      - name: total_amount
        tests:
          - not_null
          - dbt_utils.accepted_range:
              min_value: 0
              max_value: 1000000
              inclusive: true
          - dbt_expectations.expect_column_values_to_be_between:
              min_value: 0
              row_condition: "status != 'cancelled'"

      - name: customer_id
        tests:
          - not_null
          - relationships:
              to: ref('dim_customers')
              field: customer_id
              severity: warn

Step 4: Implement Data Contracts

# contracts/orders_contract.yaml
contract:
  name: orders_data_contract
  version: "1.0.0"
  owner: data-team@company.com

schema:
  type: object
  properties:
    order_id:
      type: string
      format: uuid
      description: "Unique order identifier"
    customer_id:
      type: string
      not_null: true
    order_date:
      type: date
      not_null: true
    total_amount:
      type: decimal
      precision: 10
      scale: 2
      minimum: 0
    status:
      type: string
      enum: ["pending", "confirmed", "shipped", "delivered", "cancelled"]

sla:
  freshness:
    max_delay_hours: 1
  completeness:
    min_percentage: 99.9
  accuracy:
    duplicate_tolerance: 0.01

consumers:
  - name: analytics-team
    usage: "Daily reporting dashboards"
  - name: ml-team
    usage: "Churn prediction model"

Step 5: Set Up Quality Monitoring Dashboard

# monitoring/quality_dashboard.py
from datetime import datetime, timedelta
import pandas as pd

def generate_quality_report(connection, table_name: str) -> dict:
    """Generate comprehensive data quality report."""

    report = {
        "table": table_name,
        "timestamp": datetime.now().isoformat(),
        "checks": {}
    }

    # Row count check
    row_count = connection.execute(
        f"SELECT COUNT(*) FROM {table_name}"
    ).fetchone()[0]
    report["checks"]["row_count"] = {
        "value": row_count,
        "status": "pass" if row_count > 0 else "fail"
    }

    # Freshness check
    max_date = connection.execute(
        f"SELECT MAX(created_at) FROM {table_name}"
    ).fetchone()[0]
    hours_old = (datetime.now() - max_date).total_seconds() / 3600
    report["checks"]["freshness"] = {
        "max_timestamp": max_date.isoformat(),
        "hours_old": round(hours_old, 2),
        "status": "pass" if hours_old < 24 else "fail"
    }

    # Null rate check
    null_query = f"""
    SELECT
        SUM(CASE WHEN order_id IS NULL THEN 1 ELSE 0 END) as null_order_id,
        SUM(CASE WHEN customer_id IS NULL THEN 1 ELSE 0 END) as null_customer_id,
        COUNT(*) as total
    FROM {table_name}
    """
    null_result = connection.execute(null_query).fetchone()
    report["checks"]["null_rates"] = {
        "order_id": null_result[0] / null_result[2] if null_result[2] > 0 else 0,
        "customer_id": null_result[1] / null_result[2] if null_result[2] > 0 else 0,
        "status": "pass" if null_result[0] == 0 and null_result[1] == 0 else "fail"
    }

    # Duplicate check
    dup_query = f"""
    SELECT COUNT(*) - COUNT(DISTINCT order_id) as duplicates
    FROM {table_name}
    """
    duplicates = connection.execute(dup_query).fetchone()[0]
    report["checks"]["duplicates"] = {
        "count": duplicates,
        "status": "pass" if duplicates == 0 else "fail"
    }

    # Overall status
    all_passed = all(
        check["status"] == "pass"
        for check in report["checks"].values()
    )
    report["overall_status"] = "pass" if all_passed else "fail"

    return report

Architecture Decision Framework

Use this framework to choose the right approach for your data pipeline.

Batch vs Streaming

Decision Tree:

Is real-time insight required?
├── Yes → Use streaming
│   └── Is exactly-once semantics needed?
│       ├── Yes → Kafka + Flink/Spark Structured Streaming
│       └── No → Kafka + consumer groups
└── No → Use batch
    └── Is data volume > 1TB daily?
        ├── Yes → Spark/Databricks
        └── No → dbt + warehouse compute

Lambda vs Kappa Architecture

When to choose Lambda:

• Need to train ML models on historical data
• Complex batch transformations not feasible in streaming
• Existing batch infrastructure

When to choose Kappa:

• Event-sourced architecture
• All processing can be expressed as stream operations
• Starting fresh without legacy systems

Data Warehouse vs Data Lakehouse

Tech Stack

Reference Documentation

1. Data Pipeline Architecture

See references/data_pipeline_architecture.md for:

• Lambda vs Kappa architecture patterns
• Batch processing with Spark and Airflow
• Stream processing with Kafka and Flink
• Exactly-once semantics implementation
• Error handling and dead letter queues

2. Data Modeling Patterns

See references/data_modeling_patterns.md for:

• Dimensional modeling (Star/Snowflake)
• Slowly Changing Dimensions (SCD Types 1-6)
• Data Vault modeling
• dbt best practices
• Partitioning and clustering

3. DataOps Best Practices

See references/dataops_best_practices.md for:

• Data testing frameworks
• Data contracts and schema validation
• CI/CD for data pipelines
• Observability and lineage
• Incident response

Troubleshooting

Pipeline Failures

Symptom: Airflow DAG fails with timeout

Task exceeded max execution time

Solution:

1. Check resource allocation
2. Profile slow operations
3. Add incremental processing

# Increase timeout
default_args = {
    'execution_timeout': timedelta(hours=2),
}

# Or use incremental loads
WHERE updated_at > '{{ prev_ds }}'

Symptom: Spark job OOM

java.lang.OutOfMemoryError: Java heap space

Solution:

1. Increase executor memory
2. Reduce partition size
3. Use disk spill

spark.conf.set("spark.executor.memory", "8g")
spark.conf.set("spark.sql.shuffle.partitions", "200")
spark.conf.set("spark.memory.fraction", "0.8")

Symptom: Kafka consumer lag increasing

Consumer lag: 1000000 messages

Solution:

1. Increase consumer parallelism
2. Optimize processing logic
3. Scale consumer group

# Add more partitions
kafka-topics.sh --alter \
  --bootstrap-server localhost:9092 \
  --topic user-events \
  --partitions 24

Data Quality Issues

Symptom: Duplicate records appearing

Expected unique, found 150 duplicates

Solution:

1. Add deduplication logic
2. Use merge/upsert operations

-- dbt incremental with dedup
{{
    config(
        materialized='incremental',
        unique_key='order_id'
    )
}}

SELECT * FROM (
    SELECT
        *,
        ROW_NUMBER() OVER (
            PARTITION BY order_id
            ORDER BY updated_at DESC
        ) as rn
    FROM {{ source('raw', 'orders') }}
) WHERE rn = 1

Symptom: Stale data in tables

Last update: 3 days ago

Solution:

1. Check upstream pipeline status
2. Verify source availability
3. Add freshness monitoring

# dbt freshness check
sources:
  - name: raw
    freshness:
      warn_after: {count: 12, period: hour}
      error_after: {count: 24, period: hour}
    loaded_at_field: _loaded_at

Symptom: Schema drift detected

Column 'new_field' not in expected schema

Solution:

1. Update data contract
2. Modify transformations
3. Communicate with producers

# Handle schema evolution
df = spark.read.format("delta") \
    .option("mergeSchema", "true") \
    .load("/data/orders")

Performance Issues

Symptom: Query takes hours

Query runtime: 4 hours (expected: 30 minutes)

Solution:

1. Check query plan
2. Add proper partitioning
3. Optimize joins

-- Before: Full table scan
SELECT * FROM orders WHERE order_date = '2024-01-15';

-- After: Partition pruning
-- Table partitioned by order_date
SELECT * FROM orders WHERE order_date = '2024-01-15';

-- Add clustering for frequent filters
ALTER TABLE orders CLUSTER BY (customer_id);

Symptom: dbt model takes too long

Model fct_orders completed in 45 minutes

Solution:

1. Use incremental materialization
2. Reduce upstream dependencies
3. Pre-aggregate where possible

-- Convert to incremental
{{
    config(
        materialized='incremental',
        unique_key='order_id',
        on_schema_change='sync_all_columns'
    )
}}

SELECT * FROM {{ ref('stg_orders') }}
{% if is_incremental() %}
WHERE _loaded_at > (SELECT MAX(_loaded_at) FROM {{ this }})
{% endif %}